kwimage

The Kitware Image Module (kwimage) contains functions to accomplish lower-level image operations via a high level API.

Subpackages

• kwimage.algo
  • kwimage.algo._nms_backend
    • kwimage.algo._nms_backend.py_nms
    • kwimage.algo._nms_backend.torch_nms
  • kwimage.algo.algo_nms
• kwimage.structs
  • kwimage.structs._boxes_backend
  • kwimage.structs._mask_backend
  • kwimage.structs._generic
  • kwimage.structs.boxes
  • kwimage.structs.coords
  • kwimage.structs.detections
  • kwimage.structs.heatmap
  • kwimage.structs.mask
  • kwimage.structs.points
  • kwimage.structs.polygon
  • kwimage.structs.segmentation

Submodules

• kwimage.im_alphablend
• kwimage.im_color
• kwimage.im_core
• kwimage.im_cv2
• kwimage.im_demodata
• kwimage.im_draw
• kwimage.im_filter
• kwimage.im_io
• kwimage.im_runlen
• kwimage.im_stack
• kwimage.transform
• kwimage.util_warp

Package Contents

Classes

Color	Used for converting a single color between spaces and encodings.
Boxes	Converts boxes between different formats as long as the last dimension
Coords	A data structure to store n-dimensional coordinate geometry.
Detections	Container for holding and manipulating multiple detections.
Heatmap	Keeps track of a downscaled heatmap and how to transform it to overlay the
Mask	Manages a single segmentation mask and can convert to and from
MaskList	Store and manipulate multiple masks, usually within the same image
MultiPolygon	Data structure for storing multiple polygons (typically related to the same
Points	Stores multiple keypoints for a single object.
PointsList	Stores a list of Points, each item usually corresponds to a different object.
Polygon	Represents a single polygon as set of exterior boundary points and a list
PolygonList	Stores and allows manipluation of multiple polygons, usually within the
Segmentation	Either holds a MultiPolygon, Polygon, or Mask
SegmentationList	Store and manipulate multiple segmentations (masks or polygons), usually
Affine	Helper for making affine transform matrices.
Linear	Base class for matrix-based transform.
Matrix	Base class for matrix-based transform.
Projective	Currently just a stub class that may be used to implement projective /
Transform	Inherit from this class and define __nice__ to “nicely” print your

Functions

available_nms_impls()	List available values for the impl kwarg of non_max_supression
daq_spatial_nms(ltrb, scores, diameter, thresh, max_depth=6, stop_size=2048, recsize=2048, impl='auto', device_id=None)	Divide and conquor speedup non-max-supression algorithm for when bboxes
non_max_supression(ltrb, scores, thresh, bias=0.0, classes=None, impl='auto', device_id=None)	Non-Maximum Suppression - remove redundant bounding boxes
ensure_alpha_channel(img, alpha=1.0, dtype=np.float32, copy=False)	Returns the input image with 4 channels.
overlay_alpha_images(img1, img2, keepalpha=True, dtype=np.float32, impl='inplace')	Places img1 on top of img2 respecting alpha channels.
overlay_alpha_layers(layers, keepalpha=True, dtype=np.float32)	Stacks a sequences of layers on top of one another. The first item is the
atleast_3channels(arr, copy=True)	Ensures that there are 3 channels in the image
ensure_float01(img, dtype=np.float32, copy=True)	Ensure that an image is encoded using a float32 properly
ensure_uint255(img, copy=True)	Ensure that an image is encoded using a uint8 properly. Either
make_channels_comparable(img1, img2, atleast3d=False)	Broadcasts image arrays so they can have elementwise operations applied
normalize(arr, mode='linear', alpha=None, beta=None, out=None)	Rebalance pixel intensities via contrast stretching.
normalize_intensity(imdata, return_info=False, nodata=None, axis=None, dtype=np.float32)	Normalize data intensities using heuristics to help put sensor data with
num_channels(img)	Returns the number of color channels in an image.
padded_slice(data, in_slice, pad=None, padkw=None, return_info=False)	Allows slices with out-of-bound coordinates. Any out of bounds coordinate
convert_colorspace(img, src_space, dst_space, copy=False, implicit=False, dst=None)	Converts colorspace of img.
gaussian_patch(shape=(7, 7), sigma=None)	Creates a 2D gaussian patch with a specific size and sigma
imcrop(img, dsize, about=None, origin=None, border_value=None, interpolation='nearest')	Crop an image about a specified point, padding if necessary.
imresize(img, scale=None, dsize=None, max_dim=None, min_dim=None, interpolation=None, grow_interpolation=None, letterbox=False, return_info=False, antialias=False)	Resize an image based on a scale factor, final size, or size and aspect
imscale(img, scale, interpolation=None, return_scale=False)	DEPRECATED and removed: use imresize instead
warp_affine(image, transform, dsize=None, antialias=False, interpolation='linear', border_mode=None, border_value=0, large_warp_dim=None, return_info=False)	Applies an affine transformation to an image with optional antialiasing.
checkerboard(num_squares=8, dsize=(512, 512))	Creates a checkerboard image
grab_test_image(key='astro', space='rgb', dsize=None, interpolation='lanczos')	Ensures that the test image exists (this might use the network), reads it
grab_test_image_fpath(key='astro')	Ensures that the test image exists (this might use the network) and returns
draw_boxes_on_image(img, boxes, color='blue', thickness=1, box_format=None, colorspace='rgb')	Draws boxes on an image.
draw_clf_on_image(im, classes, tcx=None, probs=None, pcx=None, border=1)	Draws classification label on an image.
draw_line_segments_on_image(img, pts1, pts2, color='blue', colorspace='rgb', thickness=1, **kwargs)	Draw line segments between pts1 and pts2 on an image.
draw_text_on_image(img, text, org, return_info=False, **kwargs)	Draws multiline text on an image using opencv
draw_vector_field(image, dx, dy, stride=0.02, thresh=0.0, scale=1.0, alpha=1.0, color='red', thickness=1, tipLength=0.1, line_type='aa')	Create an image representing a 2D vector field.
make_heatmask(probs, cmap='plasma', with_alpha=1.0, space='rgb', dsize=None)	Colorizes a single-channel intensity mask (with an alpha channel)
make_orimask(radians, mag=None, alpha=1.0)	Makes a colormap in HSV space where the orientation changes color and mag
make_vector_field(dx, dy, stride=0.02, thresh=0.0, scale=1.0, alpha=1.0, color='red', thickness=1, tipLength=0.1, line_type='aa')	Create an image representing a 2D vector field.
fourier_mask(img_hwc, mask, axis=None, clip=None)	Applies a mask to the fourier spectrum of an image
radial_fourier_mask(img_hwc, radius=11, axis=None, clip=None)	In [1] they use a radius of 11.0 on CIFAR-10.
imread(fpath, space='auto', backend='auto')	Reads image data in a specified format using some backend implementation.
imwrite(fpath, image, space='auto', backend='auto', **kwargs)	Writes image data to disk.
load_image_shape(fpath)	Determine the height/width/channels of an image without reading the entire
decode_run_length(counts, shape, binary=False, dtype=np.uint8, order='C')	Decode run length encoding back into an image.
encode_run_length(img, binary=False, order='C')	Construct the run length encoding (RLE) of an image.
rle_translate(rle, offset, output_shape=None)	Translates a run-length encoded image in RLE-space.
stack_images(images, axis=0, resize=None, interpolation=None, overlap=0, return_info=False, bg_value=None)	Make a new image with the input images side-by-side
stack_images_grid(images, chunksize=None, axis=0, overlap=0, return_info=False, bg_value=None)	Stacks images in a grid. Optionally return transforms of original image
smooth_prob(prob, k=3, inplace=False, eps=1e-09)	Smooths the probability map, but preserves the magnitude of the peaks.
add_homog(pts)	Add a homogenous coordinate to a point array
remove_homog(pts, mode='divide')	Remove homogenous coordinate to a point array.
subpixel_accum(dst, src, index, interp_axes=None)	Add the source values array into the destination array at a particular
subpixel_align(dst, src, index, interp_axes=None)	Returns an aligned version of the source tensor and destination index.
subpixel_getvalue(img, pts, coord_axes=None, interp='bilinear', bordermode='edge')	Get values at subpixel locations
subpixel_maximum(dst, src, index, interp_axes=None)	Take the max of the source values array into and the destination array at a
subpixel_minimum(dst, src, index, interp_axes=None)	Take the min of the source values array into and the destination array at a
subpixel_set(dst, src, index, interp_axes=None)	Add the source values array into the destination array at a particular
subpixel_setvalue(img, pts, value, coord_axes=None, interp='bilinear', bordermode='edge')	Set values at subpixel locations
subpixel_slice(inputs, index)	Take a subpixel slice from a larger image. The returned output is
subpixel_translate(inputs, shift, interp_axes=None, output_shape=None)	Translates an image by a subpixel shift value using bilinear interpolation
warp_image(inputs, mat, **kw)
warp_points(matrix, pts, homog_mode='divide')	Warp ND points / coordinates using a transformation matrix.
warp_tensor(inputs, mat, output_dims, mode='bilinear', padding_mode='zeros', isinv=False, ishomog=None, align_corners=False, new_mode=False)	A pytorch implementation of warp affine that works similarly to

kwimage.available_nms_impls()

List available values for the impl kwarg of non_max_supression

CommandLine:

xdoctest -m kwimage.algo.algo_nms available_nms_impls

Example

>>> impls = available_nms_impls()
>>> assert 'numpy' in impls
>>> print('impls = {!r}'.format(impls))

kwimage.daq_spatial_nms(ltrb, scores, diameter, thresh, max_depth=6, stop_size=2048, recsize=2048, impl='auto', device_id=None)

Divide and conquor speedup non-max-supression algorithm for when bboxes have a known max size

Parameters

• ltrb (ndarray) – boxes in (tlx, tly, brx, bry) format

• scores (ndarray) – scores of each box

• diameter (int or Tuple[int, int]) – Distance from split point to consider rectification. If specified as an integer, then number is used for both height and width. If specified as a tuple, then dims are assumed to be in [height, width] format.

• thresh (float) – iou threshold. Boxes are removed if they overlap greater than this threshold. 0 is the most strict, resulting in the fewest boxes, and 1 is the most permissive resulting in the most.

• max_depth (int) – maximum number of times we can divide and conquor

• stop_size (int) – number of boxes that triggers full NMS computation

• recsize (int) – number of boxes that triggers full NMS recombination

• impl (str) – algorithm to use

LookInfo:

# Didn’t read yet but it seems similar http://www.cyberneum.de/fileadmin/user_upload/files/publications/CVPR2010-Lampert_[0].pdf

https://www.researchgate.net/publication/220929789_Efficient_Non-Maximum_Suppression

# This seems very similar https://projet.liris.cnrs.fr/m2disco/pub/Congres/2006-ICPR/DATA/C03_0406.PDF

Example

>>> import kwimage
>>> # Make a bunch of boxes with the same width and height
>>> #boxes = kwimage.Boxes.random(230397, scale=1000, format='cxywh')
>>> boxes = kwimage.Boxes.random(237, scale=1000, format='cxywh')
>>> boxes.data.T[2] = 10
>>> boxes.data.T[3] = 10
>>> #
>>> ltrb = boxes.to_ltrb().data.astype(np.float32)
>>> scores = np.arange(0, len(ltrb)).astype(np.float32)
>>> #
>>> n_megabytes = (ltrb.size * ltrb.dtype.itemsize) / (2 ** 20)
>>> print('n_megabytes = {!r}'.format(n_megabytes))
>>> #
>>> thresh = iou_thresh = 0.01
>>> impl = 'auto'
>>> max_depth = 20
>>> diameter = 10
>>> stop_size = 2000
>>> recsize = 500
>>> #
>>> import ubelt as ub
>>> #
>>> with ub.Timer(label='daq'):
>>>     keep1 = daq_spatial_nms(ltrb, scores,
>>>         diameter=diameter, thresh=thresh, max_depth=max_depth,
>>>         stop_size=stop_size, recsize=recsize, impl=impl)
>>> #
>>> with ub.Timer(label='full'):
>>>     keep2 = non_max_supression(ltrb, scores,
>>>         thresh=thresh, impl=impl)
>>> #
>>> # Due to the greedy nature of the algorithm, there will be slight
>>> # differences in results, but they will be mostly similar.
>>> similarity = len(set(keep1) & set(keep2)) / len(set(keep1) | set(keep2))
>>> print('similarity = {!r}'.format(similarity))

kwimage.non_max_supression(ltrb, scores, thresh, bias=0.0, classes=None, impl='auto', device_id=None)

Non-Maximum Suppression - remove redundant bounding boxes

Parameters

• ltrb (ndarray[float32]) – Nx4 boxes in ltrb format

• scores (ndarray[float32]) – score for each bbox

• thresh (float) – iou threshold. Boxes are removed if they overlap greater than this threshold (i.e. Boxes are removed if iou > threshold). Thresh = 0 is the most strict, resulting in the fewest boxes, and 1 is the most permissive resulting in the most.

• bias (float) – bias for iou computation either 0 or 1

• classes (ndarray[int64] or None) – integer classes. If specified NMS is done on a perclass basis.

• impl (str) – implementation can be “auto”, “python”, “cython_cpu”, “gpu”, “torch”, or “torchvision”.

• device_id (int) – used if impl is gpu, device id to work on. If not specified torch.cuda.current_device() is used.

Notes

Using impl=’cython_gpu’ may result in an CUDA memory error that is not exposed to the python processes. In other words your program will hard crash if impl=’cython_gpu’, and you feed it too many bounding boxes. Ideally this will be fixed in the future.

References

https://github.com/facebookresearch/Detectron/blob/master/detectron/utils/cython_nms.pyx https://www.pyimagesearch.com/2015/02/16/faster-non-maximum-suppression-python/ https://github.com/bharatsingh430/soft-nms/blob/master/lib/nms/cpu_nms.pyx <- TODO

CommandLine:

xdoctest -m ~/code/kwimage/kwimage/algo/algo_nms.py non_max_supression

Example

>>> from kwimage.algo.algo_nms import *
>>> from kwimage.algo.algo_nms import _impls
>>> ltrb = np.array([
>>>     [0, 0, 100, 100],
>>>     [100, 100, 10, 10],
>>>     [10, 10, 100, 100],
>>>     [50, 50, 100, 100],
>>> ], dtype=np.float32)
>>> scores = np.array([.1, .5, .9, .1])
>>> keep = non_max_supression(ltrb, scores, thresh=0.5, impl='numpy')
>>> print('keep = {!r}'.format(keep))
>>> assert keep == [2, 1, 3]
>>> thresh = 0.0
>>> non_max_supression(ltrb, scores, thresh, impl='numpy')
>>> if 'numpy' in available_nms_impls():
>>>     keep = non_max_supression(ltrb, scores, thresh, impl='numpy')
>>>     assert list(keep) == [2, 1]
>>> if 'cython_cpu' in available_nms_impls():
>>>     keep = non_max_supression(ltrb, scores, thresh, impl='cython_cpu')
>>>     assert list(keep) == [2, 1]
>>> if 'cython_gpu' in available_nms_impls():
>>>     keep = non_max_supression(ltrb, scores, thresh, impl='cython_gpu')
>>>     assert list(keep) == [2, 1]
>>> if 'torch' in available_nms_impls():
>>>     keep = non_max_supression(ltrb, scores, thresh, impl='torch')
>>>     assert set(keep.tolist()) == {2, 1}
>>> if 'torchvision' in available_nms_impls():
>>>     keep = non_max_supression(ltrb, scores, thresh, impl='torchvision')  # note torchvision has no bias
>>>     assert list(keep) == [2]
>>> thresh = 1.0
>>> if 'numpy' in available_nms_impls():
>>>     keep = non_max_supression(ltrb, scores, thresh, impl='numpy')
>>>     assert list(keep) == [2, 1, 3, 0]
>>> if 'cython_cpu' in available_nms_impls():
>>>     keep = non_max_supression(ltrb, scores, thresh, impl='cython_cpu')
>>>     assert list(keep) == [2, 1, 3, 0]
>>> if 'cython_gpu' in available_nms_impls():
>>>     keep = non_max_supression(ltrb, scores, thresh, impl='cython_gpu')
>>>     assert list(keep) == [2, 1, 3, 0]
>>> if 'torch' in available_nms_impls():
>>>     keep = non_max_supression(ltrb, scores, thresh, impl='torch')
>>>     assert set(keep.tolist()) == {2, 1, 3, 0}
>>> if 'torchvision' in available_nms_impls():
>>>     keep = non_max_supression(ltrb, scores, thresh, impl='torchvision')  # note torchvision has no bias
>>>     assert set(kwarray.ArrayAPI.tolist(keep)) == {2, 1, 3, 0}

Example

>>> import ubelt as ub
>>> ltrb = np.array([
>>>     [0, 0, 100, 100],
>>>     [100, 100, 10, 10],
>>>     [10, 10, 100, 100],
>>>     [50, 50, 100, 100],
>>>     [100, 100, 150, 101],
>>>     [120, 100, 180, 101],
>>>     [150, 100, 200, 101],
>>> ], dtype=np.float32)
>>> scores = np.linspace(0, 1, len(ltrb))
>>> thresh = .2
>>> solutions = {}
>>> if not _impls._funcs:
>>>     _impls._lazy_init()
>>> for impl in _impls._funcs:
>>>     keep = non_max_supression(ltrb, scores, thresh, impl=impl)
>>>     solutions[impl] = sorted(keep)
>>> assert 'numpy' in solutions
>>> print('solutions = {}'.format(ub.repr2(solutions, nl=1)))
>>> assert ub.allsame(solutions.values())

CommandLine:

xdoctest -m ~/code/kwimage/kwimage/algo/algo_nms.py non_max_supression

Example

>>> import ubelt as ub
>>> # Check that zero-area boxes are ok
>>> ltrb = np.array([
>>>     [0, 0, 0, 0],
>>>     [0, 0, 0, 0],
>>>     [10, 10, 10, 10],
>>> ], dtype=np.float32)
>>> scores = np.array([1, 2, 3], dtype=np.float32)
>>> thresh = .2
>>> solutions = {}
>>> if not _impls._funcs:
>>>     _impls._lazy_init()
>>> for impl in _impls._funcs:
>>>     keep = non_max_supression(ltrb, scores, thresh, impl=impl)
>>>     solutions[impl] = sorted(keep)
>>> assert 'numpy' in solutions
>>> print('solutions = {}'.format(ub.repr2(solutions, nl=1)))
>>> assert ub.allsame(solutions.values())

kwimage.ensure_alpha_channel(img, alpha=1.0, dtype=np.float32, copy=False)

Returns the input image with 4 channels.

Parameters

• img (ndarray) – an image with shape [H, W], [H, W, 1], [H, W, 3], or [H, W, 4].

• alpha (float, default=1.0) – default value for missing alpha channel

• dtype (type, default=np.float32) – a numpy floating type

• copy (bool, default=False) – always copy if True, else copy if needed.

Returns

an image with specified dtype with shape [H, W, 4].

Raises

ValueError - if the input image does not have 1, 3, or 4 input channels – or if the image cannot be converted into a float01 representation

kwimage.overlay_alpha_images(img1, img2, keepalpha=True, dtype=np.float32, impl='inplace')

Places img1 on top of img2 respecting alpha channels. Works like the Photoshop layers with opacity.

Parameters

• img1 (ndarray) – top image to overlay over img2

• img2 (ndarray) – base image to superimpose on

• keepalpha (bool) – if False, the alpha channel is removed after blending

• dtype (np.dtype) – format for blending computation (defaults to float32)

• impl (str, default=inplace) – code specifying the backend implementation

Returns

raster: the blended images

Return type

ndarray

Todo

• [ ] Make fast C++ version of this function

References

http://stackoverflow.com/questions/25182421/overlay-numpy-alpha https://en.wikipedia.org/wiki/Alpha_compositing#Alpha_blending

Example

>>> import kwimage
>>> img1 = kwimage.grab_test_image('astro', dsize=(100, 100))
>>> img2 = kwimage.grab_test_image('carl', dsize=(100, 100))
>>> img1 = kwimage.ensure_alpha_channel(img1, alpha=.5)
>>> img3 = overlay_alpha_images(img1, img2)
>>> # xdoctest: +REQUIRES(--show)
>>> import kwplot
>>> kwplot.autompl()
>>> kwplot.imshow(img3)
>>> kwplot.show_if_requested()

kwimage.overlay_alpha_layers(layers, keepalpha=True, dtype=np.float32)

Stacks a sequences of layers on top of one another. The first item is the topmost layer and the last item is the bottommost layer.

Parameters

• layers (Sequence[ndarray]) – stack of images

• keepalpha (bool) – if False, the alpha channel is removed after blending

• dtype (np.dtype) – format for blending computation (defaults to float32)

Returns

raster: the blended images

Return type

ndarray

References

http://stackoverflow.com/questions/25182421/overlay-numpy-alpha https://en.wikipedia.org/wiki/Alpha_compositing#Alpha_blending

Example

>>> import kwimage
>>> keys = ['astro', 'carl', 'stars']
>>> layers = [kwimage.grab_test_image(k, dsize=(100, 100)) for k in keys]
>>> layers = [kwimage.ensure_alpha_channel(g, alpha=.5) for g in layers]
>>> stacked = overlay_alpha_layers(layers)
>>> # xdoctest: +REQUIRES(--show)
>>> import kwplot
>>> kwplot.autompl()
>>> kwplot.imshow(stacked)
>>> kwplot.show_if_requested()

class kwimage.Color(color, alpha=None, space=None)

Bases: ubelt.NiceRepr

Used for converting a single color between spaces and encodings. This should only be used when handling small numbers of colors(e.g. 1), don’t use this to represent an image.

move to colorutil?

Parameters

space (str) – colorspace of wrapped color. Assume RGB if not specified and it cannot be inferred

CommandLine:

xdoctest -m ~/code/kwimage/kwimage/im_color.py Color

Example

>>> print(Color('g'))
>>> print(Color('orangered'))
>>> print(Color('#AAAAAA').as255())
>>> print(Color([0, 255, 0]))
>>> print(Color([1, 1, 1.]))
>>> print(Color([1, 1, 1]))
>>> print(Color(Color([1, 1, 1])).as255())
>>> print(Color(Color([1., 0, 1, 0])).ashex())
>>> print(Color([1, 1, 1], alpha=255))
>>> print(Color([1, 1, 1], alpha=255, space='lab'))

__nice__(self)

_forimage(self, image, space='rgb')

Experimental function.

Create a numeric color tuple that agrees with the format of the input image (i.e. float or int, with 3 or 4 channels).

Parameters

• image (ndarray) – image to return color for

• space (str, default=rgb) – colorspace of the input image.

Example

>>> img_f3 = np.zeros([8, 8, 3], dtype=np.float32)
>>> img_u3 = np.zeros([8, 8, 3], dtype=np.uint8)
>>> img_f4 = np.zeros([8, 8, 4], dtype=np.float32)
>>> img_u4 = np.zeros([8, 8, 4], dtype=np.uint8)
>>> Color('red')._forimage(img_f3)
(1.0, 0.0, 0.0)
>>> Color('red')._forimage(img_f4)
(1.0, 0.0, 0.0, 1.0)
>>> Color('red')._forimage(img_u3)
(255, 0, 0)
>>> Color('red')._forimage(img_u4)
(255, 0, 0, 255)
>>> Color('red', alpha=0.5)._forimage(img_f4)
(1.0, 0.0, 0.0, 0.5)
>>> Color('red', alpha=0.5)._forimage(img_u4)
(255, 0, 0, 127)

ashex(self, space=None)

as255(self, space=None)

as01(self, space=None)

self = mplutil.Color(‘red’) mplutil.Color(‘green’).as01(‘rgba’)

classmethod _is_base01(channels)

check if a color is in base 01

classmethod _is_base255(Color, channels)

there is a one corner case where all pixels are 1 or less

classmethod _hex_to_01(Color, hex_color)

hex_color = ‘#6A5AFFAF’

_ensure_color01(Color, color)

Infer what type color is and normalize to 01

classmethod _255_to_01(Color, color255)

converts base 255 color to base 01 color

classmethod _string_to_01(Color, color)

mplutil.Color._string_to_01(‘green’) mplutil.Color._string_to_01(‘red’)

classmethod named_colors(cls)

Returns

names of colors that Color accepts

Return type

List[str]

Example

>>> import kwimage
>>> named_colors = kwimage.Color.named_colors()
>>> color_lut = {name: kwimage.Color(name).as01() for name in named_colors}
>>> # xdoctest: +REQUIRES(module:kwplot)
>>> import kwplot
>>> kwplot.autompl()
>>> canvas = kwplot.make_legend_img(color_lut)
>>> kwplot.imshow(canvas)

classmethod distinct(Color, num, space='rgb')

Make multiple distinct colors

classmethod random(Color, pool='named')

kwimage.atleast_3channels(arr, copy=True)

Ensures that there are 3 channels in the image

Parameters

• arr (ndarray[N, M, …]) – the image

• copy (bool) – Always copies if True, if False, then copies only when the size of the array must change.

Returns

with shape (N, M, C), where C in {3, 4}

Return type

ndarray

Doctest:

>>> assert atleast_3channels(np.zeros((10, 10))).shape[-1] == 3
>>> assert atleast_3channels(np.zeros((10, 10, 1))).shape[-1] == 3
>>> assert atleast_3channels(np.zeros((10, 10, 3))).shape[-1] == 3
>>> assert atleast_3channels(np.zeros((10, 10, 4))).shape[-1] == 4

kwimage.ensure_float01(img, dtype=np.float32, copy=True)

Ensure that an image is encoded using a float32 properly

Parameters

• img (ndarray) – an image in uint255 or float01 format. Other formats will raise errors.

• dtype (type, default=np.float32) – a numpy floating type

• copy (bool, default=False) – always copy if True, else copy if needed.

Returns

an array of floats in the range 0-1

Return type

ndarray

Raises

ValueError – if the image type is integer and not in [0-255]

Example

>>> ensure_float01(np.array([[0, .5, 1.0]]))
array([[0. , 0.5, 1. ]], dtype=float32)
>>> ensure_float01(np.array([[0, 1, 200]]))
array([[0..., 0.0039..., 0.784...]], dtype=float32)

kwimage.ensure_uint255(img, copy=True)

Ensure that an image is encoded using a uint8 properly. Either

Parameters

• img (ndarray) – an image in uint255 or float01 format. Other formats will raise errors.

• copy (bool, default=False) – always copy if True, else copy if needed.

Returns

an array of bytes in the range 0-255

Return type

ndarray

Raises

• ValueError – if the image type is float and not in [0-1]

• ValueError – if the image type is integer and not in [0-255]

Example

>>> ensure_uint255(np.array([[0, .5, 1.0]]))
array([[  0, 127, 255]], dtype=uint8)
>>> ensure_uint255(np.array([[0, 1, 200]]))
array([[  0,   1, 200]], dtype=uint8)

kwimage.make_channels_comparable(img1, img2, atleast3d=False)

Broadcasts image arrays so they can have elementwise operations applied

Parameters

• img1 (ndarray) – first image

• img2 (ndarray) – second image

• atleast3d (bool, default=False) – if true we ensure that the channel dimension exists (only relevant for 1-channel images)

Example

>>> import itertools as it
>>> wh_basis = [(5, 5), (3, 5), (5, 3), (1, 1), (1, 3), (3, 1)]
>>> for w, h in wh_basis:
>>>     shape_basis = [(w, h), (w, h, 1), (w, h, 3)]
>>>     # Test all permutations of shap inputs
>>>     for shape1, shape2 in it.product(shape_basis, shape_basis):
>>>         print('*    input shapes: %r, %r' % (shape1, shape2))
>>>         img1 = np.empty(shape1)
>>>         img2 = np.empty(shape2)
>>>         img1, img2 = make_channels_comparable(img1, img2)
>>>         print('... output shapes: %r, %r' % (img1.shape, img2.shape))
>>>         elem = (img1 + img2)
>>>         print('... elem(+) shape: %r' % (elem.shape,))
>>>         assert elem.size == img1.size, 'outputs should have same size'
>>>         assert img1.size == img2.size, 'new imgs should have same size'
>>>         print('--------')

kwimage.normalize(arr, mode='linear', alpha=None, beta=None, out=None)

Rebalance pixel intensities via contrast stretching.

By default linearly stretches pixel intensities to minimum and maximum values.

Notes

DEPRECATED: this function has been MOVED to kwarray.normalize

kwimage.normalize_intensity(imdata, return_info=False, nodata=None, axis=None, dtype=np.float32)

Normalize data intensities using heuristics to help put sensor data with extremely high or low contrast into a visible range.

This function is designed with an emphasis on getting something that is reasonable for visualization.

Parameters

• imdata (ndarray) – raw intensity data

• return_info (bool, default=False) – if True, return information about the chosen normalization heuristic.

• nodata – A value representing nodata to leave unchanged during normalization, for example 0

• dtype – can be float32 or float64

Returns

a floating point array with values between 0 and 1.

Return type

ndarray

Example

>>> from kwimage.im_core import *  # NOQA
>>> import ubelt as ub
>>> import kwimage
>>> import kwarray
>>> s = 512
>>> bit_depth = 11
>>> dtype = np.uint16
>>> max_val = int(2 ** bit_depth)
>>> min_val = int(0)
>>> rng = kwarray.ensure_rng(0)
>>> background = np.random.randint(min_val, max_val, size=(s, s), dtype=dtype)
>>> poly1 = kwimage.Polygon.random(rng=rng).scale(s / 2)
>>> poly2 = kwimage.Polygon.random(rng=rng).scale(s / 2).translate(s / 2)
>>> forground = np.zeros_like(background, dtype=np.uint8)
>>> forground = poly1.fill(forground, value=255)
>>> forground = poly2.fill(forground, value=122)
>>> forground = (kwimage.ensure_float01(forground) * max_val).astype(dtype)
>>> imdata = background + forground
>>> normed, info = normalize_intensity(imdata, return_info=True)
>>> print('info = {}'.format(ub.repr2(info, nl=1)))
>>> # xdoctest: +REQUIRES(--show)
>>> import kwplot
>>> kwplot.autompl()
>>> kwplot.imshow(imdata, pnum=(1, 2, 1), fnum=1)
>>> kwplot.imshow(normed, pnum=(1, 2, 2), fnum=1)

Example

>>> from kwimage.im_core import *  # NOQA
>>> import ubelt as ub
>>> import kwimage
>>> # Test on an image that is already normalized to test how it
>>> # degrades
>>> imdata = kwimage.grab_test_image()
>>> normed, info = normalize_intensity(imdata, return_info=True)
>>> print('info = {}'.format(ub.repr2(info, nl=1)))
>>> # xdoctest: +REQUIRES(--show)
>>> import kwplot
>>> kwplot.autompl()
>>> kwplot.imshow(imdata, pnum=(1, 2, 1), fnum=1)
>>> kwplot.imshow(normed, pnum=(1, 2, 2), fnum=1)

kwimage.num_channels(img)

Returns the number of color channels in an image.

Assumes images are 2D and the the channels are the trailing dimension. Returns 1 in the case with no trailing channel dimension, otherwise simply returns img.shape[2].

Parameters

img (ndarray) – an image with 2 or 3 dimensions.

Returns

the number of color channels (1, 3, or 4)

Return type

int

Example

>>> H = W = 3
>>> assert num_channels(np.empty((W, H))) == 1
>>> assert num_channels(np.empty((W, H, 1))) == 1
>>> assert num_channels(np.empty((W, H, 3))) == 3
>>> assert num_channels(np.empty((W, H, 4))) == 4
>>> assert num_channels(np.empty((W, H, 2))) == 2

kwimage.padded_slice(data, in_slice, pad=None, padkw=None, return_info=False)

Allows slices with out-of-bound coordinates. Any out of bounds coordinate will be sampled via padding.

DEPRECATED FOR THE VERSION IN KWARRAY (slices are more array-ish than image-ish)

Note

Negative slices have a different meaning here then they usually do. Normally, they indicate a wrap-around or a reversed stride, but here they index into out-of-bounds space (which depends on the pad mode). For example a slice of -2:1 literally samples two pixels to the left of the data and one pixel from the data, so you get two padded values and one data value.

Parameters

• data (Sliceable[T]) – data to slice into. Any channels must be the last dimension.

• in_slice (slice | Tuple[slice, …]) – slice for each dimensions

• ndim (int) – number of spatial dimensions

• pad (List[int|Tuple]) – additional padding of the slice

• padkw (Dict) – if unspecified defaults to {'mode': 'constant'}

• return_info (bool, default=False) – if True, return extra information about the transform.

SeeAlso:

_padded_slice_embed - finds the embedded slice and padding _padded_slice_apply - applies padding to sliced data

Returns

data_sliced: subregion of the input data (possibly with padding,

depending on if the original slice went out of bounds)

Tuple[Sliceable, Dict] :

data_sliced : as above

transform : information on how to return to the original coordinates

Currently a dict containing:

st_dims: a list indicating the low and high space-time

coordinate values of the returned data slice.

The structure of this dictionary mach change in the future

Return type

Sliceable

Example

>>> data = np.arange(5)
>>> in_slice = [slice(-2, 7)]

>>> data_sliced = padded_slice(data, in_slice)
>>> print(ub.repr2(data_sliced, with_dtype=False))
np.array([0, 0, 0, 1, 2, 3, 4, 0, 0])

>>> data_sliced = padded_slice(data, in_slice, pad=(3, 3))
>>> print(ub.repr2(data_sliced, with_dtype=False))
np.array([0, 0, 0, 0, 0, 0, 1, 2, 3, 4, 0, 0, 0, 0, 0])

>>> data_sliced = padded_slice(data, slice(3, 4), pad=[(1, 0)])
>>> print(ub.repr2(data_sliced, with_dtype=False))
np.array([2, 3])

kwimage.convert_colorspace(img, src_space, dst_space, copy=False, implicit=False, dst=None)

Converts colorspace of img. Convenience function around cv2.cvtColor

Parameters

• img (ndarray) – image data with float32 or uint8 precision

• src_space (str) – input image colorspace. (e.g. BGR, GRAY)

• dst_space (str) – desired output colorspace. (e.g. RGB, HSV, LAB)

• implicit (bool) –

  if False, the user must correctly specify if the input/output

  colorspaces contain alpha channels.

  If True and the input image has an alpha channel, we modify

  src_space and dst_space to ensure they both end with “A”.

• dst (ndarray[uint8_t, ndim=2], optional) – inplace-output array.

Returns

img - image data

Return type

ndarray

Note

Note the LAB and HSV colorspaces in float do not go into the 0-1 range.

For HSV the floating point range is:

0:360, 0:1, 0:1

For LAB the floating point range is:

0:100, -86.1875:98.234375, -107.859375:94.46875 (Note, that some extreme combinations of a and b are not valid)

Example

>>> import numpy as np
>>> convert_colorspace(np.array([[[0, 0, 1]]], dtype=np.float32), 'RGB', 'LAB')
>>> convert_colorspace(np.array([[[0, 1, 0]]], dtype=np.float32), 'RGB', 'LAB')
>>> convert_colorspace(np.array([[[1, 0, 0]]], dtype=np.float32), 'RGB', 'LAB')
>>> convert_colorspace(np.array([[[1, 1, 1]]], dtype=np.float32), 'RGB', 'LAB')
>>> convert_colorspace(np.array([[[0, 0, 1]]], dtype=np.float32), 'RGB', 'HSV')

Ignore:

# Check LAB output ranges import itertools as it s = 1 _iter = it.product(range(0, 256, s), range(0, 256, s), range(0, 256, s)) minvals = np.full(3, np.inf) maxvals = np.full(3, -np.inf) for r, g, b in ub.ProgIter(_iter, total=(256 // s) ** 3):

img255 = np.array([[[r, g, b]]], dtype=np.uint8) img01 = (img255 / 255.0).astype(np.float32) lab = convert_colorspace(img01, ‘rgb’, ‘lab’) np.minimum(lab[0, 0], minvals, out=minvals) np.maximum(lab[0, 0], maxvals, out=maxvals)

print(‘minvals = {}’.format(ub.repr2(minvals, nl=0))) print(‘maxvals = {}’.format(ub.repr2(maxvals, nl=0)))

kwimage.gaussian_patch(shape=(7, 7), sigma=None)

Creates a 2D gaussian patch with a specific size and sigma

Parameters

• shape (Tuple[int, int]) – patch height and width

• sigma (float | Tuple[float, float]) – Gaussian standard deviation

References

http://docs.opencv.org/modules/imgproc/doc/filtering.html#getgaussiankernel

Todo

• [ ] Look into this C-implementation

https://kwgitlab.kitware.com/computer-vision/heatmap/blob/master/heatmap/heatmap.c

CommandLine:

xdoctest -m kwimage.im_cv2 gaussian_patch –show

Example

>>> import numpy as np
>>> shape = (88, 24)
>>> sigma = None  # 1.0
>>> gausspatch = gaussian_patch(shape, sigma)
>>> sum_ = gausspatch.sum()
>>> assert np.all(np.isclose(sum_, 1.0))
>>> # xdoc: +REQUIRES(--show)
>>> import kwplot
>>> kwplot.autompl()
>>> norm = (gausspatch - gausspatch.min()) / (gausspatch.max() - gausspatch.min())
>>> kwplot.imshow(norm)
>>> kwplot.show_if_requested()

Example

>>> import numpy as np
>>> shape = (24, 24)
>>> sigma = 3.0
>>> gausspatch = gaussian_patch(shape, sigma)
>>> sum_ = gausspatch.sum()
>>> assert np.all(np.isclose(sum_, 1.0))
>>> # xdoc: +REQUIRES(--show)
>>> import kwplot
>>> kwplot.autompl()
>>> norm = (gausspatch - gausspatch.min()) / (gausspatch.max() - gausspatch.min())
>>> kwplot.imshow(norm)
>>> kwplot.show_if_requested()

kwimage.imcrop(img, dsize, about=None, origin=None, border_value=None, interpolation='nearest')

Crop an image about a specified point, padding if necessary.

This is like PIL.Image.Image.crop with more convenient arguments, or cv2.getRectSubPix without the baked-in bilinear interpolation.

Parameters

• img (ndarray) – image to crop

• dsize (Tuple[None | int, None | int]) – the desired width and height of the new image. If a dimension is None, then it is automatically computed to preserve aspect ratio. This can be larger than the original dims; if so, the cropped image is padded with border_value.

• about (Tuple[str | int, str | int]) – the location to crop about. Mutually exclusive with origin. Defaults to top left. If ints (w,h) are provided, that will be the center of the cropped image. There are also string codes available: ‘lt’: make the top left point of the image the top left point of

  the cropped image. This is equivalent to img[:dsize[1], :dsize[0]], plus padding.

  ‘rb’: make the bottom right point of the image the bottom right

  point of the cropped image. This is equivalent to img[-dsize[1]:, -dsize[0]:], plus padding.

  ‘cc’: make the center of the image the center of the cropped image. Any combination of these codes can be used, ex. ‘lb’, ‘ct’, (‘r’, 200), …

• origin (Tuple[int, int] | None) – the origin of the crop in (x,y) order (same order as dsize/about). Mutually exclusive with about. Defaults to top left.

• border_value (Numeric | Tuple | str, default=0) – any border border_value accepted by cv2.copyMakeBorder, ex. [255, 0, 0] (blue). Default is 0.

• interpolation (str, default=’nearest’) – Can be ‘nearest’, in which case integral cropping is used. Can also be ‘linear’, in which case cv2.getRectSubPix is used.

Returns

the cropped image

Return type

ndarray

SeeAlso:

kwarray.padded_slice() - a similar function for working with

“negative slices”.

Example

>>> import kwimage
>>> import numpy as np
>>> #
>>> img = kwimage.grab_test_image('astro', dsize=(32, 32))
>>> #
>>> # regular crop
>>> new_img1 = kwimage.imcrop(img, dsize=(5,6))
>>> assert new_img1.shape == (6, 5, 3)
>>> #
>>> # padding for coords outside the image bounds
>>> new_img2 = kwimage.imcrop(img, dsize=(5,6),
>>>             origin=(-1,0), border_value=[1, 0, 0])
>>> assert np.all(new_img2[:, 0] == [1, 0, 0])
>>> #
>>> # codes for corner- and edge-centered cropping
>>> new_img3 = kwimage.imcrop(img, dsize=(5,6),
>>>             about='cb')
>>> #
>>> # special code for bilinear interpolation
>>> # with floating-point coordinates
>>> new_img4 = kwimage.imcrop(img, dsize=(5,6),
>>>             about=(5.5, 8.5), interpolation='linear')
>>> #
>>> # use with bounding boxes
>>> bbox = kwimage.Boxes.random(scale=5, rng=132).to_xywh().quantize()
>>> origin, dsize = np.split(bbox.data[0], 2)
>>> new_img5 = kwimage.imcrop(img, dsize=dsize,
>>>             origin=origin)
>>> # xdoctest: +REQUIRES(--show)
>>> import kwplot
>>> kwplot.autompl()
>>> pnum_ = kwplot.PlotNums(nSubplots=6)
>>> kwplot.imshow(img, pnum=pnum_())
>>> kwplot.imshow(new_img1, pnum=pnum_())
>>> kwplot.imshow(new_img2, pnum=pnum_())
>>> kwplot.imshow(new_img3, pnum=pnum_())
>>> kwplot.imshow(new_img4, pnum=pnum_())
>>> kwplot.imshow(new_img5, pnum=pnum_())
>>> kwplot.show_if_requested()

kwimage.imresize(img, scale=None, dsize=None, max_dim=None, min_dim=None, interpolation=None, grow_interpolation=None, letterbox=False, return_info=False, antialias=False)

Resize an image based on a scale factor, final size, or size and aspect ratio.

Slightly more general than cv2.resize, allows for specification of either a scale factor, a final size, or the final size for a particular dimension.

Parameters

• img (ndarray) – image to resize

• scale (float or Tuple[float, float]) – Desired floating point scale factor. If a tuple, the dimension ordering is x,y. Mutually exclusive with dsize, max_dim, and min_dim.

• dsize (Tuple[int] | None) – The desired with and height of the new image. If a dimension is None, then it is automatically computed to preserve aspect ratio. Mutually exclusive with size, max_dim, and min_dim.

• max_dim (int) – New size of the maximum dimension, the other dimension is scaled to maintain aspect ratio. Mutually exclusive with size, dsize, and min_dim.

• min_dim (int) – New size of the minimum dimension, the other dimension is scaled to maintain aspect ratio.Mutually exclusive with size, dsize, and max_dim.

• interpolation (str | int) – The interpolation key or code (e.g. linear lanczos). By default “area” is used if the image is shrinking and “lanczos” is used if the image is growing. Note, if this is explicitly set, then it will be used regardless of if the image is growing or shrinking. Set grow_interpolation to change the default for an enlarging interpolation.

• grow_interpolation (str | int, default=”lanczos”) – The interpolation key or code to use when the image is being enlarged. Does nothing if “interpolation” is explicitly given. If “interpolation” is not specified “area” is used when shrinking.

• letterbox (bool, default=False) – If used in conjunction with dsize, then the image is scaled and translated to fit in the center of the new image while maintaining aspect ratio. Zero padding is added if necessary.

• return_info (bool, default=False) – if True returns information about the final transformation in a dictionary. If there is an offset, the scale is applied before the offset when transforming to the new resized space.

• antialias (bool, default=False) – if True blurs to anti-alias before downsampling.

Returns

the new image and optionally an info dictionary if return_info=True

Return type

ndarray | Tuple[ndarray, Dict]

Example

>>> import kwimage
>>> import numpy as np
>>> # Test scale
>>> img = np.zeros((16, 10, 3), dtype=np.uint8)
>>> new_img, info = kwimage.imresize(img, scale=.85,
>>>                                  interpolation='area',
>>>                                  return_info=True)
>>> print('info = {!r}'.format(info))
>>> assert info['scale'].tolist() == [.8, 0.875]
>>> # Test dsize without None
>>> new_img, info = kwimage.imresize(img, dsize=(5, 12),
>>>                                  interpolation='area',
>>>                                  return_info=True)
>>> print('info = {!r}'.format(info))
>>> assert info['scale'].tolist() == [0.5 , 0.75]
>>> # Test dsize with None
>>> new_img, info = kwimage.imresize(img, dsize=(6, None),
>>>                                  interpolation='area',
>>>                                  return_info=True)
>>> print('info = {!r}'.format(info))
>>> assert info['scale'].tolist() == [0.6, 0.625]
>>> # Test max_dim
>>> new_img, info = kwimage.imresize(img, max_dim=6,
>>>                                  interpolation='area',
>>>                                  return_info=True)
>>> print('info = {!r}'.format(info))
>>> assert info['scale'].tolist() == [0.4  , 0.375]
>>> # Test min_dim
>>> new_img, info = kwimage.imresize(img, min_dim=6,
>>>                                  interpolation='area',
>>>                                  return_info=True)
>>> print('info = {!r}'.format(info))
>>> assert info['scale'].tolist() == [0.6  , 0.625]

Example

>>> import kwimage
>>> import numpy as np
>>> # Test letterbox resize
>>> img = np.ones((5, 10, 3), dtype=np.float32)
>>> new_img, info = kwimage.imresize(img, dsize=(19, 19),
>>>                                  letterbox=True,
>>>                                  return_info=True)
>>> print('info = {!r}'.format(info))
>>> assert info['offset'].tolist() == [0, 4]
>>> img = np.ones((10, 5, 3), dtype=np.float32)
>>> new_img, info = kwimage.imresize(img, dsize=(19, 19),
>>>                                  letterbox=True,
>>>                                  return_info=True)
>>> print('info = {!r}'.format(info))
>>> assert info['offset'].tolist() == [4, 0]

>>> import kwimage
>>> import numpy as np
>>> # Test letterbox resize
>>> img = np.random.rand(100, 200)
>>> new_img, info = kwimage.imresize(img, dsize=(300, 300), letterbox=True, return_info=True)

Example

>>> # Check aliasing
>>> import kwimage
>>> img = kwimage.grab_test_image('checkerboard')
>>> img = kwimage.grab_test_image('astro')
>>> # xdoctest: +REQUIRES(--show)
>>> import kwplot
>>> kwplot.autompl()
>>> dsize = (14, 14)
>>> dsize = (64, 64)
>>> # When we set "grow_interpolation" for a "shrinking" resize it should
>>> # still do the "area" interpolation to antialias the results. But if we
>>> # use explicit interpolation it should alias.
>>> pnum_ = kwplot.PlotNums(nSubplots=12, nCols=4)
>>> kwplot.imshow(kwimage.imresize(img, dsize=dsize, antialias=True,  interpolation='area'), pnum=pnum_(), title='resize aa area')
>>> kwplot.imshow(kwimage.imresize(img, dsize=dsize, antialias=True, interpolation='linear'), pnum=pnum_(), title='resize aa linear')
>>> kwplot.imshow(kwimage.imresize(img, dsize=dsize, antialias=True, interpolation='nearest'), pnum=pnum_(), title='resize aa nearest')
>>> kwplot.imshow(kwimage.imresize(img, dsize=dsize, antialias=True, interpolation='cubic'), pnum=pnum_(), title='resize aa cubic')

>>> kwplot.imshow(kwimage.imresize(img, dsize=dsize, antialias=True, grow_interpolation='area'), pnum=pnum_(), title='resize aa grow area')
>>> kwplot.imshow(kwimage.imresize(img, dsize=dsize, antialias=True, grow_interpolation='linear'), pnum=pnum_(), title='resize aa grow linear')
>>> kwplot.imshow(kwimage.imresize(img, dsize=dsize, antialias=True, grow_interpolation='nearest'), pnum=pnum_(), title='resize aa grow nearest')
>>> kwplot.imshow(kwimage.imresize(img, dsize=dsize, antialias=True, grow_interpolation='cubic'), pnum=pnum_(), title='resize aa grow cubic')

>>> kwplot.imshow(kwimage.imresize(img, dsize=dsize, antialias=False, interpolation='area'), pnum=pnum_(), title='resize no-aa area')
>>> kwplot.imshow(kwimage.imresize(img, dsize=dsize, antialias=False, interpolation='linear'), pnum=pnum_(), title='resize no-aa linear')
>>> kwplot.imshow(kwimage.imresize(img, dsize=dsize, antialias=False, interpolation='nearest'), pnum=pnum_(), title='resize no-aa nearest')
>>> kwplot.imshow(kwimage.imresize(img, dsize=dsize, antialias=False, interpolation='cubic'), pnum=pnum_(), title='resize no-aa cubic')

Todo

• [X] When interpolation is area and the number of channels > 4

  cv2.resize will error but it is fine for linear interpolation

• [ ] TODO: add padding options when letterbox=True

kwimage.imscale(img, scale, interpolation=None, return_scale=False)

DEPRECATED and removed: use imresize instead

kwimage.warp_affine(image, transform, dsize=None, antialias=False, interpolation='linear', border_mode=None, border_value=0, large_warp_dim=None, return_info=False)

Applies an affine transformation to an image with optional antialiasing.

Parameters

• image (ndarray) – the input image as a numpy array. Note: this is passed directly to cv2, so it is best to ensure that it is contiguous and using a dtype that cv2 can handle.

• transform (ndarray | Affine) – a coercable affine matrix. See kwimage.Affine for details on what can be coerced.

• dsize (Tuple[int, int] | None | str, default=None) – A integer width and height tuple of the resulting “canvas” image. If None, then the input image size is used.

  If specified as a string, dsize is computed based on the given heuristic.

  If ‘positive’ (or ‘auto’), dsize is computed such that the positive coordinates of the warped image will fit in the new canvas. In this case, any pixel that maps to a negative coordinate will be clipped. This has the property that the input transformation is not modified.

  If ‘content’ (or ‘max’), the transform is modified with an extra translation such that both the positive and negative coordinates of the warped image will fit in the new canvas.

• antialias (bool, default=False) – if True determines if the transform is downsampling and applies antialiasing via gaussian a blur.

• interpolation (str, default=”linear”) – interpolation code or cv2 integer. Interpolation codes are linear, nearest, cubic, lancsoz, and area.

• border_mode (str) – Border code or cv2 integer. Border codes are constant replicate, reflect, wrap, reflect101, and transparent.

• border_value (int | float) – Used as the fill value if border_mode is constant. Otherwise this is ignored.

• large_warp_dim (int | None | str, default=None) – If specified, perform the warp piecewise in chunks of the specified size. If “auto”, it is set to the maximum “short” value in numpy. This works around a limitation of cv2.warpAffine, which must have image dimensions < SHRT_MAX (=32767 in version 4.5.3)

• return_info (bool, default=Fasle) – if True, returns information about the operation. In the case where dsize=”content”, this includes the modified transformation.

Returns

the warped image, or if return info is True, the warped image and the info dictionary.

Return type

ndarray | Tuple[ndarray, Dict]

Example

>>> from kwimage.im_cv2 import *  # NOQA
>>> import kwimage
>>> from kwimage.transform import Affine
>>> image = kwimage.grab_test_image('astro')
>>> #image = kwimage.grab_test_image('checkerboard')
>>> transform = Affine.random() @ Affine.scale(0.05)
>>> transform = Affine.scale(0.02)
>>> warped1 = warp_affine(image, transform, dsize='positive', antialias=1, interpolation='nearest')
>>> warped2 = warp_affine(image, transform, dsize='positive', antialias=0)
>>> # xdoctest: +REQUIRES(--show)
>>> import kwplot
>>> kwplot.autompl()
>>> pnum_ = kwplot.PlotNums(nRows=1, nCols=2)
>>> kwplot.imshow(warped1, pnum=pnum_(), title='antialias=True')
>>> kwplot.imshow(warped2, pnum=pnum_(), title='antialias=False')
>>> kwplot.show_if_requested()

Example

>>> from kwimage.im_cv2 import *  # NOQA
>>> import kwimage
>>> from kwimage.transform import Affine
>>> image = kwimage.grab_test_image('astro')
>>> image = kwimage.grab_test_image('checkerboard')
>>> transform = Affine.random() @ Affine.scale((.1, 1.2))
>>> warped1 = warp_affine(image, transform, dsize='positive', antialias=1)
>>> warped2 = warp_affine(image, transform, dsize='positive', antialias=0)
>>> # xdoctest: +REQUIRES(--show)
>>> import kwplot
>>> kwplot.autompl()
>>> pnum_ = kwplot.PlotNums(nRows=1, nCols=2)
>>> kwplot.imshow(warped1, pnum=pnum_(), title='antialias=True')
>>> kwplot.imshow(warped2, pnum=pnum_(), title='antialias=False')
>>> kwplot.show_if_requested()

Example

>>> # Test the case where the input data is empty or the target canvas
>>> # is empty, this should be handled like boundary effects
>>> import kwimage
>>> image = np.random.rand(1, 1, 3)
>>> transform = kwimage.Affine.random()
>>> result = kwimage.warp_affine(image, transform, dsize=(0, 0))
>>> assert result.shape == (0, 0, 3)
>>> #
>>> empty_image = np.random.rand(0, 1, 3)
>>> result = kwimage.warp_affine(empty_image, transform, dsize=(10, 10))
>>> assert result.shape == (10, 10, 3)
>>> #
>>> empty_image = np.random.rand(0, 1, 3)
>>> result = kwimage.warp_affine(empty_image, transform, dsize=(10, 0))
>>> assert result.shape == (0, 10, 3)

Example

>>> # Demo difference between positive and content dsize
>>> from kwimage.im_cv2 import *  # NOQA
>>> import kwimage
>>> from kwimage.transform import Affine
>>> image = kwimage.grab_test_image('astro', dsize=(512, 512))
>>> transform = Affine.coerce(offset=(-100, -50), scale=2, theta=0.1)
>>> # When warping other images or geometry along with this image
>>> # it is important to account for the modified transform when
>>> # setting dsize='content'. If dsize='positive', the transform
>>> # will remain unchanged wrt other aligned images / geometries.
>>> poly = kwimage.Boxes([[350, 5, 130, 290]], 'xywh').to_polygons()[0]
>>> # Apply the warping to the images
>>> warped_pos, info_pos = warp_affine(image, transform, dsize='positive', return_info=True)
>>> warped_con, info_con = warp_affine(image, transform, dsize='content', return_info=True)
>>> assert info_pos['dsize'] == (919, 1072)
>>> assert info_con['dsize'] == (1122, 1122)
>>> assert info_pos['transform'] == transform
>>> # Demo the correct and incorrect way to apply transforms
>>> poly_pos = poly.warp(transform)
>>> poly_con = poly.warp(info_con['transform'])
>>> # xdoctest: +REQUIRES(--show)
>>> import kwplot
>>> kwplot.autompl()
>>> # show original
>>> kwplot.imshow(image, pnum=(1, 3, 1), title='original')
>>> poly.draw(color='green', alpha=0.5, border=True)
>>> # show positive warped
>>> kwplot.imshow(warped_pos, pnum=(1, 3, 2), title='dsize=positive')
>>> poly_pos.draw(color='purple', alpha=0.5, border=True)
>>> # show content warped
>>> ax = kwplot.imshow(warped_con, pnum=(1, 3, 3), title='dsize=content')[1]
>>> poly_con.draw(color='dodgerblue', alpha=0.5, border=True)   # correct
>>> poly_pos.draw(color='orangered', alpha=0.5, border=True)  # incorrect
>>> cc = poly_con.to_shapely().centroid
>>> cp = poly_pos.to_shapely().centroid
>>> ax.text(cc.x, cc.y + 250, 'correctly transformed', color='dodgerblue',
>>>         backgroundcolor=(0, 0, 0, 0.7), horizontalalignment='center')
>>> ax.text(cp.x, cp.y - 250, 'incorrectly transformed', color='orangered',
>>>         backgroundcolor=(0, 0, 0, 0.7), horizontalalignment='center')
>>> kwplot.show_if_requested()

Example

>>> # Demo piecewise transform
>>> from kwimage.im_cv2 import *  # NOQA
>>> import kwimage
>>> from kwimage.transform import Affine
>>> image = kwimage.grab_test_image('astro', dsize=(512, 512))
>>> transform = Affine.coerce(offset=(-100, -50), scale=2, theta=0.1)
>>> warped_piecewise, info = warp_affine(image, transform, dsize='positive', return_info=True, large_warp_dim=32)
>>> warped_normal, info = warp_affine(image, transform, dsize='positive', return_info=True, large_warp_dim=None)
>>> # xdoctest: +REQUIRES(--show)
>>> import kwplot
>>> kwplot.autompl()
>>> kwplot.imshow(image, pnum=(1, 3, 1), title='original')
>>> kwplot.imshow(warped_normal, pnum=(1, 3, 2), title='normal warp')
>>> kwplot.imshow(warped_piecewise, pnum=(1, 3, 3), title='piecewise warp')

kwimage.checkerboard(num_squares=8, dsize=(512, 512))

Creates a checkerboard image

Parameters

• num_squares (int) – number of squares in a row

• dsize (Tuple[int, int]) – width and height

References

https://stackoverflow.com/questions/2169478/how-to-make-a-checkerboard-in-numpy

Example

>>> from kwimage.im_demodata import *  # NOQA
>>> img = checkerboard()

kwimage.grab_test_image(key='astro', space='rgb', dsize=None, interpolation='lanczos')

Ensures that the test image exists (this might use the network), reads it and returns the the image pixels.

Parameters

• key (str) – which test image to grab. Valid choices are: astro - an astronaught carl - Carl Sagan paraview - ParaView logo stars - picture of stars in the sky airport - SkySat image of Beijing Capital International Airport on 18 February 2018 See kwimage.grab_test_image.keys for a full list.

• space (str, default=’rgb’) – which colorspace to return in

• dsize (Tuple[int, int], default=None) – if specified resizes image to this size

Returns

the requested image

Return type

ndarray

CommandLine:

xdoctest -m kwimage.im_demodata grab_test_image

Example

>>> # xdoctest: +REQUIRES(--network)
>>> import kwimage
>>> for key in kwimage.grab_test_image.keys():
>>>     print('attempt to grab key = {!r}'.format(key))
>>>     kwimage.grab_test_image(key)
>>>     print('grabbed key = {!r}'.format(key))
>>> kwimage.grab_test_image('astro', dsize=(255, 255)).shape
(255, 255, 3)

kwimage.grab_test_image_fpath(key='astro')

Ensures that the test image exists (this might use the network) and returns the cached filepath to the requested image.

Parameters

key (str) – which test image to grab. Valid choices are: astro - an astronaught carl - Carl Sagan paraview - ParaView logo stars - picture of stars in the sky

Returns

path to the requested image

Return type

str

CommandLine:

python -c “import kwimage; print(kwimage.grab_test_image_fpath(‘airport’))”

Example

>>> # xdoctest: +REQUIRES(--network)
>>> import kwimage
>>> for key in kwimage.grab_test_image.keys():
...     print('attempt to grab key = {!r}'.format(key))
...     kwimage.grab_test_image_fpath(key)
...     print('grabbed grab key = {!r}'.format(key))

kwimage.draw_boxes_on_image(img, boxes, color='blue', thickness=1, box_format=None, colorspace='rgb')

Draws boxes on an image.

Parameters

• img (ndarray) – image to copy and draw on

• boxes (nh.util.Boxes) – boxes to draw

• colorspace (str) – string code of the input image colorspace

Example

>>> import kwimage
>>> import numpy as np
>>> img = np.zeros((10, 10, 3), dtype=np.uint8)
>>> color = 'dodgerblue'
>>> thickness = 1
>>> boxes = kwimage.Boxes([[1, 1, 8, 8]], 'ltrb')
>>> img2 = draw_boxes_on_image(img, boxes, color, thickness)
>>> assert tuple(img2[1, 1]) == (30, 144, 255)
>>> # xdoc: +REQUIRES(--show)
>>> import kwplot
>>> kwplot.autompl()  # xdoc: +SKIP
>>> kwplot.figure(doclf=True, fnum=1)
>>> kwplot.imshow(img2)

kwimage.draw_clf_on_image(im, classes, tcx=None, probs=None, pcx=None, border=1)

Draws classification label on an image.

Works best with image chips sized between 200x200 and 500x500

Parameters

• im (ndarray) – the image

• classes (Sequence | CategoryTree) – list of class names

• tcx (int, default=None) – true class index if known

• probs (ndarray) – predicted class probs for each class

• pcx (int, default=None) – predicted class index. (if None but probs is specified uses argmax of probs)

Example

>>> # xdoctest: +REQUIRES(module:torch)
>>> import torch
>>> import kwarray
>>> import kwimage
>>> rng = kwarray.ensure_rng(0)
>>> im = (rng.rand(300, 300) * 255).astype(np.uint8)
>>> classes = ['cls_a', 'cls_b', 'cls_c']
>>> tcx = 1
>>> probs = rng.rand(len(classes))
>>> probs[tcx] = 0
>>> probs = torch.FloatTensor(probs).softmax(dim=0).numpy()
>>> im1_ = kwimage.draw_clf_on_image(im, classes, tcx, probs)
>>> probs[tcx] = .9
>>> probs = torch.FloatTensor(probs).softmax(dim=0).numpy()
>>> im2_ = kwimage.draw_clf_on_image(im, classes, tcx, probs)
>>> # xdoctest: +REQUIRES(--show)
>>> import kwplot
>>> kwplot.autompl()
>>> kwplot.imshow(im1_, colorspace='rgb', pnum=(1, 2, 1), fnum=1, doclf=True)
>>> kwplot.imshow(im2_, colorspace='rgb', pnum=(1, 2, 2), fnum=1)
>>> kwplot.show_if_requested()

kwimage.draw_line_segments_on_image(img, pts1, pts2, color='blue', colorspace='rgb', thickness=1, **kwargs)

Draw line segments between pts1 and pts2 on an image.

Parameters

• pts1 (ndarray) – xy coordinates of starting points

• pts2 (ndarray) – corresponding xy coordinates of ending points

• color (str | List) – color code or a list of colors for each line segment

• colorspace (str, default=’rgb’) – colorspace of image

• thickness (int, default=1)

• lineType (int, default=cv2.LINE_AA)

Returns

the modified image (inplace if possible)

Return type

ndarray

Example

>>> from kwimage.im_draw import *  # NOQA
>>> pts1 = np.array([[2, 0], [2, 20], [2.5, 30]])
>>> pts2 = np.array([[10, 5], [30, 28], [100, 50]])
>>> img = np.ones((100, 100, 3), dtype=np.uint8) * 255
>>> color = 'blue'
>>> colorspace = 'rgb'
>>> img2 = draw_line_segments_on_image(img, pts1, pts2, thickness=2)
>>> # xdoc: +REQUIRES(--show)
>>> import kwplot
>>> kwplot.autompl()  # xdoc: +SKIP
>>> kwplot.figure(doclf=True, fnum=1)
>>> kwplot.imshow(img2)

Example

>>> import kwimage
>>> pts1 = kwimage.Points.random(10).scale(512).xy
>>> pts2 = kwimage.Points.random(10).scale(512).xy
>>> img = np.ones((512, 512, 3), dtype=np.uint8) * 255
>>> color = kwimage.Color.distinct(10)
>>> img2 = kwimage.draw_line_segments_on_image(img, pts1, pts2, color=color)
>>> # xdoc: +REQUIRES(--show)
>>> import kwplot
>>> kwplot.autompl()  # xdoc: +SKIP
>>> kwplot.figure(doclf=True, fnum=1)
>>> kwplot.imshow(img2)

kwimage.draw_text_on_image(img, text, org, return_info=False, **kwargs)

Draws multiline text on an image using opencv

Parameters

• img (ndarray | None | dict) – Generally a numpy image to draw on (inplace). Otherwise a canvas will be constructed such that the text will fit. The user may specify a dictionary with keys width and height to have more control over the constructed canvas.

• text (str) – text to draw

• org (Tuple[int, int]) – The x, y location of the text string “anchor” in the image as specified by halign and valign. For instance, If valign=’bottom’, halign=’left’, this is the bottom left corner.

• return_info (bool, default=False) – if True, also returns information about the positions the text was drawn on.

• **kwargs – color (tuple): default blue thickness (int): defaults to 2 fontFace (int): defaults to cv2.FONT_HERSHEY_SIMPLEX fontScale (float): defaults to 1.0 valign (str, default=’bottom’):

  either top, center, or bottom. NOTE: this default may change to “top” in the future.

  halign (str, default=’left’):

  either left, center, or right

  border (dict | int):

  If specified as an integer, draws a black border with that given thickness. If specified as a dictionary, draws a border with color specified parameters.

  “color”: border color, defaults to “black”. “thickness”: border thickness, defaults to 1.

Returns

the image that was drawn on

Return type

ndarray

Note

The image is modified inplace. If the image is non-contiguous then this returns a UMat instead of a ndarray, so be carefull with that.

References

https://stackoverflow.com/questions/27647424/ https://stackoverflow.com/questions/51285616/opencvs-gettextsize-and-puttext-return-wrong-size-and-chop-letters-with-low

Example

>>> import kwimage
>>> img = kwimage.grab_test_image(space='rgb')
>>> img2 = kwimage.draw_text_on_image(img.copy(), 'FOOBAR', org=(0, 0), valign='top')
>>> assert img2.shape == img.shape
>>> assert np.any(img2 != img)
>>> # xdoc: +REQUIRES(--show)
>>> import kwplot
>>> kwplot.autompl()
>>> kwplot.imshow(img2)
>>> kwplot.show_if_requested()

Example

>>> import kwimage
>>> # Test valign
>>> img = kwimage.grab_test_image(space='rgb', dsize=(500, 500))
>>> img2 = kwimage.draw_text_on_image(img, 'FOOBAR\nbazbiz\nspam', org=(0, 0), valign='top', border=2)
>>> img2 = kwimage.draw_text_on_image(img, 'FOOBAR\nbazbiz\nspam', org=(150, 0), valign='center', border=2)
>>> img2 = kwimage.draw_text_on_image(img, 'FOOBAR\nbazbiz\nspam', org=(300, 0), valign='bottom', border=2)
>>> # Test halign
>>> img2 = kwimage.draw_text_on_image(img, 'FOOBAR\nbazbiz\nspam', org=(250, 100), halign='right', border=2)
>>> img2 = kwimage.draw_text_on_image(img, 'FOOBAR\nbazbiz\nspam', org=(250, 250), halign='center', border=2)
>>> img2 = kwimage.draw_text_on_image(img, 'FOOBAR\nbazbiz\nspam', org=(250, 400), halign='left', border=2)
>>> # xdoc: +REQUIRES(--show)
>>> import kwplot
>>> kwplot.autompl()
>>> kwplot.imshow(img2)
>>> kwplot.show_if_requested()

Example

>>> # Ensure the function works with float01 or uint255 images
>>> import kwimage
>>> img = kwimage.grab_test_image(space='rgb')
>>> img = kwimage.ensure_float01(img)
>>> img2 = kwimage.draw_text_on_image(img, 'FOOBAR\nbazbiz\nspam', org=(0, 0), valign='top', border=2)

Example

>>> # Test dictionary border
>>> import kwimage
>>> img = kwimage.draw_text_on_image(None, 'hello\neveryone', org=(100, 100), valign='top', halign='center', border={'color': 'green', 'thickness': 9})
>>> #img = kwimage.draw_text_on_image(None, 'hello\neveryone', org=(0, 0), valign='top')
>>> #img = kwimage.draw_text_on_image(None, 'hello', org=(0, 60), valign='top', halign='center', border=0)
>>> # xdoc: +REQUIRES(--show)
>>> import kwplot
>>> kwplot.autompl()
>>> kwplot.imshow(img)
>>> kwplot.show_if_requested()

Example

>>> # Test dictionary image
>>> import kwimage
>>> img = kwimage.draw_text_on_image({'width': 300}, 'good\nPropogate', org=(150, 0), valign='top', halign='center', border={'color': 'green', 'thickness': 0})
>>> print('img.shape = {!r}'.format(img.shape))
>>> # xdoc: +REQUIRES(--show)
>>> import kwplot
>>> kwplot.autompl()
>>> kwplot.imshow(img)
>>> kwplot.show_if_requested()

kwimage.draw_vector_field(image, dx, dy, stride=0.02, thresh=0.0, scale=1.0, alpha=1.0, color='red', thickness=1, tipLength=0.1, line_type='aa')

Create an image representing a 2D vector field.

Parameters

• image (ndarray) – image to draw on

• dx (ndarray) – grid of vector x components

• dy (ndarray) – grid of vector y components

• stride (int | float) – sparsity of vectors, int specifies stride step in pixels, a float specifies it as a percentage.

• thresh (float) – only plot vectors with magnitude greater than thres

• scale (float) – multiply magnitude for easier visualization

• alpha (float) – alpha value for vectors. Non-vector regions receive 0 alpha (if False, no alpha channel is used)

• color (str | tuple | kwimage.Color) – RGB color of the vectors

• thickness (int, default=1) – thickness of arrows

• tipLength (float, default=0.1) – fraction of line length

• line_type (int) – either cv2.LINE_4, cv2.LINE_8, or cv2.LINE_AA

Returns

The image with vectors overlaid. If image=None, then an

rgb/a image is created and returned.

Return type

ndarray[float32]

Example

>>> import kwimage
>>> width, height = 512, 512
>>> image = kwimage.grab_test_image(dsize=(width, height))
>>> x, y = np.meshgrid(np.arange(height), np.arange(width))
>>> dx, dy = x - width / 2, y - height / 2
>>> radians = np.arctan2(dx, dy)
>>> mag = np.sqrt(dx ** 2 + dy ** 2) + 1e-3
>>> dx, dy = dx / mag, dy / mag
>>> img = kwimage.draw_vector_field(image, dx, dy, scale=10, alpha=False)
>>> # xdoctest: +REQUIRES(--show)
>>> import kwplot
>>> kwplot.autompl()
>>> kwplot.imshow(img)
>>> kwplot.show_if_requested()

kwimage.make_heatmask(probs, cmap='plasma', with_alpha=1.0, space='rgb', dsize=None)

Colorizes a single-channel intensity mask (with an alpha channel)

Parameters

• probs (ndarray) – 2D probability map with values between 0 and 1

• cmap (str) – mpl colormap

• with_alpha (float) – between 0 and 1, uses probs as the alpha multipled by this number.

• space (str) – output colorspace

• dsize (tuple) – if not None, then output is resized to W,H=dsize

SeeAlso:

kwimage.overlay_alpha_images

Example

>>> # xdoc: +REQUIRES(module:matplotlib)
>>> probs = np.tile(np.linspace(0, 1, 10), (10, 1))
>>> heatmask = make_heatmask(probs, with_alpha=0.8, dsize=(100, 100))
>>> # xdoc: +REQUIRES(--show)
>>> import kwplot
>>> kwplot.imshow(heatmask, fnum=1, doclf=True, colorspace='rgb')
>>> kwplot.show_if_requested()

kwimage.make_orimask(radians, mag=None, alpha=1.0)

Makes a colormap in HSV space where the orientation changes color and mag changes the saturation/value.

Parameters

• radians (ndarray) – orientation in radians

• mag (ndarray) – magnitude (must be normalized between 0 and 1)

• alpha (float | ndarray) – if False or None, then the image is returned without alpha if a float, then mag is scaled by this and used as the alpha channel if an ndarray, then this is explicilty set as the alpha channel

Returns

an rgb / rgba image in 01 space

Return type

ndarray[float32]

SeeAlso:

kwimage.overlay_alpha_images

Example

>>> # xdoc: +REQUIRES(module:matplotlib)
>>> x, y = np.meshgrid(np.arange(64), np.arange(64))
>>> dx, dy = x - 32, y - 32
>>> radians = np.arctan2(dx, dy)
>>> mag = np.sqrt(dx ** 2 + dy ** 2)
>>> orimask = make_orimask(radians, mag)
>>> # xdoc: +REQUIRES(--show)
>>> import kwplot
>>> kwplot.imshow(orimask, fnum=1, doclf=True, colorspace='rgb')
>>> kwplot.show_if_requested()

kwimage.make_vector_field(dx, dy, stride=0.02, thresh=0.0, scale=1.0, alpha=1.0, color='red', thickness=1, tipLength=0.1, line_type='aa')

Create an image representing a 2D vector field.

Parameters

• dx (ndarray) – grid of vector x components

• dy (ndarray) – grid of vector y components

• stride (int | float) – sparsity of vectors, int specifies stride step in pixels, a float specifies it as a percentage.

• thresh (float) – only plot vectors with magnitude greater than thres

• scale (float) – multiply magnitude for easier visualization

• alpha (float) – alpha value for vectors. Non-vector regions receive 0 alpha (if False, no alpha channel is used)

• color (str | tuple | kwimage.Color) – RGB color of the vectors

• thickness (int, default=1) – thickness of arrows

• tipLength (float, default=0.1) – fraction of line length

• line_type (int) – either cv2.LINE_4, cv2.LINE_8, or cv2.LINE_AA

Returns

vec_img: an rgb/rgba image in 0-1 space

Return type

ndarray[float32]

SeeAlso:

kwimage.overlay_alpha_images

DEPRECATED USE: draw_vector_field instead

Example

>>> x, y = np.meshgrid(np.arange(512), np.arange(512))
>>> dx, dy = x - 256.01, y - 256.01
>>> radians = np.arctan2(dx, dy)
>>> mag = np.sqrt(dx ** 2 + dy ** 2)
>>> dx, dy = dx / mag, dy / mag
>>> img = make_vector_field(dx, dy, scale=10, alpha=False)
>>> # xdoctest: +REQUIRES(--show)
>>> import kwplot
>>> kwplot.autompl()
>>> kwplot.imshow(img)
>>> kwplot.show_if_requested()

kwimage.fourier_mask(img_hwc, mask, axis=None, clip=None)

Applies a mask to the fourier spectrum of an image

Parameters

• img_hwc (ndarray) – assumed to be float 01

• mask (ndarray) – mask used to modulate the image in the fourier domain. Usually these are boolean values (hence the name mask), but any numerical value is technically allowed.

CommandLine:

xdoctest -m kwimage.im_filter fourier_mask –show

Example

>>> from kwimage.im_filter import *  # NOQA
>>> import kwimage
>>> img_hwc = kwimage.grab_test_image(space='gray')
>>> mask = np.random.rand(*img_hwc.shape[0:2])
>>> out_hwc = fourier_mask(img_hwc, mask)
>>> # xdoc: REQUIRES(--show)
>>> import kwplot
>>> kwplot.autompl()
>>> kwplot.imshow(img_hwc, pnum=(1, 2, 1), fnum=1)
>>> kwplot.imshow(out_hwc, pnum=(1, 2, 2), fnum=1)
>>> kwplot.show_if_requested()

kwimage.radial_fourier_mask(img_hwc, radius=11, axis=None, clip=None)

In [1] they use a radius of 11.0 on CIFAR-10.

Parameters

img_hwc (ndarray) – assumed to be float 01

References

[1] Jo and Bengio “Measuring the tendency of CNNs to Learn Surface Statistical Regularities” 2017. https://docs.opencv.org/3.0-beta/doc/py_tutorials/py_imgproc/py_transforms/py_fourier_transform/py_fourier_transform.html

Example

>>> from kwimage.im_filter import *  # NOQA
>>> import kwimage
>>> img_hwc = kwimage.grab_test_image()
>>> img_hwc = kwimage.ensure_float01(img_hwc)
>>> out_hwc = radial_fourier_mask(img_hwc, radius=11)
>>> # xdoc: REQUIRES(--show)
>>> import kwplot
>>> plt = kwplot.autoplt()
>>> def keepdim(func):
>>>     def _wrap(im):
>>>         needs_transpose = (im.shape[0] == 3)
>>>         if needs_transpose:
>>>             im = im.transpose(1, 2, 0)
>>>         out = func(im)
>>>         if needs_transpose:
>>>             out = out.transpose(2, 0, 1)
>>>         return out
>>>     return _wrap
>>> @keepdim
>>> def rgb_to_lab(im):
>>>     return kwimage.convert_colorspace(im, src_space='rgb', dst_space='lab')
>>> @keepdim
>>> def lab_to_rgb(im):
>>>     return kwimage.convert_colorspace(im, src_space='lab', dst_space='rgb')
>>> @keepdim
>>> def rgb_to_yuv(im):
>>>     return kwimage.convert_colorspace(im, src_space='rgb', dst_space='yuv')
>>> @keepdim
>>> def yuv_to_rgb(im):
>>>     return kwimage.convert_colorspace(im, src_space='yuv', dst_space='rgb')
>>> def show_data(img_hwc):
>>>     # dpath = ub.ensuredir('./fouriertest')
>>>     kwplot.imshow(img_hwc, fnum=1)
>>>     pnum_ = kwplot.PlotNums(nRows=4, nCols=5)
>>>     for r in range(0, 17):
>>>         imgt = radial_fourier_mask(img_hwc, r, clip=(0, 1))
>>>         kwplot.imshow(imgt, pnum=pnum_(), fnum=2)
>>>         plt.gca().set_title('r = {}'.format(r))
>>>     kwplot.set_figtitle('RGB')
>>>     # plt.gcf().savefig(join(dpath, '{}_{:08d}.png'.format('rgb', x)))
>>>     pnum_ = kwplot.PlotNums(nRows=4, nCols=5)
>>>     for r in range(0, 17):
>>>         imgt = lab_to_rgb(radial_fourier_mask(rgb_to_lab(img_hwc), r))
>>>         kwplot.imshow(imgt, pnum=pnum_(), fnum=3)
>>>         plt.gca().set_title('r = {}'.format(r))
>>>     kwplot.set_figtitle('LAB')
>>>     # plt.gcf().savefig(join(dpath, '{}_{:08d}.png'.format('lab', x)))
>>>     pnum_ = kwplot.PlotNums(nRows=4, nCols=5)
>>>     for r in range(0, 17):
>>>         imgt = yuv_to_rgb(radial_fourier_mask(rgb_to_yuv(img_hwc), r))
>>>         kwplot.imshow(imgt, pnum=pnum_(), fnum=4)
>>>         plt.gca().set_title('r = {}'.format(r))
>>>     kwplot.set_figtitle('YUV')
>>>     # plt.gcf().savefig(join(dpath, '{}_{:08d}.png'.format('yuv', x)))
>>> show_data(img_hwc)
>>> kwplot.show_if_requested()

kwimage.imread(fpath, space='auto', backend='auto')

Reads image data in a specified format using some backend implementation.

Parameters

• fpath (str) – path to the file to be read

• space (str, default=’auto’) – The desired colorspace of the image. Can by any colorspace accepted by convert_colorspace, or it can be ‘auto’, in which case the colorspace of the image is unmodified (except in the case where a color image is read by opencv, in which case we convert BGR to RGB by default). If None, then no modification is made to whatever backend is used to read the image.

  New in version 0.7.10: when the backend does not resolve to “cv2” the “auto” space resolves to None, thus the image is read as-is.

• backend (str, default=’auto’) – which backend reader to use. By default the file extension is used to determine this, but it can be manually overridden. Valid backends are ‘gdal’, ‘skimage’, ‘itk’, and ‘cv2’.

Returns

the image data in the specified color space.

Return type

ndarray

Note

if space is something non-standard like HSV or LAB, then the file must be a normal 8-bit color image, otherwise an error will occur.

Raises

• IOError - If the image cannot be read –

• ImportError - If trying to read a nitf without gdal –

• NotImplementedError - if trying to read a corner-case image –

Example

>>> # xdoctest: +REQUIRES(--network)
>>> from kwimage.im_io import *  # NOQA
>>> import tempfile
>>> from os.path import splitext  # NOQA
>>> # Test a non-standard image, which encodes a depth map
>>> fpath = ub.grabdata(
>>>     'http://www.topcoder.com/contest/problem/UrbanMapper3D/JAX_Tile_043_DTM.tif',
>>>     hasher='sha256', hash_prefix='64522acba6f0fb7060cd4c202ed32c5163c34e63d386afdada4190cce51ff4d4')
>>> img1 = imread(fpath)
>>> # Check that write + read preserves data
>>> tmp = tempfile.NamedTemporaryFile(suffix=splitext(fpath)[1])
>>> imwrite(tmp.name, img1)
>>> img2 = imread(tmp.name)
>>> assert np.all(img2 == img1)
>>> # xdoctest: +REQUIRES(--show)
>>> import kwplot
>>> kwplot.autompl()
>>> kwplot.imshow(img1, pnum=(1, 2, 1), fnum=1, norm=True)
>>> kwplot.imshow(img2, pnum=(1, 2, 2), fnum=1, norm=True)

Example

>>> # xdoctest: +REQUIRES(--network)
>>> import tempfile
>>> img1 = imread(ub.grabdata(
>>>     'http://i.imgur.com/iXNf4Me.png', fname='ada.png', hasher='sha256',
>>>     hash_prefix='898cf2588c40baf64d6e09b6a93b4c8dcc0db26140639a365b57619e17dd1c77'))
>>> tmp_tif = tempfile.NamedTemporaryFile(suffix='.tif')
>>> tmp_png = tempfile.NamedTemporaryFile(suffix='.png')
>>> imwrite(tmp_tif.name, img1)
>>> imwrite(tmp_png.name, img1)
>>> tif_im = imread(tmp_tif.name)
>>> png_im = imread(tmp_png.name)
>>> assert np.all(tif_im == png_im)
>>> # xdoctest: +REQUIRES(--show)
>>> import kwplot
>>> kwplot.autompl()
>>> kwplot.imshow(png_im, pnum=(1, 2, 1), fnum=1)
>>> kwplot.imshow(tif_im, pnum=(1, 2, 2), fnum=1)

Example

>>> # xdoctest: +REQUIRES(--network)
>>> import tempfile
>>> tif_fpath = ub.grabdata(
>>>     'https://ghostscript.com/doc/tiff/test/images/rgb-3c-16b.tiff',
>>>     fname='pepper.tif', hasher='sha256',
>>>     hash_prefix='31ff3a1f416cb7281acfbcbb4b56ee8bb94e9f91489602ff2806e5a49abc03c0')
>>> img1 = imread(tif_fpath)
>>> tmp_tif = tempfile.NamedTemporaryFile(suffix='.tif')
>>> tmp_png = tempfile.NamedTemporaryFile(suffix='.png')
>>> imwrite(tmp_tif.name, img1)
>>> imwrite(tmp_png.name, img1)
>>> tif_im = imread(tmp_tif.name)
>>> png_im = imread(tmp_png.name)
>>> assert np.all(tif_im == png_im)
>>> # xdoctest: +REQUIRES(--show)
>>> import kwplot
>>> kwplot.autompl()
>>> kwplot.imshow(png_im / 2 ** 16, pnum=(1, 2, 1), fnum=1)
>>> kwplot.imshow(tif_im / 2 ** 16, pnum=(1, 2, 2), fnum=1)

Example

>>> # xdoctest: +REQUIRES(module:itk, --network)
>>> import kwimage
>>> import ubelt as ub
>>> # Grab an image that ITK can read
>>> fpath = ub.grabdata(
>>>     url='https://data.kitware.com/api/v1/file/606754e32fa25629b9476f9e/download',
>>>     fname='brainweb1e5a10f17Rot20Tx20.mha',
>>>     hash_prefix='08f0812591691ae24a29788ba8cd1942e91', hasher='sha512')
>>> # Read the image (this is actually a DxHxW stack of images)
>>> img1_stack = kwimage.imread(fpath)
>>> # Check that write + read preserves data
>>> import tempfile
>>> tmp_file = tempfile.NamedTemporaryFile(suffix='.mha')
>>> kwimage.imwrite(tmp_file.name, img1_stack)
>>> recon = kwimage.imread(tmp_file.name)
>>> assert not np.may_share_memory(recon, img1_stack)
>>> assert np.all(recon == img1_stack)
>>> # xdoctest: +REQUIRES(--show)
>>> import kwplot
>>> kwplot.autompl()
>>> kwplot.imshow(kwimage.stack_images_grid(recon[0::20]))
>>> kwplot.show_if_requested()

Benchmark:

>>> from kwimage.im_io import *  # NOQA
>>> import timerit
>>> import kwimage
>>> import tempfile
>>> #
>>> dsize = (1920, 1080)
>>> img1 = kwimage.grab_test_image('amazon', dsize=dsize)
>>> ti = timerit.Timerit(10, bestof=3, verbose=1, unit='us')
>>> formats = {}
>>> dpath = ub.ensure_app_cache_dir('cache')
>>> space = 'auto'
>>> formats['png'] = kwimage.imwrite(join(dpath, '.png'), img1, space=space, backend='cv2')
>>> formats['jpg'] = kwimage.imwrite(join(dpath, '.jpg'), img1, space=space, backend='cv2')
>>> formats['tif_raw'] = kwimage.imwrite(join(dpath, '.raw.tif'), img1, space=space, backend='gdal', compress='RAW')
>>> formats['tif_deflate'] = kwimage.imwrite(join(dpath, '.deflate.tif'), img1, space=space, backend='gdal', compress='DEFLATE')
>>> formats['tif_lzw'] = kwimage.imwrite(join(dpath, '.lzw.tif'), img1, space=space, backend='gdal', compress='LZW')
>>> grid = [
>>>     ('cv2', 'png'),
>>>     ('cv2', 'jpg'),
>>>     ('gdal', 'jpg'),
>>>     ('turbojpeg', 'jpg'),
>>>     ('gdal', 'tif_raw'),
>>>     ('gdal', 'tif_lzw'),
>>>     ('gdal', 'tif_deflate'),
>>>     ('skimage', 'tif_raw'),
>>> ]
>>> backend, filefmt = 'cv2', 'png'
>>> for backend, filefmt in grid:
>>>     for timer in ti.reset(f'imread-{filefmt}-{backend}'):
>>>         with timer:
>>>             kwimage.imread(formats[filefmt], space=space, backend=backend)
>>> # Test all formats in auto mode
>>> for filefmt in formats.keys():
>>>     for timer in ti.reset(f'kwimage.imread-{filefmt}-auto'):
>>>         with timer:
>>>             kwimage.imread(formats[filefmt], space=space, backend='auto')
>>> ti.measures = ub.map_vals(ub.sorted_vals, ti.measures)
>>> import netharn as nh
>>> print('ti.measures = {}'.format(nh.util.align(ub.repr2(ti.measures['min'], nl=2), ':')))
Timed best=42891.504 µs, mean=44008.439 ± 1409.2 µs for imread-png-cv2
Timed best=33146.808 µs, mean=34185.172 ± 656.3 µs for imread-jpg-cv2
Timed best=40120.306 µs, mean=41220.927 ± 1010.9 µs for imread-jpg-gdal
Timed best=30798.162 µs, mean=31573.070 ± 737.0 µs for imread-jpg-turbojpeg
Timed best=6223.170 µs, mean=6370.462 ± 150.7 µs for imread-tif_raw-gdal
Timed best=42459.404 µs, mean=46519.940 ± 5664.9 µs for imread-tif_lzw-gdal
Timed best=36271.175 µs, mean=37301.108 ± 861.1 µs for imread-tif_deflate-gdal
Timed best=5239.503 µs, mean=6566.574 ± 1086.2 µs for imread-tif_raw-skimage
ti.measures = {
    'imread-tif_raw-skimage' : 0.0052395030070329085,
    'imread-tif_raw-gdal'    : 0.006223169999429956,
    'imread-jpg-turbojpeg'   : 0.030798161998973228,
    'imread-jpg-cv2'         : 0.03314680799667258,
    'imread-tif_deflate-gdal': 0.03627117499127053,
    'imread-jpg-gdal'        : 0.040120305988239124,
    'imread-tif_lzw-gdal'    : 0.042459404008695856,
    'imread-png-cv2'         : 0.042891503995633684,
}

>>> print('ti.measures = {}'.format(nh.util.align(ub.repr2(ti.measures['mean'], nl=2), ':')))

kwimage.imwrite(fpath, image, space='auto', backend='auto', **kwargs)

Writes image data to disk.

Parameters

• fpath (PathLike) – location to save the image

• image (ndarray) – image data

• space (str | None, default=’auto’) – the colorspace of the image to save. Can by any colorspace accepted by convert_colorspace, or it can be ‘auto’, in which case we assume the input image is either RGB, RGBA or grayscale. If None, then absolutely no color modification is made and whatever backend is used writes the image as-is.

  New in version 0.7.10: when the backend does not resolve to “cv2”, the “auto” space resolves to None, thus the image is saved as-is.

• backend (str, default=’auto’) – which backend writer to use. By default the file extension is used to determine this. Valid backends are ‘gdal’, ‘skimage’, ‘itk’, and ‘cv2’.

• **kwargs – args passed to the backend writer

Returns

path to the written file

Return type

str

Notes

The image may be modified to preserve its colorspace depending on which backend is used to write the image.

When saving as a jpeg or png, the image must be encoded with the uint8 data type. When saving as a tiff, any data type is allowed.

Raises

Exception – if the image cannot be written

Doctest:

>>> # xdoctest: +REQUIRES(--network)
>>> # This should be moved to a unit test
>>> import tempfile
>>> test_image_paths = [
>>>    ub.grabdata('https://ghostscript.com/doc/tiff/test/images/rgb-3c-16b.tiff', fname='pepper.tif'),
>>>    ub.grabdata('http://i.imgur.com/iXNf4Me.png', fname='ada.png'),
>>>    #ub.grabdata('http://www.topcoder.com/contest/problem/UrbanMapper3D/JAX_Tile_043_DTM.tif'),
>>>    ub.grabdata('https://upload.wikimedia.org/wikipedia/commons/f/fa/Grayscale_8bits_palette_sample_image.png', fname='parrot.png')
>>> ]
>>> for fpath in test_image_paths:
>>>     for space in ['auto', 'rgb', 'bgr', 'gray', 'rgba']:
>>>         img1 = imread(fpath, space=space)
>>>         print('Test im-io consistency of fpath = {!r} in {} space, shape={}'.format(fpath, space, img1.shape))
>>>         # Write the image in TIF and PNG format
>>>         tmp_tif = tempfile.NamedTemporaryFile(suffix='.tif')
>>>         tmp_png = tempfile.NamedTemporaryFile(suffix='.png')
>>>         imwrite(tmp_tif.name, img1, space=space, backend='skimage')
>>>         imwrite(tmp_png.name, img1, space=space)
>>>         tif_im = imread(tmp_tif.name, space=space)
>>>         png_im = imread(tmp_png.name, space=space)
>>>         assert np.all(tif_im == png_im), 'im-read/write inconsistency'
>>>         if _have_gdal:
>>>             tmp_tif2 = tempfile.NamedTemporaryFile(suffix='.tif')
>>>             imwrite(tmp_tif2.name, img1, space=space, backend='gdal')
>>>             tif_im2 = imread(tmp_tif2.name, space=space)
>>>             assert np.all(tif_im == tif_im2), 'im-read/write inconsistency'
>>>         if space == 'gray':
>>>             assert tif_im.ndim == 2
>>>             assert png_im.ndim == 2
>>>         elif space in ['rgb', 'bgr']:
>>>             assert tif_im.shape[2] == 3
>>>             assert png_im.shape[2] == 3
>>>         elif space in ['rgba', 'bgra']:
>>>             assert tif_im.shape[2] == 4
>>>             assert png_im.shape[2] == 4

Benchmark:

>>> import timerit
>>> import os
>>> import kwimage
>>> import tempfile
>>> #
>>> img1 = kwimage.grab_test_image('astro', dsize=(1920, 1080))
>>> space = 'auto'
>>> #
>>> file_sizes = {}
>>> #
>>> ti = timerit.Timerit(10, bestof=3, verbose=2)
>>> #
>>> for timer in ti.reset('imwrite-skimage-tif'):
>>>     with timer:
>>>         tmp = tempfile.NamedTemporaryFile(suffix='.tif')
>>>         kwimage.imwrite(tmp.name, img1, space=space, backend='skimage')
>>>     file_sizes[ti.label] = os.stat(tmp.name).st_size
>>> #
>>> for timer in ti.reset('imwrite-cv2-png'):
>>>     with timer:
>>>         tmp = tempfile.NamedTemporaryFile(suffix='.png')
>>>         kwimage.imwrite(tmp.name, img1, space=space, backend='cv2')
>>>     file_sizes[ti.label] = os.stat(tmp.name).st_size
>>> #
>>> for timer in ti.reset('imwrite-cv2-jpg'):
>>>     with timer:
>>>         tmp = tempfile.NamedTemporaryFile(suffix='.jpg')
>>>         kwimage.imwrite(tmp.name, img1, space=space, backend='cv2')
>>>     file_sizes[ti.label] = os.stat(tmp.name).st_size
>>> #
>>> for timer in ti.reset('imwrite-gdal-raw'):
>>>     with timer:
>>>         tmp = tempfile.NamedTemporaryFile(suffix='.tif')
>>>         kwimage.imwrite(tmp.name, img1, space=space, backend='gdal', compress='RAW')
>>>     file_sizes[ti.label] = os.stat(tmp.name).st_size
>>> #
>>> for timer in ti.reset('imwrite-gdal-lzw'):
>>>     with timer:
>>>         tmp = tempfile.NamedTemporaryFile(suffix='.tif')
>>>         kwimage.imwrite(tmp.name, img1, space=space, backend='gdal', compress='LZW')
>>>     file_sizes[ti.label] = os.stat(tmp.name).st_size
>>> #
>>> for timer in ti.reset('imwrite-gdal-zstd'):
>>>     with timer:
>>>         tmp = tempfile.NamedTemporaryFile(suffix='.tif')
>>>         kwimage.imwrite(tmp.name, img1, space=space, backend='gdal', compress='ZSTD')
>>>     file_sizes[ti.label] = os.stat(tmp.name).st_size
>>> #
>>> for timer in ti.reset('imwrite-gdal-deflate'):
>>>     with timer:
>>>         tmp = tempfile.NamedTemporaryFile(suffix='.tif')
>>>         kwimage.imwrite(tmp.name, img1, space=space, backend='gdal', compress='DEFLATE')
>>>     file_sizes[ti.label] = os.stat(tmp.name).st_size
>>> #
>>> for timer in ti.reset('imwrite-gdal-jpeg'):
>>>     with timer:
>>>         tmp = tempfile.NamedTemporaryFile(suffix='.tif')
>>>         kwimage.imwrite(tmp.name, img1, space=space, backend='gdal', compress='JPEG')
>>>     file_sizes[ti.label] = os.stat(tmp.name).st_size
>>> #
>>> file_sizes = ub.sorted_vals(file_sizes)
>>> import xdev
>>> file_sizes_human = ub.map_vals(lambda x: xdev.byte_str(x, 'MB'), file_sizes)
>>> print('ti.rankings = {}'.format(ub.repr2(ti.rankings, nl=2)))
>>> print('file_sizes = {}'.format(ub.repr2(file_sizes_human, nl=1)))

Example

>>> # Test saving a multi-band file
>>> import kwimage
>>> import tempfile
>>> # In this case the backend will not resolve to cv2, so
>>> # we should not need to specify space.
>>> data = np.random.rand(32, 32, 13).astype(np.float32)
>>> temp = tempfile.NamedTemporaryFile(suffix='.tif')
>>> fpath = temp.name
>>> kwimage.imwrite(fpath, data)
>>> recon = kwimage.imread(fpath)
>>> assert np.all(recon == data)

>>> kwimage.imwrite(fpath, data, backend='skimage')
>>> recon = kwimage.imread(fpath)
>>> assert np.all(recon == data)

>>> import pytest
>>> # In this case the backend will resolve to cv2, and thus we expect
>>> # a failure
>>> temp = tempfile.NamedTemporaryFile(suffix='.png')
>>> fpath = temp.name
>>> with pytest.raises(NotImplementedError):
>>>     kwimage.imwrite(fpath, data)

kwimage.load_image_shape(fpath)

Determine the height/width/channels of an image without reading the entire file.

Parameters

fpath (str) – path to an image

Returns

Tuple - shape of the dataset.

Recall this library uses the convention that “shape” is refers to height,width,channels and “size” is width,height ordering.

Benchmark:

>>> # For large files, PIL is much faster
>>> import gdal
>>> from PIL import Image
>>> #
>>> import kwimage
>>> fpath = kwimage.grab_test_image_fpath()
>>> #
>>> ti = ub.Timerit(100, bestof=10, verbose=2)
>>> for timer in ti.reset('gdal'):
>>>     with timer:
>>>         gdal_dset = gdal.Open(fpath, gdal.GA_ReadOnly)
>>>         width = gdal_dset.RasterXSize
>>>         height = gdal_dset.RasterYSize
>>>         gdal_dset = None
>>> #
>>> for timer in ti.reset('PIL'):
>>>     with timer:
>>>         pil_img = Image.open(fpath)
>>>         width, height = pil_img.size
>>>         pil_img.close()
Timed gdal for: 100 loops, best of 10
    time per loop: best=62.967 µs, mean=63.991 ± 0.8 µs
Timed PIL for: 100 loops, best of 10
    time per loop: best=46.640 µs, mean=47.314 ± 0.4 µs

kwimage.decode_run_length(counts, shape, binary=False, dtype=np.uint8, order='C')

Decode run length encoding back into an image.

Parameters

• counts (ndarray) – the run-length encoding

• shape (Tuple[int, int])

• binary (bool) – if the RLE is binary or non-binary. Set to True for compatibility with COCO.

• dtype (dtype, default=np.uint8) – data type for decoded image

• order ({‘C’, ‘F’}, default=’C’) – row-major (C) or column-major (F)

Returns

the reconstructed image

Return type

ndarray

Example

>>> from kwimage.im_runlen import *  # NOQA
>>> img = np.array([[1, 0, 1, 1, 1, 0, 0, 1, 0]])
>>> encoded = encode_run_length(img, binary=True)
>>> recon = decode_run_length(**encoded)
>>> assert np.all(recon == img)

>>> import ubelt as ub
>>> lines = ub.codeblock(
>>>     '''
>>>     ..........
>>>     ......111.
>>>     ..2...111.
>>>     .222..111.
>>>     22222.....
>>>     .222......
>>>     ..2.......
>>>     ''').replace('.', '0').splitlines()
>>> img = np.array([list(map(int, line)) for line in lines])
>>> encoded = encode_run_length(img)
>>> recon = decode_run_length(**encoded)
>>> assert np.all(recon == img)

kwimage.encode_run_length(img, binary=False, order='C')

Construct the run length encoding (RLE) of an image.

Parameters

• img (ndarray) – 2D image

• binary (bool, default=False) – If true, assume that the input image only contains 0’s and 1’s. Set to True for compatibility with COCO (which does not support multi-value RLE encodings).

• order ({‘C’, ‘F’}, default=’C’) – row-major (C) or column-major (F)

Returns

encoding: dictionary items are:

counts (ndarray): the run length encoding

shape (Tuple): the original image shape.

This should be in standard shape row-major (e.g. h/w) order

binary (bool):

if True, the counts are assumed to encode only 0’s and 1’s, otherwise the counts encoding specifies any numeric values.

order ({‘C’, ‘F’}, default=’C’): encoding order

Return type

Dict[str, object]

SeeAlso:

• kwimage.Mask - a cython-backed data structure to handle coco-style RLEs

Example

>>> import ubelt as ub
>>> lines = ub.codeblock(
>>>     '''
>>>     ..........
>>>     ......111.
>>>     ..2...111.
>>>     .222..111.
>>>     22222.....
>>>     .222......
>>>     ..2.......
>>>     ''').replace('.', '0').splitlines()
>>> img = np.array([list(map(int, line)) for line in lines])
>>> encoding = encode_run_length(img)
>>> target = np.array([0,16,1,3,0,3,2,1,0,3,1,3,0,2,2,3,0,2,1,3,0,1,2,5,0,6,2,3,0,8,2,1,0,7])
>>> assert np.all(target == encoding['counts'])

Example

>>> binary = True
>>> img = np.array([[1, 0, 1, 1, 1, 0, 0, 1, 0]])
>>> encoding = encode_run_length(img, binary=True)
>>> assert encoding['counts'].tolist() == [0, 1, 1, 3, 2, 1, 1]

kwimage.rle_translate(rle, offset, output_shape=None)

Translates a run-length encoded image in RLE-space.

Parameters

• rle (dict) – an enconding dict returned by encode_run_length

• offset (Tuple) – x,y offset, CAREFUL, this can only accept integers

• output_shape (Tuple, optional) – h,w of transformed mask. If unspecified the input rle shape is used.

SeeAlso:

# ITK has some RLE code that looks like it can perform translations https://github.com/KitwareMedical/ITKRLEImage/blob/master/include/itkRLERegionOfInterestImageFilter.h

Doctest:

>>> # test that translate works on all zero images
>>> img = np.zeros((7, 8), dtype=np.uint8)
>>> rle = encode_run_length(img, binary=True, order='F')
>>> new_rle = rle_translate(rle, (1, 2), (6, 9))
>>> assert np.all(new_rle['counts'] == [54])

Example

>>> from kwimage.im_runlen import *  # NOQA
>>> img = np.array([
>>>     [1, 1, 1, 1],
>>>     [0, 1, 0, 0],
>>>     [0, 1, 0, 1],
>>>     [1, 1, 1, 1],], dtype=np.uint8)
>>> rle = encode_run_length(img, binary=True, order='C')
>>> offset = (1, -1)
>>> output_shape = (3, 5)
>>> new_rle = rle_translate(rle, offset, output_shape)
>>> decoded = decode_run_length(**new_rle)
>>> print(decoded)
[[0 0 1 0 0]
 [0 0 1 0 1]
 [0 1 1 1 1]]

Example

>>> from kwimage.im_runlen import *  # NOQA
>>> img = np.array([
>>>     [0, 0, 0],
>>>     [0, 1, 0],
>>>     [0, 0, 0]], dtype=np.uint8)
>>> rle = encode_run_length(img, binary=True, order='C')
>>> new_rle = rle_translate(rle, (1, 0))
>>> decoded = decode_run_length(**new_rle)
>>> print(decoded)
[[0 0 0]
 [0 0 1]
 [0 0 0]]
>>> new_rle = rle_translate(rle, (0, 1))
>>> decoded = decode_run_length(**new_rle)
>>> print(decoded)
[[0 0 0]
 [0 0 0]
 [0 1 0]]

kwimage.stack_images(images, axis=0, resize=None, interpolation=None, overlap=0, return_info=False, bg_value=None)

Make a new image with the input images side-by-side

Parameters

• images (Iterable[ndarray[ndim=2]]) – image data

• axis (int) – axis to stack on (either 0 or 1)

• resize (int, str, or None) – if None image sizes are not modified, otherwise resize resize can be either 0 or 1. We resize the resize-th image to match the 1 - resize-th image. Can also be strings “larger” or “smaller”.

• interpolation (int or str) – string or cv2-style interpolation type. only used if resize or overlap > 0

• overlap (int) – number of pixels to overlap. Using a negative number results in a border.

• return_info (bool) – if True, returns transforms (scales and translations) to map from original image to its new location.

Returns

an image of stacked images side by side

OR

Tuple[ndarray, List]: where the first item is the aformentioned stacked

image and the second item is a list of transformations for each input image mapping it to its location in the returned image.

Return type

ndarray

Example

>>> import kwimage
>>> img1 = kwimage.grab_test_image('carl', space='rgb')
>>> img2 = kwimage.grab_test_image('astro', space='rgb')
>>> images = [img1, img2]
>>> imgB, transforms = stack_images(images, axis=0, resize='larger',
>>>                                 overlap=-10, return_info=True)
>>> print('imgB.shape = {}'.format(imgB.shape))
>>> # xdoctest: +REQUIRES(--show)
>>> import kwplot
>>> import kwimage
>>> kwplot.autompl()
>>> kwplot.imshow(imgB, colorspace='rgb')
>>> wh1 = np.multiply(img1.shape[0:2][::-1], transforms[0].scale)
>>> wh2 = np.multiply(img2.shape[0:2][::-1], transforms[1].scale)
>>> xoff1, yoff1 = transforms[0].translation
>>> xoff2, yoff2 = transforms[1].translation
>>> xywh1 = (xoff1, yoff1, wh1[0], wh1[1])
>>> xywh2 = (xoff2, yoff2, wh2[0], wh2[1])
>>> kwplot.draw_boxes(kwimage.Boxes([xywh1], 'xywh'), color=(1.0, 0, 0))
>>> kwplot.draw_boxes(kwimage.Boxes([xywh2], 'xywh'), color=(1.0, 0, 0))
>>> kwplot.show_if_requested()
((662, 512, 3), (0.0, 0.0), (0, 150))

kwimage.stack_images_grid(images, chunksize=None, axis=0, overlap=0, return_info=False, bg_value=None)

Stacks images in a grid. Optionally return transforms of original image positions in the output image.

Parameters

• images (Iterable[ndarray[ndim=2]]) – image data

• chunksize (int, default=None) – number of rows per column or columns per row depending on the value of axis. If unspecified, computes this as int(sqrt(len(images))).

• axis (int, default=0) – If 0, chunksize is columns per row. If 1, chunksize is rows per column.

• overlap (int) – number of pixels to overlap. Using a negative number results in a border.

• return_info (bool) – if True, returns transforms (scales and translations) to map from original image to its new location.

Returns

an image of stacked images in a grid pattern

OR

Tuple[ndarray, List]: where the first item is the aformentioned stacked

image and the second item is a list of transformations for each input image mapping it to its location in the returned image.

Return type

ndarray

class kwimage.Boxes(data, format=None, check=True)

Bases: _BoxConversionMixins, _BoxPropertyMixins, _BoxTransformMixins, _BoxDrawMixins, ubelt.NiceRepr

Converts boxes between different formats as long as the last dimension contains 4 coordinates and the format is specified.

This is a convinience class, and should not not store the data for very long. The general idiom should be create class, convert data, and then get the raw data and let the class be garbage collected. This will help ensure that your code is portable and understandable if this class is not available.

Example

>>> # xdoctest: +IGNORE_WHITESPACE
>>> import kwimage
>>> import numpy as np
>>> # Given an array / tensor that represents one or more boxes
>>> data = np.array([[ 0,  0, 10, 10],
>>>                  [ 5,  5, 50, 50],
>>>                  [20,  0, 30, 10]])
>>> # The kwimage.Boxes data structure is a thin fast wrapper
>>> # that provides methods for operating on the boxes.
>>> # It requires that the user explicitly provide a code that denotes
>>> # the format of the boxes (i.e. what each column represents)
>>> boxes = kwimage.Boxes(data, 'ltrb')
>>> # This means that there is no ambiguity about box format
>>> # The representation string of the Boxes object demonstrates this
>>> print('boxes = {!r}'.format(boxes))
boxes = <Boxes(ltrb,
    array([[ 0,  0, 10, 10],
           [ 5,  5, 50, 50],
           [20,  0, 30, 10]]))>
>>> # if you pass this data around. You can convert to other formats
>>> # For docs on available format codes see :class:`BoxFormat`.
>>> # In this example we will convert (left, top, right, bottom)
>>> # to (left-x, top-y, width, height).
>>> boxes.toformat('xywh')
<Boxes(xywh,
    array([[ 0,  0, 10, 10],
           [ 5,  5, 45, 45],
           [20,  0, 10, 10]]))>
>>> # In addition to format conversion there are other operations
>>> # We can quickly (using a C-backend) find IoUs
>>> ious = boxes.ious(boxes)
>>> print('{}'.format(ub.repr2(ious, nl=1, precision=2, with_dtype=False)))
np.array([[1.  , 0.01, 0.  ],
          [0.01, 1.  , 0.02],
          [0.  , 0.02, 1.  ]])
>>> # We can ask for the area of each box
>>> print('boxes.area = {}'.format(ub.repr2(boxes.area, nl=0, with_dtype=False)))
boxes.area = np.array([[ 100],[2025],[ 100]])
>>> # We can ask for the center of each box
>>> print('boxes.center = {}'.format(ub.repr2(boxes.center, nl=1, with_dtype=False)))
boxes.center = (
    np.array([[ 5. ],[27.5],[25. ]]),
    np.array([[ 5. ],[27.5],[ 5. ]]),
)
>>> # We can translate / scale the boxes
>>> boxes.translate((10, 10)).scale(100)
<Boxes(ltrb,
    array([[1000., 1000., 2000., 2000.],
           [1500., 1500., 6000., 6000.],
           [3000., 1000., 4000., 2000.]]))>
>>> # We can clip the bounding boxes
>>> boxes.translate((10, 10)).scale(100).clip(1200, 1200, 1700, 1800)
<Boxes(ltrb,
    array([[1200., 1200., 1700., 1800.],
           [1500., 1500., 1700., 1800.],
           [1700., 1200., 1700., 1800.]]))>
>>> # We can perform arbitrary warping of the boxes
>>> # (note that if the transform is not axis aligned, the axis aligned
>>> #  bounding box of the transform result will be returned)
>>> transform = np.array([[-0.83907153,  0.54402111,  0. ],
>>>                       [-0.54402111, -0.83907153,  0. ],
>>>                       [ 0.        ,  0.        ,  1. ]])
>>> boxes.warp(transform)
<Boxes(ltrb,
    array([[ -8.3907153 , -13.8309264 ,   5.4402111 ,   0.        ],
           [-39.23347095, -69.154632  ,  23.00569785,  -6.9154632 ],
           [-25.1721459 , -24.7113486 , -11.3412195 , -10.8804222 ]]))>
>>> # Note, that we can transform the box to a Polygon for more
>>> # accurate warping.
>>> transform = np.array([[-0.83907153,  0.54402111,  0. ],
>>>                       [-0.54402111, -0.83907153,  0. ],
>>>                       [ 0.        ,  0.        ,  1. ]])
>>> warped_polys = boxes.to_polygons().warp(transform)
>>> print(ub.repr2(warped_polys.data, sv=1))
[
    <Polygon({
        'exterior': <Coords(data=
                        array([[  0.       ,   0.       ],
                               [  5.4402111,  -8.3907153],
                               [ -2.9505042, -13.8309264],
                               [ -8.3907153,  -5.4402111],
                               [  0.       ,   0.       ]]))>,
        'interiors': [],
    })>,
    <Polygon({
        'exterior': <Coords(data=
                        array([[ -1.4752521 ,  -6.9154632 ],
                               [ 23.00569785, -44.67368205],
                               [-14.752521  , -69.154632  ],
                               [-39.23347095, -31.39641315],
                               [ -1.4752521 ,  -6.9154632 ]]))>,
        'interiors': [],
    })>,
    <Polygon({
        'exterior': <Coords(data=
                        array([[-16.7814306, -10.8804222],
                               [-11.3412195, -19.2711375],
                               [-19.7319348, -24.7113486],
                               [-25.1721459, -16.3206333],
                               [-16.7814306, -10.8804222]]))>,
        'interiors': [],
    })>,
]
>>> # The kwimage.Boxes data structure is also convertable to
>>> # several alternative data structures, like shapely, coco, and imgaug.
>>> print(ub.repr2(boxes.to_shapely(), sv=1))
[
    POLYGON ((0 0, 0 10, 10 10, 10 0, 0 0)),
    POLYGON ((5 5, 5 50, 50 50, 50 5, 5 5)),
    POLYGON ((20 0, 20 10, 30 10, 30 0, 20 0)),
]
>>> # xdoctest: +REQUIRES(module:imgaug)
>>> print(ub.repr2(boxes[0:1].to_imgaug(shape=(100, 100)), sv=1))
BoundingBoxesOnImage([BoundingBox(x1=0.0000, y1=0.0000, x2=10.0000, y2=10.0000, label=None)], shape=(100, 100))
>>> # xdoctest: -REQUIRES(module:imgaug)
>>> print(ub.repr2(list(boxes.to_coco()), sv=1))
[
    [0, 0, 10, 10],
    [5, 5, 45, 45],
    [20, 0, 10, 10],
]
>>> # Finally, when you are done with your boxes object, you can
>>> # unwrap the raw data by using the ``.data`` attribute
>>> # all operations are done on this data, which gives the
>>> # kwiamge.Boxes data structure almost no overhead when
>>> # inserted into existing code.
>>> print('boxes.data =\n{}'.format(ub.repr2(boxes.data, nl=1)))
boxes.data =
np.array([[ 0,  0, 10, 10],
          [ 5,  5, 50, 50],
          [20,  0, 30, 10]], dtype=np.int64)
>>> # xdoctest: +REQUIRES(module:torch)
>>> # This data structure was designed for use with both torch
>>> # and numpy, the underlying data can be either an array or tensor.
>>> boxes.tensor()
<Boxes(ltrb,
    tensor([[ 0,  0, 10, 10],
            [ 5,  5, 50, 50],
            [20,  0, 30, 10]]))>
>>> boxes.numpy()
<Boxes(ltrb,
    array([[ 0,  0, 10, 10],
           [ 5,  5, 50, 50],
           [20,  0, 30, 10]]))>

Example

>>> # xdoctest: +IGNORE_WHITESPACE
>>> from kwimage.structs.boxes import *  # NOQA
>>> # Demo of conversion methods
>>> import kwimage
>>> kwimage.Boxes([[25, 30, 15, 10]], 'xywh')
<Boxes(xywh, array([[25, 30, 15, 10]]))>
>>> kwimage.Boxes([[25, 30, 15, 10]], 'xywh').to_xywh()
<Boxes(xywh, array([[25, 30, 15, 10]]))>
>>> kwimage.Boxes([[25, 30, 15, 10]], 'xywh').to_cxywh()
<Boxes(cxywh, array([[32.5, 35. , 15. , 10. ]]))>
>>> kwimage.Boxes([[25, 30, 15, 10]], 'xywh').to_ltrb()
<Boxes(ltrb, array([[25, 30, 40, 40]]))>
>>> kwimage.Boxes([[25, 30, 15, 10]], 'xywh').scale(2).to_ltrb()
<Boxes(ltrb, array([[50., 60., 80., 80.]]))>
>>> # xdoctest: +REQUIRES(module:torch)
>>> kwimage.Boxes(torch.FloatTensor([[25, 30, 15, 20]]), 'xywh').scale(.1).to_ltrb()
<Boxes(ltrb, tensor([[ 2.5000,  3.0000,  4.0000,  5.0000]]))>

Notes

In the following examples we show cases where Boxes can hold a single 1-dimensional box array. This is a holdover from an older codebase, and some functions may assume that the input is at least 2-D. Thus when representing a single bounding box it is best practice to view it as a list of 1 box. While many function will work in the 1-D case, not all functions have been tested and thus we cannot gaurentee correctness.

Example

>>> # xdoctest: +IGNORE_WHITESPACE
>>> Boxes([25, 30, 15, 10], 'xywh')
<Boxes(xywh, array([25, 30, 15, 10]))>
>>> Boxes([25, 30, 15, 10], 'xywh').to_xywh()
<Boxes(xywh, array([25, 30, 15, 10]))>
>>> Boxes([25, 30, 15, 10], 'xywh').to_cxywh()
<Boxes(cxywh, array([32.5, 35. , 15. , 10. ]))>
>>> Boxes([25, 30, 15, 10], 'xywh').to_ltrb()
<Boxes(ltrb, array([25, 30, 40, 40]))>
>>> Boxes([25, 30, 15, 10], 'xywh').scale(2).to_ltrb()
<Boxes(ltrb, array([50., 60., 80., 80.]))>
>>> # xdoctest: +REQUIRES(module:torch)
>>> Boxes(torch.FloatTensor([[25, 30, 15, 20]]), 'xywh').scale(.1).to_ltrb()
<Boxes(ltrb, tensor([[ 2.5000,  3.0000,  4.0000,  5.0000]]))>

Example

>>> datas = [
>>>     [1, 2, 3, 4],
>>>     [[1, 2, 3, 4], [4, 5, 6, 7]],
>>>     [[[1, 2, 3, 4], [4, 5, 6, 7]]],
>>> ]
>>> formats = BoxFormat.cannonical
>>> for format1 in formats:
>>>     for data in datas:
>>>         self = box1 = Boxes(data, format1)
>>>         for format2 in formats:
>>>             box2 = box1.toformat(format2)
>>>             back = box2.toformat(format1)
>>>             assert box1 == back

__getitem__(self, index)

__eq__(self, other)

Tests equality of two Boxes objects

Example

>>> box0 = box1 = Boxes([[1, 2, 3, 4]], 'xywh')
>>> box2 = Boxes(box0.data, 'ltrb')
>>> box3 = Boxes([[0, 2, 3, 4]], box0.format)
>>> box4 = Boxes(box0.data, box2.format)
>>> assert box0 == box1
>>> assert not box0 == box2
>>> assert not box2 == box3
>>> assert box2 == box4

__len__(self)

__nice__(self)

__repr__(self)

Return repr(self).

classmethod random(Boxes, num=1, scale=1.0, format=BoxFormat.XYWH, anchors=None, anchor_std=1.0 / 6, tensor=False, rng=None)

Makes random boxes; typically for testing purposes

Parameters

• num (int) – number of boxes to generate

• scale (float | Tuple[float, float]) – size of imgdims

• format (str) – format of boxes to be created (e.g. ltrb, xywh)

• anchors (ndarray) – normalized width / heights of anchor boxes to perterb and randomly place. (must be in range 0-1)

• anchor_std (float) – magnitude of noise applied to anchor shapes

• tensor (bool) – if True, returns boxes in tensor format

• rng (None | int | RandomState) – initial random seed

Example

>>> # xdoctest: +IGNORE_WHITESPACE
>>> Boxes.random(3, rng=0, scale=100)
<Boxes(xywh,
    array([[54, 54,  6, 17],
           [42, 64,  1, 25],
           [79, 38, 17, 14]]))>
>>> # xdoctest: +REQUIRES(module:torch)
>>> Boxes.random(3, rng=0, scale=100).tensor()
<Boxes(xywh,
    tensor([[ 54,  54,   6,  17],
            [ 42,  64,   1,  25],
            [ 79,  38,  17,  14]]))>
>>> anchors = np.array([[.5, .5], [.3, .3]])
>>> Boxes.random(3, rng=0, scale=100, anchors=anchors)
<Boxes(xywh,
    array([[ 2, 13, 51, 51],
           [32, 51, 32, 36],
           [36, 28, 23, 26]]))>

Example

>>> # Boxes position/shape within 0-1 space should be uniform.
>>> # xdoctest: +REQUIRES(--show)
>>> import kwplot
>>> kwplot.autompl()
>>> fig = kwplot.figure(fnum=1, doclf=True)
>>> fig.gca().set_xlim(0, 128)
>>> fig.gca().set_ylim(0, 128)
>>> import kwimage
>>> kwimage.Boxes.random(num=10).scale(128).draw()

copy(self)

classmethod concatenate(cls, boxes, axis=0)

Concatenates multiple boxes together

Parameters

• boxes (Sequence[Boxes]) – list of boxes to concatenate

• axis (int, default=0) – axis to stack on

Returns

stacked boxes

Return type

Boxes

Example

>>> boxes = [Boxes.random(3) for _ in range(3)]
>>> new = Boxes.concatenate(boxes)
>>> assert len(new) == 9
>>> assert np.all(new.data[3:6] == boxes[1].data)

Example

>>> boxes = [Boxes.random(3) for _ in range(3)]
>>> boxes[0].data = boxes[0].data[0]
>>> boxes[1].data = boxes[0].data[0:0]
>>> new = Boxes.concatenate(boxes)
>>> assert len(new) == 4
>>> # xdoctest: +REQUIRES(module:torch)
>>> new = Boxes.concatenate([b.tensor() for b in boxes])
>>> assert len(new) == 4

compress(self, flags, axis=0, inplace=False)

Filters boxes based on a boolean criterion

Parameters

• flags (ArrayLike[bool]) – true for items to be kept

• axis (int) – you usually want this to be 0

• inplace (bool) – if True, modifies this object

Example

>>> self = Boxes([[25, 30, 15, 10]], 'ltrb')
>>> self.compress([True])
<Boxes(ltrb, array([[25, 30, 15, 10]]))>
>>> self.compress([False])
<Boxes(ltrb, array([], shape=(0, 4), dtype=int64))>

take(self, idxs, axis=0, inplace=False)

Takes a subset of items at specific indices

Parameters

• indices (ArrayLike[int]) – indexes of items to take

• axis (int) – you usually want this to be 0

• inplace (bool) – if True, modifies this object

Example

>>> self = Boxes([[25, 30, 15, 10]], 'ltrb')
>>> self.take([0])
<Boxes(ltrb, array([[25, 30, 15, 10]]))>
>>> self.take([])
<Boxes(ltrb, array([], shape=(0, 4), dtype=int64))>

is_tensor(self)

is the backend fueled by torch?

is_numpy(self)

is the backend fueled by numpy?

_impl(self)

returns the kwarray.ArrayAPI implementation for the data

Example

>>> assert Boxes.random().numpy()._impl.is_numpy
>>> # xdoctest: +REQUIRES(module:torch)
>>> assert Boxes.random().tensor()._impl.is_tensor

property device(self)

If the backend is torch returns the data device, otherwise None

astype(self, dtype)

Changes the type of the internal array used to represent the boxes

Notes

this operation is not inplace

Example

>>> # xdoctest: +IGNORE_WHITESPACE
>>> # xdoctest: +REQUIRES(module:torch)
>>> Boxes.random(3, 100, rng=0).tensor().astype('int32')
<Boxes(xywh,
    tensor([[54, 54,  6, 17],
            [42, 64,  1, 25],
            [79, 38, 17, 14]], dtype=torch.int32))>
>>> Boxes.random(3, 100, rng=0).numpy().astype('int32')
<Boxes(xywh,
    array([[54, 54,  6, 17],
           [42, 64,  1, 25],
           [79, 38, 17, 14]], dtype=int32))>
>>> Boxes.random(3, 100, rng=0).tensor().astype('float32')
>>> Boxes.random(3, 100, rng=0).numpy().astype('float32')

round(self, inplace=False)

Rounds data coordinates to the nearest integer.

This operation is applied directly to the box coordinates, so its output will depend on the format the boxes are stored in.

Parameters

inplace (bool, default=False) – if True, modifies this object

SeeAlso:

:method:`Boxes.quantize`

Example

>>> import kwimage
>>> self = kwimage.Boxes.random(3, rng=0).scale(10)
>>> new = self.round()
>>> print('self = {!r}'.format(self))
>>> print('new = {!r}'.format(new))
self = <Boxes(xywh,
    array([[5.48813522, 5.44883192, 0.53949833, 1.70306146],
           [4.23654795, 6.4589411 , 0.13932407, 2.45878875],
           [7.91725039, 3.83441508, 1.71937704, 1.45453393]]))>
new = <Boxes(xywh,
    array([[5., 5., 1., 2.],
           [4., 6., 0., 2.],
           [8., 4., 2., 1.]]))>

quantize(self, inplace=False, dtype=np.int32)

Converts the box to integer coordinates.

This operation takes the floor of the left side and the ceil of the right side. Thus the area of the box will never decreases.

Parameters

• inplace (bool, default=False) – if True, modifies this object

• dtype (type) – type to cast as

SeeAlso:

:method:`Boxes.round`

Example

>>> import kwimage
>>> self = kwimage.Boxes.random(3, rng=0).scale(10)
>>> new = self.quantize()
>>> print('self = {!r}'.format(self))
>>> print('new = {!r}'.format(new))
self = <Boxes(xywh,
    array([[5.48813522, 5.44883192, 0.53949833, 1.70306146],
           [4.23654795, 6.4589411 , 0.13932407, 2.45878875],
           [7.91725039, 3.83441508, 1.71937704, 1.45453393]]))>
new = <Boxes(xywh,
    array([[5, 5, 2, 3],
           [4, 6, 1, 3],
           [7, 3, 3, 3]], dtype=int32))>

Example

>>> import kwimage
>>> self = kwimage.Boxes.random(3, rng=0)
>>> orig = self.copy()
>>> self.quantize(inplace=True)
>>> assert np.any(self.data != orig.data)

numpy(self)

Converts tensors to numpy. Does not change memory if possible.

Example

>>> # xdoctest: +REQUIRES(module:torch)
>>> self = Boxes.random(3).tensor()
>>> newself = self.numpy()
>>> self.data[0, 0] = 0
>>> assert newself.data[0, 0] == 0
>>> self.data[0, 0] = 1
>>> assert self.data[0, 0] == 1

tensor(self, device=ub.NoParam)

Converts numpy to tensors. Does not change memory if possible.

Example

>>> # xdoctest: +REQUIRES(module:torch)
>>> self = Boxes.random(3)
>>> # xdoctest: +REQUIRES(module:torch)
>>> newself = self.tensor()
>>> self.data[0, 0] = 0
>>> assert newself.data[0, 0] == 0
>>> self.data[0, 0] = 1
>>> assert self.data[0, 0] == 1

ious(self, other, bias=0, impl='auto', mode=None)

Intersection over union.

Compute IOUs (intersection area over union area) between these boxes and another set of boxes. This is a symmetric measure of similarity between boxes.

Todo

• [ ] Add pairwise flag to toggle between one-vs-one and all-vs-all

  computation. I.E. Add option for componentwise calculation.

Parameters

• other (Boxes) – boxes to compare IoUs against

• bias (int, default=0) – either 0 or 1, does TL=BR have area of 0 or 1?

• impl (str, default=’auto’) – code to specify implementation used to ious. Can be either torch, py, c, or auto. Efficiency and the exact result will vary by implementation, but they will always be close. Some implementations only accept certain data types (e.g. impl=’c’, only accepts float32 numpy arrays). See ~/code/kwimage/dev/bench_bbox.py for benchmark details. On my system the torch impl was fastest (when the data was on the GPU).

• mode – depricated, use impl

SeeAlso:

iooas - for a measure of coverage between boxes

Examples

>>> import kwimage
>>> self = kwimage.Boxes(np.array([[ 0,  0, 10, 10],
>>>                                [10,  0, 20, 10],
>>>                                [20,  0, 30, 10]]), 'ltrb')
>>> other = kwimage.Boxes(np.array([6, 2, 20, 10]), 'ltrb')
>>> overlaps = self.ious(other, bias=1).round(2)
>>> assert np.all(np.isclose(overlaps, [0.21, 0.63, 0.04])), repr(overlaps)

Examples

>>> import kwimage
>>> boxes1 = kwimage.Boxes(np.array([[ 0,  0, 10, 10],
>>>                                  [10,  0, 20, 10],
>>>                                  [20,  0, 30, 10]]), 'ltrb')
>>> other = kwimage.Boxes(np.array([[6, 2, 20, 10],
>>>                                 [100, 200, 300, 300]]), 'ltrb')
>>> overlaps = boxes1.ious(other)
>>> print('{}'.format(ub.repr2(overlaps, precision=2, nl=1)))
np.array([[0.18, 0.  ],
          [0.61, 0.  ],
          [0.  , 0.  ]]...)

Examples

>>> # xdoctest: +IGNORE_WHITESPACE
>>> Boxes(np.empty(0), 'xywh').ious(Boxes(np.empty(4), 'xywh')).shape
(0,)
>>> #Boxes(np.empty(4), 'xywh').ious(Boxes(np.empty(0), 'xywh')).shape
>>> Boxes(np.empty((0, 4)), 'xywh').ious(Boxes(np.empty((0, 4)), 'xywh')).shape
(0, 0)
>>> Boxes(np.empty((1, 4)), 'xywh').ious(Boxes(np.empty((0, 4)), 'xywh')).shape
(1, 0)
>>> Boxes(np.empty((0, 4)), 'xywh').ious(Boxes(np.empty((1, 4)), 'xywh')).shape
(0, 1)

Examples

>>> # xdoctest: +REQUIRES(module:torch)
>>> formats = BoxFormat.cannonical
>>> istensors = [False, True]
>>> results = {}
>>> for format in formats:
>>>     for tensor in istensors:
>>>         boxes1 = Boxes.random(5, scale=10.0, rng=0, format=format, tensor=tensor)
>>>         boxes2 = Boxes.random(7, scale=10.0, rng=1, format=format, tensor=tensor)
>>>         ious = boxes1.ious(boxes2)
>>>         results[(format, tensor)] = ious
>>> results = {k: v.numpy() if torch.is_tensor(v) else v for k, v in results.items() }
>>> results = {k: v.tolist() for k, v in results.items()}
>>> print(ub.repr2(results, sk=True, precision=3, nl=2))
>>> from functools import partial
>>> assert ub.allsame(results.values(), partial(np.allclose, atol=1e-07))

Ignore:

>>> # does this work with backprop?
>>> # xdoctest: +REQUIRES(module:torch)
>>> import torch
>>> import kwimage
>>> num = 1000
>>> true_boxes = kwimage.Boxes.random(num).tensor()
>>> inputs = torch.rand(num, 10)
>>> regress = torch.nn.Linear(10, 4)
>>> energy = regress(inputs)
>>> energy.retain_grad()
>>> outputs = energy.sigmoid()
>>> outputs.retain_grad()
>>> out_boxes = kwimage.Boxes(outputs, 'cxywh')
>>> ious = out_boxes.ious(true_boxes)
>>> loss = ious.sum()
>>> loss.backward()

iooas(self, other, bias=0)

Intersection over other area.

This is an asymetric measure of coverage. How much of the “other” boxes are covered by these boxes. It is the area of intersection between each pair of boxes and the area of the “other” boxes.

SeeAlso:

ious - for a measure of similarity between boxes

Parameters

• other (Boxes) – boxes to compare IoOA against

• bias (int, default=0) – either 0 or 1, does TL=BR have area of 0 or 1?

Examples

>>> self = Boxes(np.array([[ 0,  0, 10, 10],
>>>                        [10,  0, 20, 10],
>>>                        [20,  0, 30, 10]]), 'ltrb')
>>> other = Boxes(np.array([[6, 2, 20, 10], [0, 0, 0, 3]]), 'xywh')
>>> coverage = self.iooas(other, bias=0).round(2)
>>> print('coverage = {!r}'.format(coverage))

isect_area(self, other, bias=0)

Intersection part of intersection over union computation

Examples

>>> # xdoctest: +IGNORE_WHITESPACE
>>> self = Boxes.random(5, scale=10.0, rng=0, format='ltrb')
>>> other = Boxes.random(3, scale=10.0, rng=1, format='ltrb')
>>> isect = self.isect_area(other, bias=0)
>>> ious_v1 = isect / ((self.area + other.area.T) - isect)
>>> ious_v2 = self.ious(other, bias=0)
>>> assert np.allclose(ious_v1, ious_v2)

intersection(self, other)

Componentwise intersection between two sets of Boxes

intersections of boxes are always boxes, so this works

Returns

intersected boxes

Return type

Boxes

Examples

>>> # xdoctest: +IGNORE_WHITESPACE
>>> from kwimage.structs.boxes import *  # NOQA
>>> self = Boxes.random(5, rng=0).scale(10.)
>>> other = self.translate(1)
>>> new = self.intersection(other)
>>> new_area = np.nan_to_num(new.area).ravel()
>>> alt_area = np.diag(self.isect_area(other))
>>> close = np.isclose(new_area, alt_area)
>>> assert np.all(close)

union_hull(self, other)

Componentwise hull union between two sets of Boxes

NOTE: convert to polygon to do a real union.

Returns

unioned boxes

Return type

Boxes

Examples

>>> # xdoctest: +IGNORE_WHITESPACE
>>> from kwimage.structs.boxes import *  # NOQA
>>> self = Boxes.random(5, rng=0).scale(10.)
>>> other = self.translate(1)
>>> new = self.union_hull(other)
>>> new_area = np.nan_to_num(new.area).ravel()

bounding_box(self)

Returns the box that bounds all of the contained boxes

Returns

a single box

Return type

Boxes

Examples

>>> # xdoctest: +IGNORE_WHITESPACE
>>> from kwimage.structs.boxes import *  # NOQA
>>> self = Boxes.random(5, rng=0).scale(10.)
>>> other = self.translate(1)
>>> new = self.union_hull(other)
>>> new_area = np.nan_to_num(new.area).ravel()

contains(self, other)

Determine of points are completely contained by these boxes

Parameters

other (Points) – points to test for containment. TODO: support generic data types

Returns

N x M boolean matrix indicating which box

contains which points, where N is the number of boxes and M is the number of points.

Return type

flags (ArrayLike)

Examples

>>> import kwimage
>>> self = kwimage.Boxes.random(10).scale(10).round()
>>> other = kwimage.Points.random(10).scale(10).round()
>>> flags = self.contains(other)
>>> flags = self.contains(self.xy_center)
>>> assert np.all(np.diag(flags))

view(self, *shape)

Passthrough method to view or reshape

Example

>>> # xdoctest: +REQUIRES(module:torch)
>>> self = Boxes.random(6, scale=10.0, rng=0, format='xywh').tensor()
>>> assert list(self.view(3, 2, 4).data.shape) == [3, 2, 4]
>>> self = Boxes.random(6, scale=10.0, rng=0, format='ltrb').tensor()
>>> assert list(self.view(3, 2, 4).data.shape) == [3, 2, 4]

class kwimage.Coords(data=None, meta=None)

Bases: kwimage.structs._generic.Spatial, ubelt.NiceRepr

A data structure to store n-dimensional coordinate geometry.

Currently it is up to the user to maintain what coordinate system this geometry belongs to.

Note

This class was designed to hold coordinates in r/c format, but in general this class is anostic to dimension ordering as long as you are consistent. However, there are two places where this matters:

(1) drawing and (2) gdal/imgaug-warping. In these places we will assume x/y for legacy reasons. This may change in the future.

The term axes with resepct to Coords always refers to the final numpy axis. In other words the final numpy-axis represents ALL of the coordinate-axes.

CommandLine:

xdoctest -m kwimage.structs.coords Coords

Example

>>> from kwimage.structs.coords import *  # NOQA
>>> import kwarray
>>> rng = kwarray.ensure_rng(0)
>>> self = Coords.random(num=4, dim=3, rng=rng)
>>> print('self = {}'.format(self))
self = <Coords(data=
    array([[0.5488135 , 0.71518937, 0.60276338],
           [0.54488318, 0.4236548 , 0.64589411],
           [0.43758721, 0.891773  , 0.96366276],
           [0.38344152, 0.79172504, 0.52889492]]))>
>>> matrix = rng.rand(4, 4)
>>> self.warp(matrix)
<Coords(data=
    array([[0.71037426, 1.25229659, 1.39498435],
           [0.60799503, 1.26483447, 1.42073131],
           [0.72106004, 1.39057144, 1.38757508],
           [0.68384299, 1.23914654, 1.29258196]]))>
>>> self.translate(3, inplace=True)
<Coords(data=
    array([[3.5488135 , 3.71518937, 3.60276338],
           [3.54488318, 3.4236548 , 3.64589411],
           [3.43758721, 3.891773  , 3.96366276],
           [3.38344152, 3.79172504, 3.52889492]]))>
>>> self.translate(3, inplace=True)
<Coords(data=
    array([[6.5488135 , 6.71518937, 6.60276338],
           [6.54488318, 6.4236548 , 6.64589411],
           [6.43758721, 6.891773  , 6.96366276],
           [6.38344152, 6.79172504, 6.52889492]]))>
>>> self.scale(2)
<Coords(data=
    array([[13.09762701, 13.43037873, 13.20552675],
           [13.08976637, 12.8473096 , 13.29178823],
           [12.87517442, 13.783546  , 13.92732552],
           [12.76688304, 13.58345008, 13.05778984]]))>
>>> # xdoctest: +REQUIRES(module:torch)
>>> self.tensor()
>>> self.tensor().tensor().numpy().numpy()
>>> self.numpy()
>>> #self.draw_on()

__repr__

__nice__(self)

__len__(self)

property dtype(self)

property dim(self)

property shape(self)

copy(self)

classmethod random(Coords, num=1, dim=2, rng=None, meta=None)

Makes random coordinates; typically for testing purposes

is_numpy(self)

is_tensor(self)

compress(self, flags, axis=0, inplace=False)

Filters items based on a boolean criterion

Parameters

• flags (ArrayLike[bool]) – true for items to be kept

• axis (int) – you usually want this to be 0

• inplace (bool, default=False) – if True, modifies this object

Returns

filtered coords

Return type

Coords

Example

>>> from kwimage.structs.coords import *  # NOQA
>>> self = Coords.random(10, rng=0)
>>> self.compress([True] * len(self))
>>> self.compress([False] * len(self))
<Coords(data=array([], shape=(0, 2), dtype=float64))>
>>> # xdoctest: +REQUIRES(module:torch)
>>> self = self.tensor()
>>> self.compress([True] * len(self))
>>> self.compress([False] * len(self))

take(self, indices, axis=0, inplace=False)

Takes a subset of items at specific indices

Parameters

• indices (ArrayLike[int]) – indexes of items to take

• axis (int) – you usually want this to be 0

• inplace (bool, default=False) – if True, modifies this object

Returns

filtered coords

Return type

Coords

Example

>>> self = Coords(np.array([[25, 30, 15, 10]]))
>>> self.take([0])
<Coords(data=array([[25, 30, 15, 10]]))>
>>> self.take([])
<Coords(data=array([], shape=(0, 4), dtype=int64))>

astype(self, dtype, inplace=False)

Changes the data type

Parameters

• dtype – new type

• inplace (bool, default=False) – if True, modifies this object

Returns

modified coordinates

Return type

Coords

round(self, inplace=False)

Rounds data to the nearest integer

Parameters

inplace (bool, default=False) – if True, modifies this object

Example

>>> import kwimage
>>> self = kwimage.Coords.random(3).scale(10)
>>> self.round()

view(self, *shape)

Passthrough method to view or reshape

Parameters

*shape – new shape of the data

Returns

modified coordinates

Return type

Coords

Example

>>> self = Coords.random(6, dim=4).numpy()
>>> assert list(self.view(3, 2, 4).data.shape) == [3, 2, 4]
>>> # xdoctest: +REQUIRES(module:torch)
>>> self = Coords.random(6, dim=4).tensor()
>>> assert list(self.view(3, 2, 4).data.shape) == [3, 2, 4]

classmethod concatenate(cls, coords, axis=0)

Concatenates lists of coordinates together

Parameters

• coords (Sequence[Coords]) – list of coords to concatenate

• axis (int, default=0) – axis to stack on

Returns

stacked coords

Return type

Coords

CommandLine:

xdoctest -m kwimage.structs.coords Coords.concatenate

Example

>>> coords = [Coords.random(3) for _ in range(3)]
>>> new = Coords.concatenate(coords)
>>> assert len(new) == 9
>>> assert np.all(new.data[3:6] == coords[1].data)

property device(self)

If the backend is torch returns the data device, otherwise None

property _impl(self)

Returns the internal tensor/numpy ArrayAPI implementation

tensor(self, device=ub.NoParam)

Converts numpy to tensors. Does not change memory if possible.

Returns

modified coordinates

Return type

Coords

Example

>>> # xdoctest: +REQUIRES(module:torch)
>>> self = Coords.random(3).numpy()
>>> newself = self.tensor()
>>> self.data[0, 0] = 0
>>> assert newself.data[0, 0] == 0
>>> self.data[0, 0] = 1
>>> assert self.data[0, 0] == 1

numpy(self)

Converts tensors to numpy. Does not change memory if possible.

Returns

modified coordinates

Return type

Coords

Example

>>> # xdoctest: +REQUIRES(module:torch)
>>> self = Coords.random(3).tensor()
>>> newself = self.numpy()
>>> self.data[0, 0] = 0
>>> assert newself.data[0, 0] == 0
>>> self.data[0, 0] = 1
>>> assert self.data[0, 0] == 1

reorder_axes(self, new_order, inplace=False)

Change the ordering of the coordinate axes.

Parameters

• new_order (Tuple[int]) – new_order[i] should specify which axes in the original coordinates should be mapped to the i-th position in the returned axes.

• inplace (bool, default=False) – if True, modifies data inplace

Returns

modified coordinates

Return type

Coords

Note

This is the ordering of the “columns” in final numpy axis, not the numpy axes themselves.

Example

>>> from kwimage.structs.coords import *  # NOQA
>>> self = Coords(data=np.array([
>>>     [7, 11],
>>>     [13, 17],
>>>     [21, 23],
>>> ]))
>>> new = self.reorder_axes((1, 0))
>>> print('new = {!r}'.format(new))
new = <Coords(data=
    array([[11,  7],
           [17, 13],
           [23, 21]]))>

Example

>>> from kwimage.structs.coords import *  # NOQA
>>> self = Coords.random(10, rng=0)
>>> new = self.reorder_axes((1, 0))
>>> # Remapping using 1, 0 reverses the axes
>>> assert np.all(new.data[:, 0] == self.data[:, 1])
>>> assert np.all(new.data[:, 1] == self.data[:, 0])
>>> # Remapping using 0, 1 does nothing
>>> eye = self.reorder_axes((0, 1))
>>> assert np.all(eye.data == self.data)
>>> # Remapping using 0, 0, destroys the 1-th column
>>> bad = self.reorder_axes((0, 0))
>>> assert np.all(bad.data[:, 0] == self.data[:, 0])
>>> assert np.all(bad.data[:, 1] == self.data[:, 0])

warp(self, transform, input_dims=None, output_dims=None, inplace=False)

Generalized coordinate transform.

Parameters

• transform (GeometricTransform | ArrayLike | Augmenter | callable) – scikit-image tranform, a 3x3 transformation matrix, an imgaug Augmenter, or generic callable which transforms an NxD ndarray.

• input_dims (Tuple) – shape of the image these objects correspond to (only needed / used when transform is an imgaug augmenter)

• output_dims (Tuple) – unused in non-raster structures, only exists for compatibility.

• inplace (bool, default=False) – if True, modifies data inplace

Returns

modified coordinates

Return type

Coords

Notes

Let D = self.dims

transformation matrices can be either:

• (D + 1) x (D + 1) # for homog

• D x D # for scale / rotate

• D x (D + 1) # for affine

Example

>>> from kwimage.structs.coords import *  # NOQA
>>> self = Coords.random(10, rng=0)
>>> transform = skimage.transform.AffineTransform(scale=(2, 2))
>>> new = self.warp(transform)
>>> assert np.all(new.data == self.scale(2).data)

Doctest:

>>> self = Coords.random(10, rng=0)
>>> assert np.all(self.warp(np.eye(3)).data == self.data)
>>> assert np.all(self.warp(np.eye(2)).data == self.data)

Doctest:

>>> # xdoctest: +REQUIRES(module:osgeo)
>>> from osgeo import osr
>>> wgs84_crs = osr.SpatialReference()
>>> wgs84_crs.ImportFromEPSG(4326)
>>> dst_crs = osr.SpatialReference()
>>> dst_crs.ImportFromEPSG(2927)
>>> transform = osr.CoordinateTransformation(wgs84_crs, dst_crs)
>>> self = Coords.random(10, rng=0)
>>> new = self.warp(transform)
>>> assert np.all(new.data != self.data)

>>> # Alternative using generic func
>>> def _gdal_coord_tranform(pts):
...     return np.array([transform.TransformPoint(x, y, 0)[0:2]
...                      for x, y in pts])
>>> alt = self.warp(_gdal_coord_tranform)
>>> assert np.all(alt.data != self.data)
>>> assert np.all(alt.data == new.data)

Doctest:

>>> # can use a generic function
>>> def func(xy):
...     return np.zeros_like(xy)
>>> self = Coords.random(10, rng=0)
>>> assert np.all(self.warp(func).data == 0)

_warp_imgaug(self, augmenter, input_dims, inplace=False)

Warps by applying an augmenter from the imgaug library

Note

We are assuming you are using X/Y coordinates here.

Parameters

• augmenter (imgaug.augmenters.Augmenter)

• input_dims (Tuple) – h/w of the input image

• inplace (bool, default=False) – if True, modifies data inplace

CommandLine:

xdoctest -m ~/code/kwimage/kwimage/structs/coords.py Coords._warp_imgaug

Example

>>> # xdoctest: +REQUIRES(module:imgaug)
>>> from kwimage.structs.coords import *  # NOQA
>>> import imgaug
>>> input_dims = (10, 10)
>>> self = Coords.random(10).scale(input_dims)
>>> augmenter = imgaug.augmenters.Fliplr(p=1)
>>> new = self._warp_imgaug(augmenter, input_dims)
>>> # y coordinate should not change
>>> assert np.allclose(self.data[:, 1], new.data[:, 1])
>>> assert np.allclose(input_dims[0] - self.data[:, 0], new.data[:, 0])

>>> # xdoc: +REQUIRES(--show)
>>> import kwplot
>>> kwplot.autompl()
>>> kwplot.figure(fnum=1, doclf=True)
>>> from matplotlib import pyplot as pl
>>> ax = plt.gca()
>>> ax.set_xlim(0, input_dims[0])
>>> ax.set_ylim(0, input_dims[1])
>>> self.draw(color='red', alpha=.4, radius=0.1)
>>> new.draw(color='blue', alpha=.4, radius=0.1)

Example

>>> # xdoctest: +REQUIRES(module:imgaug)
>>> from kwimage.structs.coords import *  # NOQA
>>> import imgaug
>>> input_dims = (32, 32)
>>> inplace = 0
>>> self = Coords.random(1000, rng=142).scale(input_dims).scale(.8)
>>> self.data = self.data.astype(np.int32).astype(np.float32)
>>> augmenter = imgaug.augmenters.CropAndPad(px=(-4, 4), keep_size=1).to_deterministic()
>>> new = self._warp_imgaug(augmenter, input_dims)
>>> # Change should be linear
>>> norm1 = (self.data - self.data.min(axis=0)) / (self.data.max(axis=0) - self.data.min(axis=0))
>>> norm2 = (new.data - new.data.min(axis=0)) / (new.data.max(axis=0) - new.data.min(axis=0))
>>> diff = norm1 - norm2
>>> assert np.allclose(diff, 0, atol=1e-6, rtol=1e-4)
>>> #assert np.allclose(self.data[:, 1], new.data[:, 1])
>>> #assert np.allclose(input_dims[0] - self.data[:, 0], new.data[:, 0])
>>> # xdoc: +REQUIRES(--show)
>>> import kwimage
>>> im = kwimage.imresize(kwimage.grab_test_image(), dsize=input_dims[::-1])
>>> new_im = augmenter.augment_image(im)
>>> import kwplot
>>> plt = kwplot.autoplt()
>>> kwplot.figure(fnum=1, doclf=True)
>>> kwplot.imshow(im, pnum=(1, 2, 1), fnum=1)
>>> self.draw(color='red', alpha=.8, radius=0.5)
>>> kwplot.imshow(new_im, pnum=(1, 2, 2), fnum=1)
>>> new.draw(color='blue', alpha=.8, radius=0.5, coord_axes=[1, 0])

to_imgaug(self, input_dims)

Translate to an imgaug object

Returns

imgaug data structure

Return type

imgaug.KeypointsOnImage

Example

>>> # xdoctest: +REQUIRES(module:imgaug)
>>> from kwimage.structs.coords import *  # NOQA
>>> self = Coords.random(10)
>>> input_dims = (10, 10)
>>> kpoi = self.to_imgaug(input_dims)
>>> new = Coords.from_imgaug(kpoi)
>>> assert np.allclose(new.data, self.data)

classmethod from_imgaug(cls, kpoi)

scale(self, factor, about=None, output_dims=None, inplace=False)

Scale coordinates by a factor

Parameters

• factor (float or Tuple[float, float]) – scale factor as either a scalar or per-dimension tuple.

• about (Tuple | None) – if unspecified scales about the origin (0, 0), otherwise the rotation is about this point.

• output_dims (Tuple) – unused in non-raster spatial structures

• inplace (bool, default=False) – if True, modifies data inplace

Returns

modified coordinates

Return type

Coords

Example

>>> from kwimage.structs.coords import *  # NOQA
>>> self = Coords.random(10, rng=0)
>>> new = self.scale(10)
>>> assert new.data.max() <= 10

>>> self = Coords.random(10, rng=0)
>>> self.data = (self.data * 10).astype(int)
>>> new = self.scale(10)
>>> assert new.data.dtype.kind == 'i'
>>> new = self.scale(10.0)
>>> assert new.data.dtype.kind == 'f'

translate(self, offset, output_dims=None, inplace=False)

Shift the coordinates

Parameters

• offset (float or Tuple[float]) – transation offset as either a scalar or a per-dimension tuple.

• output_dims (Tuple) – unused in non-raster spatial structures

• inplace (bool, default=False) – if True, modifies data inplace

Returns

modified coordinates

Return type

Coords

Example

>>> from kwimage.structs.coords import *  # NOQA
>>> self = Coords.random(10, dim=3, rng=0)
>>> new = self.translate(10)
>>> assert new.data.min() >= 10
>>> assert new.data.max() <= 11
>>> Coords.random(3, dim=3, rng=0)
>>> Coords.random(3, dim=3, rng=0).translate((1, 2, 3))

rotate(self, theta, about=None, output_dims=None, inplace=False)

Rotate the coordinates about a point.

Parameters

• theta (float) – rotation angle in radians

• about (Tuple | None) – if unspecified rotates about the origin (0, 0), otherwise the rotation is about this point.

• output_dims (Tuple) – unused in non-raster spatial structures

• inplace (bool, default=False) – if True, modifies data inplace

Returns

modified coordinates

Return type

Coords

Todo

• [ ] Generalized ND Rotations?

References

https://math.stackexchange.com/questions/197772/gen-rot-matrix

Example

>>> from kwimage.structs.coords import *  # NOQA
>>> self = Coords.random(10, dim=2, rng=0)
>>> theta = np.pi / 2
>>> new = self.rotate(theta)

>>> # Test rotate agrees with warp
>>> sin_ = np.sin(theta)
>>> cos_ = np.cos(theta)
>>> rot_ = np.array([[cos_, -sin_], [sin_,  cos_]])
>>> new2 = self.warp(rot_)
>>> assert np.allclose(new.data, new2.data)

>>> #
>>> # Rotate about a custom point
>>> theta = np.pi / 2
>>> new3 = self.rotate(theta, about=(0.5, 0.5))
>>> #
>>> # Rotate about the center of mass
>>> about = self.data.mean(axis=0)
>>> new4 = self.rotate(theta, about=about)
>>> # xdoc: +REQUIRES(--show)
>>> # xdoc: +REQUIRES(module:kwplot)
>>> import kwplot
>>> kwplot.figure(fnum=1, doclf=True)
>>> plt = kwplot.autoplt()
>>> self.draw(radius=0.01, color='blue', alpha=.5, coord_axes=[1, 0], setlim='grow')
>>> plt.gca().set_aspect('equal')
>>> new3.draw(radius=0.01, color='red', alpha=.5, coord_axes=[1, 0], setlim='grow')

_rectify_about(self, about)

Ensures that about returns a specified point. Allows for special keys like center to be used.

Example

>>> from kwimage.structs.coords import *  # NOQA
>>> self = Coords.random(10, dim=2, rng=0)

fill(self, image, value, coord_axes=None, interp='bilinear')

Sets sub-coordinate locations in a grid to a particular value

Parameters

coord_axes (Tuple) – specify which image axes each coordinate dim corresponds to. For 2D images, if you are storing r/c data, set to [0,1], if you are storing x/y data, set to [1,0].

Returns

image with coordinates rasterized on it

Return type

ndarray

soft_fill(self, image, coord_axes=None, radius=5)

Used for drawing keypoint truth in heatmaps

Parameters

coord_axes (Tuple) – specify which image axes each coordinate dim corresponds to. For 2D images, if you are storing r/c data, set to [0,1], if you are storing x/y data, set to [1,0].

In other words the i-th entry in coord_axes specifies which row-major spatial dimension the i-th column of a coordinate corresponds to. The index is the coordinate dimension and the value is the axes dimension.

Returns

image with coordinates rasterized on it

Return type

ndarray

References

https://stackoverflow.com/questions/54726703/generating-keypoint-heatmaps-in-tensorflow

Example

>>> from kwimage.structs.coords import *  # NOQA
>>> s = 64
>>> self = Coords.random(10, meta={'shape': (s, s)}).scale(s)
>>> # Put points on edges to to verify "edge cases"
>>> self.data[1] = [0, 0]       # top left
>>> self.data[2] = [s, s]       # bottom right
>>> self.data[3] = [0, s + 10]  # bottom left
>>> self.data[4] = [-3, s // 2] # middle left
>>> self.data[5] = [s + 1, -1]  # top right
>>> # Put points in the middle to verify overlap blending
>>> self.data[6] = [32.5, 32.5] # middle
>>> self.data[7] = [34.5, 34.5] # middle
>>> fill_value = 1
>>> coord_axes = [1, 0]
>>> radius = 10
>>> image1 = np.zeros((s, s))
>>> self.soft_fill(image1, coord_axes=coord_axes, radius=radius)
>>> radius = 3.0
>>> image2 = np.zeros((s, s))
>>> self.soft_fill(image2, coord_axes=coord_axes, radius=radius)
>>> # xdoc: +REQUIRES(--show)
>>> # xdoc: +REQUIRES(module:kwplot)
>>> import kwplot
>>> kwplot.autompl()
>>> kwplot.imshow(image1, pnum=(1, 2, 1))
>>> kwplot.imshow(image2, pnum=(1, 2, 2))

draw_on(self, image=None, fill_value=1, coord_axes=[1, 0], interp='bilinear')

Note

unlike other methods, the defaults assume x/y internal data

Parameters

coord_axes (Tuple) – specify which image axes each coordinate dim corresponds to. For 2D images, if you are storing r/c data, set to [0,1], if you are storing x/y data, set to [1,0].

In other words the i-th entry in coord_axes specifies which row-major spatial dimension the i-th column of a coordinate corresponds to. The index is the coordinate dimension and the value is the axes dimension.

Returns

image with coordinates drawn on it

Return type

ndarray

Example

>>> # xdoc: +REQUIRES(module:kwplot)
>>> from kwimage.structs.coords import *  # NOQA
>>> s = 256
>>> self = Coords.random(10, meta={'shape': (s, s)}).scale(s)
>>> self.data[0] = [10, 10]
>>> self.data[1] = [20, 40]
>>> image = np.zeros((s, s))
>>> fill_value = 1
>>> image = self.draw_on(image, fill_value, coord_axes=[1, 0], interp='bilinear')
>>> # image = self.draw_on(image, fill_value, coord_axes=[0, 1], interp='nearest')
>>> # image = self.draw_on(image, fill_value, coord_axes=[1, 0], interp='bilinear')
>>> # image = self.draw_on(image, fill_value, coord_axes=[1, 0], interp='nearest')
>>> # xdoc: +REQUIRES(--show)
>>> # xdoc: +REQUIRES(module:kwplot)
>>> import kwplot
>>> kwplot.autompl()
>>> kwplot.figure(fnum=1, doclf=True)
>>> kwplot.imshow(image)
>>> self.draw(radius=3, alpha=.5, coord_axes=[1, 0])

draw(self, color='blue', ax=None, alpha=None, coord_axes=[1, 0], radius=1, setlim=False)

Note

unlike other methods, the defaults assume x/y internal data

Parameters

• setlim (bool) – if True ensures the limits of the axes contains the polygon

• coord_axes (Tuple) – specify which image axes each coordinate dim corresponds to. For 2D images,

  if you are storing r/c data, set to [0,1], if you are storing x/y data, set to [1,0].

Returns

drawn matplotlib objects

Return type

List[mpl.collections.PatchCollection]

Example

>>> # xdoc: +REQUIRES(module:kwplot)
>>> from kwimage.structs.coords import *  # NOQA
>>> self = Coords.random(10)
>>> # xdoc: +REQUIRES(--show)
>>> self.draw(radius=3.0, setlim=True)
>>> import kwplot
>>> kwplot.autompl()
>>> self.draw(radius=3.0)

class kwimage.Detections(data=None, meta=None, datakeys=None, metakeys=None, checks=True, **kwargs)

Bases: ubelt.NiceRepr, _DetAlgoMixin, _DetDrawMixin

Container for holding and manipulating multiple detections.

Variables

• data (Dict) –

  dictionary containing corresponding lists. The length of each list is the number of detections. This contains the bounding boxes, confidence scores, and class indices. Details of the most common keys and types are as follows:

  boxes (kwimage.Boxes[ArrayLike]): multiple bounding boxes scores (ArrayLike): associated scores class_idxs (ArrayLike): associated class indices segmentations (ArrayLike): segmentations masks for each box,

  members can be Mask or MultiPolygon.

  keypoints (ArrayLike): keypoints for each box. Members should

  be Points.

  Additional custom keys may be specified as long as (a) the values are array-like and the first axis corresponds to the standard data values and (b) are custom keys are listed in the datakeys kwargs when constructing the Detections.

• meta (Dict) – This contains contextual information about the detections. This includes the class names, which can be indexed into via the class indexes.

Example

>>> import kwimage
>>> dets = kwimage.Detections(
>>>     # there are expected keys that do not need registration
>>>     boxes=kwimage.Boxes.random(3),
>>>     class_idxs=[0, 1, 1],
>>>     classes=['a', 'b'],
>>>     # custom data attrs must align with boxes
>>>     myattr1=np.random.rand(3),
>>>     myattr2=np.random.rand(3, 2, 8),
>>>     # there are no restrictions on metadata
>>>     mymeta='a custom metadata string',
>>>     # Note that any key not in kwimage.Detections.__datakeys__ or
>>>     # kwimage.Detections.__metakeys__ must be registered at the
>>>     # time of construction.
>>>     datakeys=['myattr1', 'myattr2'],
>>>     metakeys=['mymeta'],
>>>     checks=True,
>>> )
>>> print('dets = {}'.format(dets))
dets = <Detections(3)>

__datakeys__ = ['boxes', 'scores', 'class_idxs', 'probs', 'weights', 'keypoints', 'segmentations']

__metakeys__ = ['classes']

__nice__(self)

__len__(self)

copy(self)

Returns a deep copy of this Detections object

classmethod coerce(cls, data=None, **kwargs)

The “try-anything to get what I want” constructor

Parameters

• data

• **kwargs – currently boxes and cnames

Example

>>> from kwimage.structs.detections import *  # NOQA
>>> import kwimage
>>> kwargs = dict(
>>>     boxes=kwimage.Boxes.random(4),
>>>     cnames=['a', 'b', 'c', 'c'],
>>> )
>>> data = {}
>>> self = kwimage.Detections.coerce(data, **kwargs)

classmethod from_coco_annots(cls, anns, cats=None, classes=None, kp_classes=None, shape=None, dset=None)

Create a Detections object from a list of coco-like annotations.

Parameters

• anns (List[Dict]) – list of coco-like annotation objects

• dset (CocoDataset) – if specified, cats, classes, and kp_classes can are ignored.

• cats (List[Dict]) – coco-format category information. Used only if dset is not specified.

• classes (ndsampler.CategoryTree) – category tree with coco class info. Used only if dset is not specified.

• kp_classes (ndsampler.CategoryTree) – keypoint category tree with coco keypoint class info. Used only if dset is not specified.

• shape (tuple) – shape of parent image

Returns

a detections object

Return type

Detections

Example

>>> from kwimage.structs.detections import *  # NOQA
>>> # xdoctest: +REQUIRES(--module:ndsampler)
>>> anns = [{
>>>     'id': 0,
>>>     'image_id': 1,
>>>     'category_id': 2,
>>>     'bbox': [2, 3, 10, 10],
>>>     'keypoints': [4.5, 4.5, 2],
>>>     'segmentation': {
>>>         'counts': '_11a04M2O0O20N101N3L_5',
>>>         'size': [20, 20],
>>>     },
>>> }]
>>> dataset = {
>>>     'images': [],
>>>     'annotations': [],
>>>     'categories': [
>>>         {'id': 0, 'name': 'background'},
>>>         {'id': 2, 'name': 'class1', 'keypoints': ['spot']}
>>>     ]
>>> }
>>> #import ndsampler
>>> #dset = ndsampler.CocoDataset(dataset)
>>> cats = dataset['categories']
>>> dets = Detections.from_coco_annots(anns, cats)

Example

>>> # xdoctest: +REQUIRES(--module:ndsampler)
>>> # Test case with no category information
>>> from kwimage.structs.detections import *  # NOQA
>>> anns = [{
>>>     'id': 0,
>>>     'image_id': 1,
>>>     'category_id': None,
>>>     'bbox': [2, 3, 10, 10],
>>>     'prob': [.1, .9],
>>> }]
>>> cats = [
>>>     {'id': 0, 'name': 'background'},
>>>     {'id': 2, 'name': 'class1'}
>>> ]
>>> dets = Detections.from_coco_annots(anns, cats)

Example

>>> import kwimage
>>> # xdoctest: +REQUIRES(--module:ndsampler)
>>> import ndsampler
>>> sampler = ndsampler.CocoSampler.demo('photos')
>>> iminfo, anns = sampler.load_image_with_annots(1)
>>> shape = iminfo['imdata'].shape[0:2]
>>> kp_classes = sampler.dset.keypoint_categories()
>>> dets = kwimage.Detections.from_coco_annots(
>>>     anns, sampler.dset.dataset['categories'], sampler.catgraph,
>>>     kp_classes, shape=shape)

to_coco(self, cname_to_cat=None, style='orig', image_id=None, dset=None)

Converts this set of detections into coco-like annotation dictionaries.

Notes

Not all aspects of the MS-COCO format can be accurately represented, so some liberties are taken. The MS-COCO standard defines that annotations should specifiy a category_id field, but in some cases this information is not available so we will populate a ‘category_name’ field if possible and in the worst case fall back to ‘category_index’.

Additionally, detections may contain additional information beyond the MS-COCO standard, and this information (e.g. weight, prob, score) is added as forign fields.

Parameters

• cname_to_cat – currently ignored.

• style (str, default=’orig’) – either ‘orig’ (for the original coco format) or ‘new’ for the more general kwcoco-style coco format.

• image_id (int, default=None) – if specified, populates the image_id field of each image

• dset (CocoDataset, default=None) – if specified, attempts to populate the category_id field to be compatible with this coco dataset.

Yields

dict – coco-like annotation structures

Example

>>> # xdoctest: +REQUIRES(module:ndsampler)
>>> from kwimage.structs.detections import *
>>> self = Detections.demo()[0]
>>> cname_to_cat = None
>>> list(self.to_coco())

property boxes(self)

property class_idxs(self)

property scores(self)

typically only populated for predicted detections

property probs(self)

typically only populated for predicted detections

property weights(self)

typically only populated for groundtruth detections

property classes(self)

num_boxes(self)

warp(self, transform, input_dims=None, output_dims=None, inplace=False)

Spatially warp the detections.

Example

>>> import skimage
>>> transform = skimage.transform.AffineTransform(scale=(2, 3), translation=(4, 5))
>>> self = Detections.random(2)
>>> new = self.warp(transform)
>>> assert new.boxes == self.boxes.warp(transform)
>>> assert new != self

scale(self, factor, output_dims=None, inplace=False)

Spatially warp the detections.

Example

>>> import skimage
>>> transform = skimage.transform.AffineTransform(scale=(2, 3), translation=(4, 5))
>>> self = Detections.random(2)
>>> new = self.warp(transform)
>>> assert new.boxes == self.boxes.warp(transform)
>>> assert new != self

translate(self, offset, output_dims=None, inplace=False)

Spatially warp the detections.

Example

>>> import skimage
>>> self = Detections.random(2)
>>> new = self.translate(10)

classmethod concatenate(cls, dets)

Parameters

boxes (Sequence[Detections]) – list of detections to concatenate

Returns

stacked detections

Return type

Detections

Example

>>> self = Detections.random(2)
>>> other = Detections.random(3)
>>> dets = [self, other]
>>> new = Detections.concatenate(dets)
>>> assert new.num_boxes() == 5

>>> self = Detections.random(2, segmentations=True)
>>> other = Detections.random(3, segmentations=True)
>>> dets = [self, other]
>>> new = Detections.concatenate(dets)
>>> assert new.num_boxes() == 5

argsort(self, reverse=True)

Sorts detection indices by descending (or ascending) scores

Returns

sorted indices

Return type

ndarray[int]

sort(self, reverse=True)

Sorts detections by descending (or ascending) scores

Returns

sorted copy of self

Return type

kwimage.structs.Detections

compress(self, flags, axis=0)

Returns a subset where corresponding locations are True.

Parameters

flags (ndarray[bool]) – mask marking selected items

Returns

subset of self

Return type

kwimage.structs.Detections

CommandLine:

xdoctest -m kwimage.structs.detections Detections.compress

Example

>>> # xdoctest: +REQUIRES(module:torch)
>>> import kwimage
>>> dets = kwimage.Detections.random(keypoints='dense')
>>> flags = np.random.rand(len(dets)) > 0.5
>>> subset = dets.compress(flags)
>>> assert len(subset) == flags.sum()
>>> subset = dets.tensor().compress(flags)
>>> assert len(subset) == flags.sum()

take(self, indices, axis=0)

Returns a subset specified by indices

Parameters

indices (ndarray[int]) – indices to select

Returns

subset of self

Return type

kwimage.structs.Detections

Example

>>> import kwimage
>>> dets = kwimage.Detections(boxes=kwimage.Boxes.random(10))
>>> subset = dets.take([2, 3, 5, 7])
>>> assert len(subset) == 4
>>> # xdoctest: +REQUIRES(module:torch)
>>> subset = dets.tensor().take([2, 3, 5, 7])
>>> assert len(subset) == 4

__getitem__(self, index)

Fancy slicing / subset / indexing.

Note: scalar indices are always coerced into index lists of length 1.

Example

>>> import kwimage
>>> import kwarray
>>> dets = kwimage.Detections(boxes=kwimage.Boxes.random(10))
>>> indices = [2, 3, 5, 7]
>>> flags = kwarray.boolmask(indices, len(dets))
>>> assert dets[flags].data == dets[indices].data

property device(self)

If the backend is torch returns the data device, otherwise None

is_tensor(self)

is the backend fueled by torch?

is_numpy(self)

is the backend fueled by numpy?

numpy(self)

Converts tensors to numpy. Does not change memory if possible.

Example

>>> # xdoctest: +REQUIRES(module:torch)
>>> self = Detections.random(3).tensor()
>>> newself = self.numpy()
>>> self.scores[0] = 0
>>> assert newself.scores[0] == 0
>>> self.scores[0] = 1
>>> assert self.scores[0] == 1
>>> self.numpy().numpy()

property dtype(self)

tensor(self, device=ub.NoParam)

Converts numpy to tensors. Does not change memory if possible.

Example

>>> # xdoctest: +REQUIRES(module:torch)
>>> from kwimage.structs.detections import *
>>> self = Detections.random(3)
>>> newself = self.tensor()
>>> self.scores[0] = 0
>>> assert newself.scores[0] == 0
>>> self.scores[0] = 1
>>> assert self.scores[0] == 1
>>> self.tensor().tensor()

classmethod demo(Detections)

classmethod random(cls, num=10, scale=1.0, classes=3, keypoints=False, segmentations=False, tensor=False, rng=None)

Creates dummy data, suitable for use in tests and benchmarks

Parameters

• num (int) – number of boxes

• scale (float | tuple, default=1.0) – bounding image size

• classes (int | Sequence) – list of class labels or number of classes

• keypoints (bool, default=False) – if True include random keypoints for each box.

• segmentations (bool, default=False) – if True include random segmentations for each box.

• tensor (bool, default=False) – determines backend. DEPRECATED. Call tensor on resulting object instead.

• rng (np.random.RandomState) – random state

Example

>>> import kwimage
>>> dets = kwimage.Detections.random(keypoints='jagged')
>>> dets.data['keypoints'].data[0].data
>>> dets.data['keypoints'].meta
>>> dets = kwimage.Detections.random(keypoints='dense')
>>> dets = kwimage.Detections.random(keypoints='dense', segmentations=True).scale(1000)
>>> # xdoctest:+REQUIRES(--show)
>>> import kwplot
>>> kwplot.autompl()
>>> dets.draw(setlim=True)

Example

>>> import kwimage
>>> dets = kwimage.Detections.random(
>>>     keypoints='jagged', segmentations=True, rng=0).scale(1000)
>>> print('dets = {}'.format(dets))
dets = <Detections(10)>
>>> dets.data['boxes'].quantize(inplace=True)
>>> print('dets.data = {}'.format(ub.repr2(
>>>     dets.data, nl=1, with_dtype=False, strvals=True)))
dets.data = {
    'boxes': <Boxes(xywh,
                 array([[548, 544,  55, 172],
                        [423, 645,  15, 247],
                        [791, 383, 173, 146],
                        [ 71,  87, 498, 839],
                        [ 20, 832, 759,  39],
                        [461, 780, 518,  20],
                        [118, 639,  26, 306],
                        [264, 414, 258, 361],
                        [ 18, 568, 439,  50],
                        [612, 616, 332,  66]], dtype=int32))>,
    'class_idxs': [1, 2, 0, 0, 2, 0, 0, 0, 0, 0],
    'keypoints': <PointsList(n=10)>,
    'scores': [0.3595079 , 0.43703195, 0.6976312 , 0.06022547, 0.66676672, 0.67063787,0.21038256, 0.1289263 , 0.31542835, 0.36371077],
    'segmentations': <SegmentationList(n=10)>,
}
>>> # xdoctest:+REQUIRES(--show)
>>> import kwplot
>>> kwplot.autompl()
>>> dets.draw(setlim=True)

Example

>>> # Boxes position/shape within 0-1 space should be uniform.
>>> # xdoctest: +REQUIRES(--show)
>>> import kwplot
>>> kwplot.autompl()
>>> fig = kwplot.figure(fnum=1, doclf=True)
>>> fig.gca().set_xlim(0, 128)
>>> fig.gca().set_ylim(0, 128)
>>> import kwimage
>>> kwimage.Detections.random(num=10, segmentations=True).scale(128).draw()

class kwimage.Heatmap(data=None, meta=None, **kwargs)

Bases: kwimage.structs._generic.Spatial, _HeatmapDrawMixin, _HeatmapWarpMixin, _HeatmapAlgoMixin

Keeps track of a downscaled heatmap and how to transform it to overlay the original input image. Heatmaps generally are used to estimate class probabilites at each pixel. This data struction additionally contains logic to augment pixel with offset (dydx) and scale (diamter) information.

Variables

• data (Dict[str, ArrayLike]) –

  dictionary containing spatially aligned heatmap data. Valid keys are as follows.

  class_probs (ArrayLike[C, H, W] | ArrayLike[C, D, H, W]):

  A probability map for each class. C is the number of classes.

  offset (ArrayLike[2, H, W] | ArrayLike[3, D, H, W], optional):

  object center position offset in y,x / t,y,x coordinates

  diamter (ArrayLike[2, H, W] | ArrayLike[3, D, H, W], optional):

  object bounding box sizes in h,w / d,h,w coordinates

  keypoints (ArrayLike[2, K, H, W] | ArrayLike[3, K, D, H, W], optional):

  y/x offsets for K different keypoint classes

• meta (Dict[str, object]) –

  dictionary containing miscellanious metadata about the heatmap data. Valid keys are as follows.

  img_dims (Tuple[H, W] | Tuple[D, H, W]):

  original image dimension

  tf_data_to_image (skimage.transform._geometric.GeometricTransform):

  transformation matrix (typically similarity or affine) that projects the given, heatmap onto the image dimensions such that the image and heatmap are spatially aligned.

  classes (List[str] | ndsampler.CategoryTree):

  information about which index in data[‘class_probs’] corresponds to which semantic class.

• dims (Tuple) – dimensions of the heatmap (See `image_dims) for the original image dimensions.

• **kwargs – any key that is accepted by the data or meta dictionaries can be specified as a keyword argument to this class and it will be properly placed in the appropriate internal dictionary.

CommandLine:

xdoctest -m ~/code/kwimage/kwimage/structs/heatmap.py Heatmap –show

Example

>>> # xdoctest: +REQUIRES(module:torch)
>>> from kwimage.structs.heatmap import *  # NOQA
>>> import kwimage
>>> class_probs = kwimage.grab_test_image(dsize=(32, 32), space='gray')[None, ] / 255.0
>>> img_dims = (220, 220)
>>> tf_data_to_img = skimage.transform.AffineTransform(translation=(-18, -18), scale=(8, 8))
>>> self = Heatmap(class_probs=class_probs, img_dims=img_dims,
>>>                tf_data_to_img=tf_data_to_img)
>>> aligned = self.upscale()
>>> # xdoctest: +REQUIRES(--show)
>>> import kwplot
>>> kwplot.autompl()
>>> kwplot.imshow(aligned[0])
>>> kwplot.show_if_requested()

Example

>>> # xdoctest: +REQUIRES(module:torch)
>>> import kwimage
>>> self = Heatmap.random()
>>> # xdoctest: +REQUIRES(--show)
>>> import kwplot
>>> kwplot.autompl()
>>> self.draw()

__datakeys__ = ['class_probs', 'offset', 'diameter', 'keypoints', 'class_idx', 'class_energy']

__metakeys__ = ['img_dims', 'tf_data_to_img', 'classes', 'kp_classes']

__spatialkeys__ = ['offset', 'diameter', 'keypoints']

__nice__(self)

__getitem__(self, index)

__len__(self)

property shape(self)

property bounds(self)

property dims(self)

space-time dimensions of this heatmap

is_numpy(self)

is_tensor(self)

property _impl(self)

Returns the internal tensor/numpy ArrayAPI implementation

Returns

kwarray.ArrayAPI

classmethod random(cls, dims=(10, 10), classes=3, diameter=True, offset=True, keypoints=False, img_dims=None, dets=None, nblips=10, noise=0.0, rng=None)

Creates dummy data, suitable for use in tests and benchmarks

Parameters

• dims (Tuple) – dimensions of the heatmap

• img_dims (Tuple) – dimensions of the image the heatmap corresponds to

Example

>>> from kwimage.structs.heatmap import *  # NOQA
>>> self = Heatmap.random((128, 128), img_dims=(200, 200),
>>>     classes=3, nblips=10, rng=0, noise=0.1)
>>> # xdoctest: +REQUIRES(--show)
>>> import kwplot
>>> kwplot.autompl()
>>> kwplot.imshow(self.colorize(0, imgspace=0), fnum=1, pnum=(1, 4, 1), doclf=1)
>>> kwplot.imshow(self.colorize(1, imgspace=0), fnum=1, pnum=(1, 4, 2))
>>> kwplot.imshow(self.colorize(2, imgspace=0), fnum=1, pnum=(1, 4, 3))
>>> kwplot.imshow(self.colorize(3, imgspace=0), fnum=1, pnum=(1, 4, 4))

Ignore:

self.detect(0).sort().non_max_supress()[-np.arange(1, 4)].draw() from kwimage.structs.heatmap import * # NOQA import xdev globals().update(xdev.get_func_kwargs(Heatmap.random))

Example

>>> # xdoctest: +REQUIRES(module:ndsampler)
>>> import kwimage
>>> self = kwimage.Heatmap.random(dims=(50, 200), dets='coco',
>>>                               keypoints=True)
>>> image = np.zeros(self.img_dims)
>>> # xdoctest: +REQUIRES(module:kwplot)
>>> toshow = self.draw_on(image, 1, vecs=True, kpts=0, with_alpha=0.85)
>>> # xdoctest: +REQUIRES(--show)
>>> import kwplot
>>> kwplot.autompl()
>>> kwplot.figure(fnum=1, doclf=True)
>>> kwplot.imshow(toshow)

Ignore:

>>> kwplot.figure(fnum=1, doclf=True)
>>> kwplot.imshow(image)
>>> dets.draw()
>>> dets.data['keypoints'].draw(radius=6)
>>> dets.data['segmentations'].draw()

>>> self.draw()

property class_probs(self)

property offset(self)

property diameter(self)

property img_dims(self)

property tf_data_to_img(self)

property classes(self)

numpy(self)

Converts underlying data to numpy arrays

tensor(self, device=ub.NoParam)

Converts underlying data to torch tensors

class kwimage.Mask(data=None, format=None)

Bases: ubelt.NiceRepr, _MaskConversionMixin, _MaskConstructorMixin, _MaskTransformMixin, _MaskDrawMixin

Manages a single segmentation mask and can convert to and from multiple formats including:

• bytes_rle - byte encoded run length encoding

• array_rle - raw run length encoding

• c_mask - c-style binary mask

• f_mask - fortran-style binary mask

Example

>>> # xdoc: +REQUIRES(--mask)
>>> # a ms-coco style compressed bytes rle segmentation
>>> segmentation = {'size': [5, 9], 'counts': ';?1B10O30O4'}
>>> mask = Mask(segmentation, 'bytes_rle')
>>> # convert to binary numpy representation
>>> binary_mask = mask.to_c_mask().data
>>> print(ub.repr2(binary_mask.tolist(), nl=1, nobr=1))
[0, 0, 0, 1, 1, 1, 1, 1, 0],
[0, 0, 1, 1, 1, 0, 0, 0, 0],
[0, 0, 1, 1, 1, 1, 1, 1, 0],
[0, 0, 1, 1, 1, 0, 1, 1, 0],
[0, 0, 1, 1, 1, 0, 1, 1, 0],

property dtype(self)

__nice__(self)

classmethod random(Mask, rng=None, shape=(32, 32))

Create a random binary mask object

Parameters

• rng (int | RandomState | None) – the random seed

• shape (Tuple[int, int]) – the height / width of the returned mask

Returns

the random mask

Return type

Mask

Example

>>> import kwimage
>>> mask = kwimage.Mask.random()
>>> # xdoc: +REQUIRES(--show)
>>> import kwplot
>>> kwplot.autompl()
>>> mask.draw()
>>> kwplot.show_if_requested()

classmethod demo(cls)

Demo mask with holes and disjoint shapes

Returns

the demo mask

Return type

Mask

copy(self)

Performs a deep copy of the mask data

Returns

the copied mask

Return type

Mask

Example

>>> self = Mask.random(shape=(8, 8), rng=0)
>>> other = self.copy()
>>> assert other.data is not self.data

union(self, *others)

This can be used as a staticmethod or an instancemethod

Parameters

*others – multiple input masks to union

Returns

the unioned mask

Return type

Mask

Example

>>> # xdoc: +REQUIRES(--mask)
>>> from kwimage.structs.mask import *  # NOQA
>>> masks = [Mask.random(shape=(8, 8), rng=i) for i in range(2)]
>>> mask = Mask.union(*masks)
>>> print(mask.area)
>>> masks = [m.to_c_mask() for m in masks]
>>> mask = Mask.union(*masks)
>>> print(mask.area)

>>> masks = [m.to_bytes_rle() for m in masks]
>>> mask = Mask.union(*masks)
>>> print(mask.area)

Benchmark:

import ubelt as ub ti = ub.Timerit(100, bestof=10, verbose=2)

masks = [Mask.random(shape=(172, 172), rng=i) for i in range(2)]

for timer in ti.reset(‘native rle union’):

masks = [m.to_bytes_rle() for m in masks] with timer:

mask = Mask.union(*masks)

for timer in ti.reset(‘native cmask union’):

masks = [m.to_c_mask() for m in masks] with timer:

mask = Mask.union(*masks)

for timer in ti.reset(‘cmask->rle union’):

masks = [m.to_c_mask() for m in masks] with timer:

mask = Mask.union(*[m.to_bytes_rle() for m in masks])

intersection(self, *others)

This can be used as a staticmethod or an instancemethod

Parameters

*others – multiple input masks to intersect

Returns

the intersection of the masks

Return type

Mask

Example

>>> n = 3
>>> masks = [Mask.random(shape=(8, 8), rng=i) for i in range(n)]
>>> items = masks
>>> mask = Mask.intersection(*masks)
>>> areas = [item.area for item in items]
>>> print('areas = {!r}'.format(areas))
>>> print(mask.area)
>>> print(Mask.intersection(*masks).area / Mask.union(*masks).area)

property shape(self)

property area(self)

Returns the number of non-zero pixels

Returns

the number of non-zero pixels

Return type

int

Example

>>> self = Mask.demo()
>>> self.area
150

get_patch(self)

Extract the patch with non-zero data

Example

>>> # xdoc: +REQUIRES(--mask)
>>> from kwimage.structs.mask import *  # NOQA
>>> self = Mask.random(shape=(8, 8), rng=0)
>>> self.get_patch()

get_xywh(self)

Gets the bounding xywh box coordinates of this mask

Returns

x, y, w, h: Note we dont use a Boxes object because

a general singular version does not yet exist.

Return type

ndarray

Example

>>> # xdoc: +REQUIRES(--mask)
>>> self = Mask.random(shape=(8, 8), rng=0)
>>> self.get_xywh().tolist()
>>> self = Mask.random(rng=0).translate((10, 10))
>>> self.get_xywh().tolist()

Example

>>> # test empty case
>>> import kwimage
>>> self = kwimage.Mask(np.empty((0, 0), dtype=np.uint8), format='c_mask')
>>> assert self.get_xywh().tolist() == [0, 0, 0, 0]

Ignore:

>>> import kwimage
>>> self = kwimage.Mask(np.zeros((768, 768), dtype=np.uint8), format='c_mask')
>>> x_coords = np.array([621, 752])
>>> y_coords = np.array([366, 292])
>>> self.data[y_coords, x_coords] = 1
>>> self.get_xywh()

>>> # References:
>>> # https://stackoverflow.com/questions/33281957/faster-alternative-to-numpy-where
>>> # https://answers.opencv.org/question/4183/what-is-the-best-way-to-find-bounding-box-for-binary-mask/
>>> import timerit
>>> ti = timerit.Timerit(100, bestof=10, verbose=2)
>>> for timer in ti.reset('time'):
>>>     with timer:
>>>         y_coords, x_coords = np.where(self.data)
>>> #
>>> for timer in ti.reset('time'):
>>>     with timer:
>>>         cv2.findNonZero(data)

self.data = np.random.rand(800, 700) > 0.5

import timerit ti = timerit.Timerit(100, bestof=10, verbose=2) for timer in ti.reset(‘time’):

with timer:

y_coords, x_coords = np.where(self.data)

# for timer in ti.reset(‘time’):

with timer:

data = np.ascontiguousarray(self.data).astype(np.uint8) cv2_coords = cv2.findNonZero(data)

>>> poly = self.to_multi_polygon()

get_polygon(self)

DEPRECATED: USE to_multi_polygon

Returns a list of (x,y)-coordinate lists. The length of the list is equal to the number of disjoint regions in the mask.

Returns

polygon around each connected component of the

mask. Each ndarray is an Nx2 array of xy points.

Return type

List[ndarray]

Note

The returned polygon may not surround points that are only one pixel thick.

Example

>>> # xdoc: +REQUIRES(--mask)
>>> from kwimage.structs.mask import *  # NOQA
>>> self = Mask.random(shape=(8, 8), rng=0)
>>> polygons = self.get_polygon()
>>> print('polygons = ' + ub.repr2(polygons))
>>> polygons = self.get_polygon()
>>> self = self.to_bytes_rle()
>>> other = Mask.from_polygons(polygons, self.shape)
>>> # xdoc: +REQUIRES(--show)
>>> import kwplot
>>> kwplot.autompl()
>>> image = np.ones(self.shape)
>>> image = self.draw_on(image, color='blue')
>>> image = other.draw_on(image, color='red')
>>> kwplot.imshow(image)

polygons = [

np.array([[6, 4],[7, 4]], dtype=np.int32), np.array([[0, 1],[0, 3],[2, 3],[2, 1]], dtype=np.int32),

]

to_mask(self, dims=None)

Converts to a mask object (which does nothing because this already is mask object!)

Returns

kwimage.Mask

to_boxes(self)

Returns the bounding box of the mask.

Returns

kwimage.Boxes

to_multi_polygon(self)

Returns a MultiPolygon object fit around this raster including disjoint pieces and holes.

Returns

vectorized representation

Return type

MultiPolygon

Example

>>> # xdoc: +REQUIRES(--mask)
>>> from kwimage.structs.mask import *  # NOQA
>>> self = Mask.demo()
>>> self = self.scale(5)
>>> multi_poly = self.to_multi_polygon()
>>> # xdoc: +REQUIRES(module:kwplot)
>>> # xdoc: +REQUIRES(--show)
>>> self.draw(color='red')
>>> multi_poly.scale(1.1).draw(color='blue')

>>> # xdoc: +REQUIRES(--show)
>>> import kwplot
>>> kwplot.autompl()
>>> image = np.ones(self.shape)
>>> image = self.draw_on(image, color='blue')
>>> #image = other.draw_on(image, color='red')
>>> kwplot.imshow(image)
>>> multi_poly.draw()

Example

>>> import kwimage
>>> self = kwimage.Mask(np.empty((0, 0), dtype=np.uint8), format='c_mask')
>>> poly = self.to_multi_polygon()
>>> poly.to_multi_polygon()

Example

# Corner case, only two pixels are on >>> import kwimage >>> self = kwimage.Mask(np.zeros((768, 768), dtype=np.uint8), format=’c_mask’) >>> x_coords = np.array([621, 752]) >>> y_coords = np.array([366, 292]) >>> self.data[y_coords, x_coords] = 1 >>> poly = self.to_multi_polygon()

poly.to_mask(self.shape).data.sum()

self.to_array_rle().to_c_mask().data.sum() temp.to_c_mask().data.sum()

Example

>>> # TODO: how do we correctly handle the 1 or 2 point to a poly
>>> # case?
>>> import kwimage
>>> data = np.zeros((8, 8), dtype=np.uint8)
>>> data[0, 3:5] = 1
>>> data[7, 3:5] = 1
>>> data[3:5, 0:2] = 1
>>> self = kwimage.Mask.coerce(data)
>>> polys = self.to_multi_polygon()
>>> # xdoc: +REQUIRES(--show)
>>> import kwplot
>>> kwplot.autompl()
>>> kwplot.imshow(data)
>>> polys.draw(border=True, linewidth=5, alpha=0.5, radius=0.2)

get_convex_hull(self)

Returns a list of xy points around the convex hull of this mask

Note

The returned polygon may not surround points that are only one pixel thick.

Example

>>> # xdoc: +REQUIRES(--mask)
>>> self = Mask.random(shape=(8, 8), rng=0)
>>> polygons = self.get_convex_hull()
>>> print('polygons = ' + ub.repr2(polygons))
>>> other = Mask.from_polygons(polygons, self.shape)

iou(self, other)

The area of intersection over the area of union

Todo

• [ ] Write plural Masks version of this class, which should

  be able to perform this operation more efficiently.

CommandLine:

xdoctest -m kwimage.structs.mask Mask.iou

Example

>>> # xdoc: +REQUIRES(--mask)
>>> self = Mask.demo()
>>> other = self.translate(1)
>>> iou = self.iou(other)
>>> print('iou = {:.4f}'.format(iou))
iou = 0.0830
>>> iou2 = self.intersection(other).area / self.union(other).area
>>> print('iou2 = {:.4f}'.format(iou2))

classmethod coerce(Mask, data, dims=None)

Attempts to auto-inspect the format of the data and conver to Mask

Parameters

• data – the data to coerce

• dims (Tuple) – required for certain formats like polygons height / width of the source image

Returns

the constructed mask object

Return type

Mask

Example

>>> # xdoc: +REQUIRES(--mask)
>>> segmentation = {'size': [5, 9], 'counts': ';?1B10O30O4'}
>>> polygon = [
>>>     [np.array([[3, 0],[2, 1],[2, 4],[4, 4],[4, 3],[7, 0]])],
>>>     [np.array([[2, 1],[2, 2],[4, 2],[4, 1]])],
>>> ]
>>> dims = (9, 5)
>>> mask = (np.random.rand(32, 32) > .5).astype(np.uint8)
>>> Mask.coerce(polygon, dims).to_bytes_rle()
>>> Mask.coerce(segmentation).to_bytes_rle()
>>> Mask.coerce(mask).to_bytes_rle()

_to_coco(self)

use to_coco instead

to_coco(self, style='orig')

Convert the Mask to a COCO json representation based on the current format.

A COCO mask is formatted as a run-length-encoding (RLE), of which there are two variants: (1) a array RLE, which is slightly more readable and extensible, and (2) a bytes RLE, which is slightly more concise. The returned format will depend on the current format of the Mask object. If it is in “bytes_rle” format, it will be returned in that format, otherwise it will be converted to the “array_rle” format and returned as such.

Parameters

style (str) – Does nothing for this particular method, exists for API compatibility and if alternate encoding styles are implemented in the future.

Returns

either a bytes-rle or array-rle encoding, depending

on the current mask format. The keys in this dictionary are as follows:

counts (List[int] | str): the array or bytes rle encoding

size (Tuple[int]): the height and width of the encoded mask

see note.

shape (Tuple[int]): only present in array-rle mode. This

is also the height/width of the underlying encoded array. This exists for semantic consistency with other kwimage conventions, and is not part of the original coco spec.

order (str): only present in array-rle mode.

Either C or F, indicating if counts is aranged in row-major or column-major order. For COCO-compatibility this is always returned in F (column-major) order.

binary (bool): only present in array-rle mode.

For COCO-compatibility this is always returned as False, indicating the mask only contains binary 0 or 1 values.

Return type

dict

Note

The output dictionary will contain a key named “size”, this is the only location in kwimage where “size” refers to a tuple in (height/width) order, in order to be backwards compatible with the original coco spec. In all other locations in kwimage a “size” will refer to a (width/height) ordered tuple.

SeeAlso:

func

kwimage.im_runlen.encode_run_length - backend function that does array-style run length encoding.

Example

>>> # xdoc: +REQUIRES(--mask)
>>> from kwimage.structs.mask import *  # NOQA
>>> self = Mask.demo()
>>> coco_data1 = self.toformat('array_rle').to_coco()
>>> coco_data2 = self.toformat('bytes_rle').to_coco()
>>> print('coco_data1 = {}'.format(ub.repr2(coco_data1, nl=1)))
>>> print('coco_data2 = {}'.format(ub.repr2(coco_data2, nl=1)))
coco_data1 = {
    'binary': True,
    'counts': [47, 5, 3, 1, 14, ... 1, 4, 19, 141],
    'order': 'F',
    'shape': (23, 32),
    'size': (23, 32),
}
coco_data2 = {
    'counts': '_153L;4EL...ON3060L0N060L0Nb0Y4',
    'size': [23, 32],
}

class kwimage.MaskList

Bases: kwimage.structs._generic.ObjectList

Store and manipulate multiple masks, usually within the same image

to_polygon_list(self)

Converts all mask objects to multi-polygon objects

Returns

kwimage.PolygonList

to_segmentation_list(self)

Converts all items to segmentation objects

Returns

kwimage.SegmentationList

to_mask_list(self)

returns this object

Returns

kwimage.MaskList

class kwimage.MultiPolygon

Bases: kwimage.structs._generic.ObjectList

Data structure for storing multiple polygons (typically related to the same underlying but potentitally disjoing object)

Variables

data (List[Polygon]) –

classmethod random(self, n=3, n_holes=0, rng=None, tight=False)

Create a random MultiPolygon

Returns

MultiPolygon

fill(self, image, value=1)

Inplace fill in an image based on this multi-polyon.

Parameters

• image (ndarray) – image to draw on (inplace)

• value (int | Tuple[int], default=1) – value fill in with

Returns

the image that has been modified in place

Return type

ndarray

to_multi_polygon(self)

to_boxes(self)

Deprecated: lossy conversion use ‘bounding_box’ instead

bounding_box(self)

Return the bounding box of the multi polygon

Returns

a Boxes object with one box that encloses all

polygons

Return type

kwimage.Boxes

Example

>>> from kwimage.structs.polygon import *  # NOQA
>>> self = MultiPolygon.random(rng=0, n=10)
>>> boxes = self.to_boxes()
>>> sub_boxes = [d.to_boxes() for d in self.data]
>>> areas1 = np.array([s.intersection(boxes).area[0] for s in sub_boxes])
>>> areas2 = np.array([s.area[0] for s in sub_boxes])
>>> assert np.allclose(areas1, areas2)

to_mask(self, dims=None)

Returns a mask object indication regions occupied by this multipolygon

Example

>>> from kwimage.structs.polygon import *  # NOQA
>>> s = 100
>>> self = MultiPolygon.random(rng=0).scale(s)
>>> dims = (s, s)
>>> mask = self.to_mask(dims)

>>> # xdoc: +REQUIRES(--show)
>>> import kwplot
>>> kwplot.autompl()
>>> kwplot.figure(fnum=1, doclf=True)
>>> from matplotlib import pyplot as pl
>>> ax = plt.gca()
>>> ax.set_xlim(0, s)
>>> ax.set_ylim(0, s)
>>> self.draw(color='red', alpha=.4)
>>> mask.draw(color='blue', alpha=.4)

to_relative_mask(self)

Returns a translated mask such the mask dimensions are minimal.

In other words, we move the polygon all the way to the top-left and return a mask just big enough to fit the polygon.

Returns

Mask

classmethod coerce(cls, data, dims=None)

Attempts to construct a MultiPolygon instance from the input data

See Mask.coerce

to_shapely(self)

Example

>>> # xdoc: +REQUIRES(module:kwplot)
>>> # xdoc: +REQUIRES(module:shapely)
>>> from kwimage.structs.polygon import *  # NOQA
>>> self = MultiPolygon.random(rng=0)
>>> geom = self.to_shapely()
>>> print('geom = {!r}'.format(geom))

classmethod from_shapely(MultiPolygon, geom)

Convert a shapely polygon or multipolygon to a kwimage.MultiPolygon

classmethod from_geojson(MultiPolygon, data_geojson)

Convert a geojson polygon or multipolygon to a kwimage.MultiPolygon

Example

>>> import kwimage
>>> orig = kwimage.MultiPolygon.random()
>>> data_geojson = orig.to_geojson()
>>> self = kwimage.MultiPolygon.from_geojson(data_geojson)

to_geojson(self)

Converts polygon to a geojson structure

classmethod from_coco(cls, data, dims=None)

Accepts either new-style or old-style coco multi-polygons

_to_coco(self, style='orig')

to_coco(self, style='orig')

Example

>>> from kwimage.structs.polygon import *  # NOQA
>>> self = MultiPolygon.random(1, rng=0)
>>> self.to_coco()

swap_axes(self, inplace=False)

class kwimage.Points(data=None, meta=None, datakeys=None, metakeys=None, **kwargs)

Bases: kwimage.structs._generic.Spatial, _PointsWarpMixin

Stores multiple keypoints for a single object.

This stores both the geometry and the class metadata if available

Ignore:

meta = {

“names” = [‘head’, ‘nose’, ‘tail’], “skeleton” = [(0, 1), (0, 2)],

}

Example

>>> from kwimage.structs.points import *  # NOQA
>>> xy = np.random.rand(10, 2)
>>> pts = Points(xy=xy)
>>> print('pts = {!r}'.format(pts))

__datakeys__ = ['xy', 'class_idxs', 'visible']

__metakeys__ = ['classes']

__repr__

__nice__(self)

__len__(self)

property shape(self)

property xy(self)

classmethod random(Points, num=1, classes=None, rng=None)

Makes random points; typically for testing purposes

Example

>>> import kwimage
>>> self = kwimage.Points.random(classes=[1, 2, 3])
>>> self.data
>>> print('self.data = {!r}'.format(self.data))

is_numpy(self)

is_tensor(self)

_impl(self)

tensor(self, device=ub.NoParam)

Example

>>> # xdoctest: +REQUIRES(module:torch)
>>> from kwimage.structs.points import *  # NOQA
>>> self = Points.random(10)
>>> self.tensor()

round(self, inplace=False)

Rounds data to the nearest integer

Parameters

inplace (bool, default=False) – if True, modifies this object

Example

>>> import kwimage
>>> self = kwimage.Points.random(3).scale(10)
>>> self.round()

numpy(self)

Example

>>> # xdoctest: +REQUIRES(module:torch)
>>> from kwimage.structs.points import *  # NOQA
>>> self = Points.random(10)
>>> self.tensor().numpy().tensor().numpy()

draw_on(self, image, color='white', radius=None, copy=False)

CommandLine:

xdoctest -m ~/code/kwimage/kwimage/structs/points.py Points.draw_on –show

Example

>>> # xdoc: +REQUIRES(module:kwplot)
>>> from kwimage.structs.points import *  # NOQA
>>> s = 128
>>> image = np.zeros((s, s))
>>> self = Points.random(10).scale(s)
>>> image = self.draw_on(image)
>>> # xdoc: +REQUIRES(--show)
>>> import kwplot
>>> kwplot.figure(fnum=1, doclf=True)
>>> kwplot.autompl()
>>> kwplot.imshow(image)
>>> self.draw(radius=3, alpha=.5)
>>> kwplot.show_if_requested()

Example

>>> # xdoc: +REQUIRES(module:kwplot)
>>> from kwimage.structs.points import *  # NOQA
>>> s = 128
>>> image = np.zeros((s, s))
>>> self = Points.random(10).scale(s)
>>> image = self.draw_on(image, radius=3, color='distinct')
>>> # xdoc: +REQUIRES(--show)
>>> import kwplot
>>> kwplot.figure(fnum=1, doclf=True)
>>> kwplot.autompl()
>>> kwplot.imshow(image)
>>> self.draw(radius=3, alpha=.5, color='classes')
>>> kwplot.show_if_requested()

Example

>>> import kwimage
>>> s = 32
>>> self = kwimage.Points.random(10).scale(s)
>>> color = 'blue'
>>> # Test drawong on all channel + dtype combinations
>>> im3 = np.zeros((s, s, 3), dtype=np.float32)
>>> im_chans = {
>>>     'im3': im3,
>>>     'im1': kwimage.convert_colorspace(im3, 'rgb', 'gray'),
>>>     'im4': kwimage.convert_colorspace(im3, 'rgb', 'rgba'),
>>> }
>>> inputs = {}
>>> for k, im in im_chans.items():
>>>     inputs[k + '_01'] = (kwimage.ensure_float01(im.copy()), {'radius': None})
>>>     inputs[k + '_255'] = (kwimage.ensure_uint255(im.copy()), {'radius': None})
>>> outputs = {}
>>> for k, v in inputs.items():
>>>     im, kw = v
>>>     outputs[k] = self.draw_on(im, color=color, **kw)
>>> # xdoc: +REQUIRES(--show)
>>> import kwplot
>>> kwplot.figure(fnum=2, doclf=True)
>>> kwplot.autompl()
>>> pnum_ = kwplot.PlotNums(nCols=2, nRows=len(inputs))
>>> for k in inputs.keys():
>>>     kwplot.imshow(inputs[k][0], fnum=2, pnum=pnum_(), title=k)
>>>     kwplot.imshow(outputs[k], fnum=2, pnum=pnum_(), title=k)
>>> kwplot.show_if_requested()

draw(self, color='blue', ax=None, alpha=None, radius=1, **kwargs)

TODO: can use kwplot.draw_points

Example

>>> # xdoc: +REQUIRES(module:kwplot)
>>> from kwimage.structs.points import *  # NOQA
>>> pts = Points.random(10)
>>> # xdoc: +REQUIRES(--show)
>>> pts.draw(radius=0.01)

>>> from kwimage.structs.points import *  # NOQA
>>> self = Points.random(10, classes=['a', 'b', 'c'])
>>> self.draw(radius=0.01, color='classes')

compress(self, flags, axis=0, inplace=False)

Filters items based on a boolean criterion

Example

>>> from kwimage.structs.points import *  # NOQA
>>> self = Points.random(4)
>>> flags = [1, 0, 1, 1]
>>> other = self.compress(flags)
>>> assert len(self) == 4
>>> assert len(other) == 3

>>> # xdoctest: +REQUIRES(module:torch)
>>> other = self.tensor().compress(flags)
>>> assert len(other) == 3

take(self, indices, axis=0, inplace=False)

Takes a subset of items at specific indices

Example

>>> from kwimage.structs.points import *  # NOQA
>>> self = Points.random(4)
>>> indices = [1, 3]
>>> other = self.take(indices)
>>> assert len(self) == 4
>>> assert len(other) == 2

>>> # xdoctest: +REQUIRES(module:torch)
>>> other = self.tensor().take(indices)
>>> assert len(other) == 2

classmethod concatenate(cls, points, axis=0)

to_coco(self, style='orig')

Converts to an mscoco-like representation

Note

items that are usually id-references to other objects may need to be rectified.

Parameters

style (str) – either orig, new, new-id, or new-name

Returns

mscoco-like representation

Return type

Dict

Example

>>> from kwimage.structs.points import *  # NOQA
>>> self = Points.random(4, classes=['a', 'b'])
>>> orig = self._to_coco(style='orig')
>>> print('orig = {!r}'.format(orig))
>>> new_name = self._to_coco(style='new-name')
>>> print('new_name = {}'.format(ub.repr2(new_name, nl=-1)))
>>> # xdoctest: +REQUIRES(module:ndsampler)
>>> import ndsampler
>>> self.meta['classes'] = ndsampler.CategoryTree.coerce(self.meta['classes'])
>>> new_id = self._to_coco(style='new-id')
>>> print('new_id = {}'.format(ub.repr2(new_id, nl=-1)))

_to_coco(self, style='orig')

See to_coco

classmethod coerce(cls, data)

Attempt to coerce data into a Points object

classmethod _from_coco(cls, coco_kpts, class_idxs=None, classes=None)

classmethod from_coco(cls, coco_kpts, class_idxs=None, classes=None, warn=False)

Parameters

• coco_kpts (list | dict) – either the original list keypoint encoding or the new dict keypoint encoding.

• class_idxs (list) – only needed if using old style

• classes (list | CategoryTree) – list of all keypoint category names

• warn (bool, default=False) – if True raise warnings

Example

>>> ##
>>> classes = ['mouth', 'left-hand', 'right-hand']
>>> coco_kpts = [
>>>     {'xy': (0, 0), 'visible': 2, 'keypoint_category': 'left-hand'},
>>>     {'xy': (1, 2), 'visible': 2, 'keypoint_category': 'mouth'},
>>> ]
>>> Points.from_coco(coco_kpts, classes=classes)
>>> # Test without classes
>>> Points.from_coco(coco_kpts)
>>> # Test without any category info
>>> coco_kpts2 = [ub.dict_diff(d, {'keypoint_category'}) for d in coco_kpts]
>>> Points.from_coco(coco_kpts2)
>>> # Test without category instead of keypoint_category
>>> coco_kpts3 = [ub.map_keys(lambda x: x.replace('keypoint_', ''), d) for d in coco_kpts]
>>> Points.from_coco(coco_kpts3)
>>> #
>>> # Old style
>>> coco_kpts = [0, 0, 2, 0, 1, 2]
>>> Points.from_coco(coco_kpts)
>>> # Fail case
>>> coco_kpts4 = [{'xy': [4686.5, 1341.5], 'category': 'dot'}]
>>> Points.from_coco(coco_kpts4, classes=[])

Example

>>> # xdoctest: +REQUIRES(module:ndsampler)
>>> import ndsampler
>>> classes = ndsampler.CategoryTree.from_coco([
>>>     {'name': 'mouth', 'id': 2}, {'name': 'left-hand', 'id': 3}, {'name': 'right-hand', 'id': 5}
>>> ])
>>> coco_kpts = [
>>>     {'xy': (0, 0), 'visible': 2, 'keypoint_category_id': 5},
>>>     {'xy': (1, 2), 'visible': 2, 'keypoint_category_id': 2},
>>> ]
>>> pts = Points.from_coco(coco_kpts, classes=classes)
>>> assert pts.data['class_idxs'].tolist() == [2, 0]

class kwimage.PointsList

Bases: kwimage.structs._generic.ObjectList

Stores a list of Points, each item usually corresponds to a different object.

Notes

# TODO: when the data is homogenous we can use a more efficient # representation, otherwise we have to use heterogenous storage.

class kwimage.Polygon(data=None, meta=None, datakeys=None, metakeys=None, **kwargs)

Bases: kwimage.structs._generic.Spatial, _PolyArrayBackend, _PolyWarpMixin, ubelt.NiceRepr

Represents a single polygon as set of exterior boundary points and a list of internal polygons representing holes.

By convention exterior boundaries should be counterclockwise and interior holes should be clockwise.

Example

>>> import kwimage
>>> data = {
>>>     'exterior': np.array([[13,  1], [13, 19], [25, 19], [25,  1]]),
>>>     'interiors': [
>>>         np.array([[13, 13], [14, 12], [24, 12], [25, 13], [25, 18],
>>>                   [24, 19], [14, 19], [13, 18]]),
>>>         np.array([[13,  2], [14,  1], [24,  1], [25, 2], [25, 11],
>>>                   [24, 12], [14, 12], [13, 11]])]
>>> }
>>> self = kwimage.Polygon(**data)
>>> # xdoc: +REQUIRES(--show)
>>> import kwplot
>>> kwplot.autompl()
>>> self.draw(setlim=True)

Example

>>> import kwimage
>>> self = kwimage.Polygon.random(
>>>     n=5, n_holes=1, convex=False, rng=0)
>>> print('self = {}'.format(self))
self = <Polygon({
    'exterior': <Coords(data=
                    array([[0.30371392, 0.97195856],
                           [0.24372304, 0.60568445],
                           [0.21408694, 0.34884262],
                           [0.5799477 , 0.44020379],
                           [0.83720288, 0.78367234]]))>,
    'interiors': [<Coords(data=
                     array([[0.50164209, 0.83520279],
                            [0.25835064, 0.40313428],
                            [0.28778562, 0.74758761],
                            [0.30341266, 0.93748088]]))>],
})>
>>> # xdoc: +REQUIRES(--show)
>>> import kwplot
>>> kwplot.autompl()
>>> self.draw(setlim=True)

__datakeys__ = ['exterior', 'interiors']

__metakeys__ = ['classes']

property exterior(self)

property interiors(self)

__nice__(self)

classmethod circle(cls, xy, r, resolution=64)

Create a circular polygon

Example

>>> xy = (0.5, 0.5)
>>> r = .3
>>> poly = Polygon.circle(xy, r)

classmethod random(cls, n=6, n_holes=0, convex=True, tight=False, rng=None)

Parameters

• n (int) – number of points in the polygon (must be 3 or more)

• n_holes (int) – number of holes

• tight (bool, default=False) – fits the minimum and maximum points between 0 and 1

• convex (bool, default=True) – force resulting polygon will be convex (may remove exterior points)

CommandLine:

xdoctest -m kwimage.structs.polygon Polygon.random

Example

>>> rng = None
>>> n = 4
>>> n_holes = 1
>>> cls = Polygon
>>> self = Polygon.random(n=n, rng=rng, n_holes=n_holes, convex=1)
>>> # xdoc: +REQUIRES(--show)
>>> import kwplot
>>> kwplot.figure(fnum=1, doclf=True)
>>> kwplot.autompl()
>>> self.draw()

References

https://gis.stackexchange.com/questions/207731/random-multipolygon https://stackoverflow.com/questions/8997099/random-polygon https://stackoverflow.com/questions/27548363/from-voronoi-tessellation-to-shapely-polygons https://stackoverflow.com/questions/8997099/algorithm-to-generate-random-2d-polygon

_impl(self)

to_mask(self, dims=None)

Convert this polygon to a mask

Todo

• [ ] currently not efficient

Parameters

dims (Tuple) – height and width of the output mask

Returns

kwimage.Mask

Example

>>> from kwimage.structs.polygon import *  # NOQA
>>> self = Polygon.random(n_holes=1).scale(128)
>>> mask = self.to_mask((128, 128))
>>> # xdoc: +REQUIRES(--show)
>>> import kwplot
>>> kwplot.autompl()
>>> kwplot.figure(fnum=1, doclf=True)
>>> mask.draw(color='blue')
>>> mask.to_multi_polygon().draw(color='red', alpha=.5)

to_relative_mask(self)

Returns a translated mask such the mask dimensions are minimal.

In other words, we move the polygon all the way to the top-left and return a mask just big enough to fit the polygon.

Returns

kwimage.Mask

Example

>>> from kwimage.structs.polygon import *  # NOQA
>>> self = Polygon.random().scale(8).translate(100, 100)
>>> mask = self.to_relative_mask()
>>> assert mask.shape <= (8, 8)
>>> # xdoc: +REQUIRES(--show)
>>> import kwplot
>>> kwplot.autompl()
>>> kwplot.figure(fnum=1, doclf=True)
>>> mask.draw(color='blue')
>>> mask.to_multi_polygon().draw(color='red', alpha=.5)

fill(self, image, value=1)

Inplace fill in an image based on this polyon.

Parameters

• image (ndarray) – image to draw on

• value (int | Tuple[int], default=1) – value fill in with

Returns

the image that has been modified in place

Return type

ndarray

_to_cv_countours(self)

OpenCV polygon representation, which is a list of points. Holes are implicitly represented. When another polygon is drawn over an existing polyon via cv2.fillPoly

Returns

where each ndarray is of shape [N, 1, 2],

where N is the number of points on the boundary, the middle dimension is always 1, and the trailing dimension represents x and y coordinates respectively.

Return type

List[ndarray]

classmethod coerce(Polygon, data)

Try to autodetermine format of input polygon and coerce it into a kwimage.Polygon.

Parameters

data (object) – some type of data that can be interpreted as a polygon.

Returns

kwimage.Polygon

Example

>>> import kwimage
>>> self = kwimage.Polygon.random()
>>> self.coerce(self)
>>> self.coerce(self.exterior)
>>> self.coerce(self.exterior.data)
>>> self.coerce(self.data)
>>> self.coerce(self.to_geojson())

classmethod from_shapely(Polygon, geom)

Convert a shapely polygon to a kwimage.Polygon

Parameters

geom (shapely.geometry.polygon.Polygon) – a shapely polygon

Returns

kwimage.Polygon

classmethod from_wkt(Polygon, data)

Convert a WKT string to a kwimage.Polygon

Parameters

data (str) – a WKT polygon string

Returns

kwimage.Polygon

Example

>>> import kwimage
>>> data = 'POLYGON ((0.11 0.61, 0.07 0.588, 0.015 0.50, 0.11 0.61))'
>>> self = kwimage.Polygon.from_wkt(data)
>>> assert len(self.exterior) == 4

classmethod from_geojson(Polygon, data_geojson)

Convert a geojson polygon to a kwimage.Polygon

Parameters

data_geojson (dict) – geojson data

References

https://geojson.org/geojson-spec.html

Example

>>> from kwimage.structs.polygon import *  # NOQA
>>> self = Polygon.random(n_holes=2)
>>> data_geojson = self.to_geojson()
>>> new = Polygon.from_geojson(data_geojson)

to_shapely(self)

Example

>>> # xdoc: +REQUIRES(module:kwplot)
>>> # xdoc: +REQUIRES(module:shapely)
>>> from kwimage.structs.polygon import *  # NOQA
>>> self = Polygon.random(n_holes=1)
>>> self = self.scale(100)
>>> geom = self.to_shapely()
>>> print('geom = {!r}'.format(geom))

to_geojson(self)

Converts polygon to a geojson structure

Returns

Dict[str, object]

Example

>>> import kwimage
>>> self = kwimage.Polygon.random()
>>> print(self.to_geojson())

to_wkt(self)

Convert a kwimage.Polygon to WKT string

Example

>>> import kwimage
>>> self = kwimage.Polygon.random()
>>> print(self.to_wkt())

classmethod from_coco(cls, data, dims=None)

Accepts either new-style or old-style coco polygons

_to_coco(self, style='orig')

to_coco(self, style='orig')

Returns

coco-style polygons

Return type

List | Dict

to_multi_polygon(self)

to_boxes(self)

Deprecated: lossy conversion use ‘bounding_box’ instead

property centroid(self)

bounding_box(self)

Returns an axis-aligned bounding box for the segmentation

Returns

kwimage.Boxes

bounding_box_polygon(self)

Returns an axis-aligned bounding polygon for the segmentation.

Notes

This Polygon will be a Box, not a convex hull! Use shapely for convex hulls.

Returns

kwimage.Polygon

copy(self)

clip(self, x_min, y_min, x_max, y_max, inplace=False)

Clip polygon to image boundaries.

Example

>>> from kwimage.structs.polygon import *
>>> self = Polygon.random().scale(10).translate(-1)
>>> self2 = self.clip(1, 1, 3, 3)
>>> # xdoc: +REQUIRES(--show)
>>> import kwplot
>>> kwplot.autompl()
>>> self2.draw(setlim=True)

draw_on(self, image, color='blue', fill=True, border=False, alpha=1.0, copy=False)

Rasterizes a polygon on an image. See draw for a vectorized matplotlib version.

Parameters

• image (ndarray) – image to raster polygon on.

• color (str | tuple) – data coercable to a color

• fill (bool, default=True) – draw the center mass of the polygon

• border (bool, default=False) – draw the border of the polygon

• alpha (float, default=1.0) – polygon transparency (setting alpha < 1 makes this function much slower).

• copy (bool, default=False) – if False only copies if necessary

Example

>>> # xdoc: +REQUIRES(module:kwplot)
>>> from kwimage.structs.polygon import *  # NOQA
>>> self = Polygon.random(n_holes=1).scale(128)
>>> image = np.zeros((128, 128), dtype=np.float32)
>>> image = self.draw_on(image)
>>> # xdoc: +REQUIRES(--show)
>>> import kwplot
>>> kwplot.autompl()
>>> kwplot.imshow(image, fnum=1)

Example

>>> import kwimage
>>> color = 'blue'
>>> self = kwimage.Polygon.random(n_holes=1).scale(128)
>>> image = np.zeros((128, 128), dtype=np.float32)
>>> # Test drawong on all channel + dtype combinations
>>> im3 = np.random.rand(128, 128, 3)
>>> im_chans = {
>>>     'im3': im3,
>>>     'im1': kwimage.convert_colorspace(im3, 'rgb', 'gray'),
>>>     'im4': kwimage.convert_colorspace(im3, 'rgb', 'rgba'),
>>> }
>>> inputs = {}
>>> for k, im in im_chans.items():
>>>     inputs[k + '_01'] = (kwimage.ensure_float01(im.copy()), {'alpha': None})
>>>     inputs[k + '_255'] = (kwimage.ensure_uint255(im.copy()), {'alpha': None})
>>>     inputs[k + '_01_a'] = (kwimage.ensure_float01(im.copy()), {'alpha': 0.5})
>>>     inputs[k + '_255_a'] = (kwimage.ensure_uint255(im.copy()), {'alpha': 0.5})
>>> outputs = {}
>>> for k, v in inputs.items():
>>>     im, kw = v
>>>     outputs[k] = self.draw_on(im, color=color, **kw)
>>> # xdoc: +REQUIRES(--show)
>>> import kwplot
>>> kwplot.figure(fnum=2, doclf=True)
>>> kwplot.autompl()
>>> pnum_ = kwplot.PlotNums(nCols=2, nRows=len(inputs))
>>> for k in inputs.keys():
>>>     kwplot.imshow(inputs[k][0], fnum=2, pnum=pnum_(), title=k)
>>>     kwplot.imshow(outputs[k], fnum=2, pnum=pnum_(), title=k)
>>> kwplot.show_if_requested()

draw(self, color='blue', ax=None, alpha=1.0, radius=1, setlim=False, border=False, linewidth=2)

Draws polygon in a matplotlib axes. See draw_on for in-memory image modification.

Parameters

• setlim (bool) – if True ensures the limits of the axes contains the polygon

• color (str | Tuple) – coercable color

• alpha (float) – fill transparency

• setlim (bool) – if True, modify the x and y limits of the matplotlib axes such that the polygon is can be seen.

• border (bool, default=False) – if True, draws an edge border on the polygon.

• linewidth (bool) – width of the border

Todo

• [ ] Rework arguments in favor of matplotlib standards

Example

>>> # xdoc: +REQUIRES(module:kwplot)
>>> from kwimage.structs.polygon import *  # NOQA
>>> self = Polygon.random(n_holes=1)
>>> self = self.scale(100)
>>> # xdoc: +REQUIRES(--show)
>>> self.draw()
>>> import kwplot
>>> kwplot.autompl()
>>> from matplotlib import pyplot as plt
>>> kwplot.figure(fnum=2)
>>> self.draw(setlim=True)

_ensure_vertex_order(self, inplace=False)

Fixes vertex ordering so the exterior ring is CCW and the interior rings are CW.

Example

>>> import kwimage
>>> self = kwimage.Polygon.random(n=3, n_holes=2, rng=0)
>>> print('self = {!r}'.format(self))
>>> new = self._ensure_vertex_order()
>>> print('new = {!r}'.format(new))

>>> self = kwimage.Polygon.random(n=3, n_holes=2, rng=0).swap_axes()
>>> print('self = {!r}'.format(self))
>>> new = self._ensure_vertex_order()
>>> print('new = {!r}'.format(new))

class kwimage.PolygonList

Bases: kwimage.structs._generic.ObjectList

Stores and allows manipluation of multiple polygons, usually within the same image.

to_mask_list(self, dims=None)

Converts all items to masks

to_polygon_list(self)

to_segmentation_list(self)

Converts all items to segmentation objects

swap_axes(self, inplace=False)

to_geojson(self, as_collection=False)

Converts a list of polygons/multipolygons to a geojson structure

Parameters

as_collection (bool) – if True, wraps the polygon geojson items in a geojson feature collection, otherwise just return a list of items.

Returns

items or geojson data

Return type

List[Dict] | Dict

Example

>>> import kwimage
>>> data = [kwimage.Polygon.random(),
>>>         kwimage.Polygon.random(n_holes=1),
>>>         kwimage.MultiPolygon.random(n_holes=1),
>>>         kwimage.MultiPolygon.random()]
>>> self = kwimage.PolygonList(data)
>>> geojson = self.to_geojson(as_collection=True)
>>> items = self.to_geojson(as_collection=False)
>>> print('geojson = {}'.format(ub.repr2(geojson, nl=-2, precision=1)))
>>> print('items = {}'.format(ub.repr2(items, nl=-2, precision=1)))

class kwimage.Segmentation(data, format=None)

Bases: _WrapperObject

Either holds a MultiPolygon, Polygon, or Mask

Parameters

• data (object) – the underlying object

• format (str) – either ‘mask’, ‘polygon’, or ‘multipolygon’

classmethod random(cls, rng=None)

Example

>>> self = Segmentation.random()
>>> print('self = {!r}'.format(self))
>>> # xdoc: +REQUIRES(--show)
>>> import kwplot
>>> kwplot.autompl()
>>> kwplot.figure(fnum=1, doclf=True)
>>> self.draw()
>>> kwplot.show_if_requested()

to_multi_polygon(self)

to_mask(self, dims=None)

property meta(self)

classmethod coerce(cls, data, dims=None)

class kwimage.SegmentationList

Bases: kwimage.structs._generic.ObjectList

Store and manipulate multiple segmentations (masks or polygons), usually within the same image

to_polygon_list(self)

Converts all mask objects to multi-polygon objects

to_mask_list(self, dims=None)

Converts all mask objects to multi-polygon objects

to_segmentation_list(self)

classmethod coerce(cls, data)

Interpret data as a list of Segmentations

kwimage.smooth_prob(prob, k=3, inplace=False, eps=1e-09)

Smooths the probability map, but preserves the magnitude of the peaks.

Notes

even if inplace is true, we still need to make a copy of the input array, however, we do ensure that it is cleaned up before we leave the function scope.

sigma=0.8 @ k=3, sigma=1.1 @ k=5, sigma=1.4 @ k=7

class kwimage.Affine(matrix)

Bases: Projective

Helper for making affine transform matrices.

Example

>>> self = Affine(np.eye(3))
>>> m1 = np.eye(3) @ self
>>> m2 = self @ np.eye(3)

Example

>>> from kwimage.transform import *  # NOQA
>>> m = {}
>>> # Works, and returns a Affine
>>> m[len(m)] = x = Affine.random() @ np.eye(3)
>>> assert isinstance(x, Affine)
>>> m[len(m)] = x = Affine.random() @ None
>>> assert isinstance(x, Affine)
>>> # Works, and returns an ndarray
>>> m[len(m)] = x = np.eye(3) @ Affine.random(3)
>>> assert isinstance(x, np.ndarray)
>>> # Works, and returns an Matrix
>>> m[len(m)] = x = Affine.random() @ Matrix.random(3)
>>> assert isinstance(x, Matrix)
>>> m[len(m)] = x = Matrix.random(3) @ Affine.random()
>>> assert isinstance(x, Matrix)
>>> print('m = {}'.format(ub.repr2(m)))

property shape(self)

__getitem__(self, index)

__json__(self)

concise(self)

Return a concise coercable dictionary representation of this matrix

Returns

a small serializable dict that can be passed

to Affine.coerce() to reconstruct this object.

Return type

Dict[str, object]

Returns

dictionary with consise parameters

Return type

Dict

Example

>>> self = Affine.random(rng=0, scale=1)
>>> params = self.concise()
>>> assert np.allclose(Affine.coerce(params).matrix, self.matrix)
>>> print('params = {}'.format(ub.repr2(params, nl=1, precision=2)))
params = {
    'offset': (0.08, 0.38),
    'theta': 0.08,
    'type': 'affine',
}

Example

>>> self = Affine.random(rng=0, scale=2, offset=0)
>>> params = self.concise()
>>> assert np.allclose(Affine.coerce(params).matrix, self.matrix)
>>> print('params = {}'.format(ub.repr2(params, nl=1, precision=2)))
params = {
    'scale': 2.00,
    'theta': 0.04,
    'type': 'affine',
}

classmethod coerce(cls, data=None, **kwargs)

Attempt to coerce the data into an affine object

Parameters

• data – some data we attempt to coerce to an Affine matrix

• **kwargs – some data we attempt to coerce to an Affine matrix, mutually exclusive with data.

Returns

Affine

Example

>>> import kwimage
>>> kwimage.Affine.coerce({'type': 'affine', 'matrix': [[1, 0, 0], [0, 1, 0]]})
>>> kwimage.Affine.coerce({'scale': 2})
>>> kwimage.Affine.coerce({'offset': 3})
>>> kwimage.Affine.coerce(np.eye(3))
>>> kwimage.Affine.coerce(None)
>>> kwimage.Affine.coerce(skimage.transform.AffineTransform(scale=30))

decompose(self)

Decompose the affine matrix into its individual scale, translation, rotation, and skew parameters.

Returns

decomposed offset, scale, theta, and shear params

Return type

Dict

References

https://math.stackexchange.com/questions/612006/decompose-affine

Example

>>> self = Affine.random()
>>> params = self.decompose()
>>> recon = Affine.coerce(**params)
>>> params2 = recon.decompose()
>>> pt = np.vstack([np.random.rand(2, 1), [1]])
>>> result1 = self.matrix[0:2] @ pt
>>> result2 = recon.matrix[0:2] @ pt
>>> assert np.allclose(result1, result2)

>>> self = Affine.scale(0.001) @ Affine.random()
>>> params = self.decompose()
>>> self.det()

classmethod scale(cls, scale)

Create a scale Affine object

Parameters

scale (float | Tuple[float, float]) – x, y scale factor

Returns

Affine

classmethod translate(cls, offset)

Create a translation Affine object

Parameters

offset (float | Tuple[float, float]) – x, y translation factor

Returns

Affine

classmethod rotate(cls, theta)

Create a rotation Affine object

Parameters

theta (float) – counter-clockwise rotation angle in radians

Returns

Affine

classmethod random(cls, rng=None, **kw)

Create a random Affine object

Parameters

• rng – random number generator

• **kw – passed to Affine.random_params(). can contain coercable random distributions for scale, offset, about, theta, and shear.

Returns

Affine

classmethod random_params(cls, rng=None, **kw)

Parameters

• rng – random number generator

• **kw – can contain coercable random distributions for scale, offset, about, theta, and shear.

Returns

affine parameters suitable to be passed to Affine.affine

Return type

Dict

Todo

• [ ] improve kwargs parameterization

classmethod affine(cls, scale=None, offset=None, theta=None, shear=None, about=None)

Create an affine matrix from high-level parameters

Parameters

• scale (float | Tuple[float, float]) – x, y scale factor

• offset (float | Tuple[float, float]) – x, y translation factor

• theta (float) – counter-clockwise rotation angle in radians

• shear (float) – counter-clockwise shear angle in radians

• about (float | Tuple[float, float]) – x, y location of the origin

Returns

the constructed Affine object

Return type

Affine

Example

>>> rng = kwarray.ensure_rng(None)
>>> scale = rng.randn(2) * 10
>>> offset = rng.randn(2) * 10
>>> about = rng.randn(2) * 10
>>> theta = rng.randn() * 10
>>> shear = rng.randn() * 10
>>> # Create combined matrix from all params
>>> F = Affine.affine(
>>>     scale=scale, offset=offset, theta=theta, shear=shear,
>>>     about=about)
>>> # Test that combining components matches
>>> S = Affine.affine(scale=scale)
>>> T = Affine.affine(offset=offset)
>>> R = Affine.affine(theta=theta)
>>> H = Affine.affine(shear=shear)
>>> O = Affine.affine(offset=about)
>>> # combine (note shear must be on the RHS of rotation)
>>> alt  = O @ T @ R @ H @ S @ O.inv()
>>> print('F    = {}'.format(ub.repr2(F.matrix.tolist(), nl=1)))
>>> print('alt  = {}'.format(ub.repr2(alt.matrix.tolist(), nl=1)))
>>> assert np.all(np.isclose(alt.matrix, F.matrix))
>>> pt = np.vstack([np.random.rand(2, 1), [[1]]])
>>> warp_pt1 = (F.matrix @ pt)
>>> warp_pt2 = (alt.matrix @ pt)
>>> assert np.allclose(warp_pt2, warp_pt1)

Sympy:

>>> # xdoctest: +SKIP
>>> import sympy
>>> # Shows the symbolic construction of the code
>>> # https://groups.google.com/forum/#!topic/sympy/k1HnZK_bNNA
>>> from sympy.abc import theta
>>> x0, y0, sx, sy, theta, shear, tx, ty = sympy.symbols(
>>>     'x0, y0, sx, sy, theta, shear, tx, ty')
>>> # move the center to 0, 0
>>> tr1_ = np.array([[1, 0,  -x0],
>>>                  [0, 1,  -y0],
>>>                  [0, 0,    1]])
>>> # Define core components of the affine transform
>>> S = np.array([  # scale
>>>     [sx,  0, 0],
>>>     [ 0, sy, 0],
>>>     [ 0,  0, 1]])
>>> H = np.array([  # shear
>>>     [1, -sympy.sin(shear), 0],
>>>     [0,  sympy.cos(shear), 0],
>>>     [0,                 0, 1]])
>>> R = np.array([  # rotation
>>>     [sympy.cos(theta), -sympy.sin(theta), 0],
>>>     [sympy.sin(theta),  sympy.cos(theta), 0],
>>>     [               0,                 0, 1]])
>>> T = np.array([  # translation
>>>     [ 1,  0, tx],
>>>     [ 0,  1, ty],
>>>     [ 0,  0,  1]])
>>> # Contruct the affine 3x3 about the origin
>>> aff0 = np.array(sympy.simplify(T @ R @ H @ S))
>>> # move 0, 0 back to the specified origin
>>> tr2_ = np.array([[1, 0,  x0],
>>>                  [0, 1,  y0],
>>>                  [0, 0,   1]])
>>> # combine transformations
>>> aff = tr2_ @ aff0 @ tr1_
>>> print('aff = {}'.format(ub.repr2(aff.tolist(), nl=1)))

class kwimage.Linear(matrix)

Bases: Matrix

Base class for matrix-based transform.

Example

>>> from kwimage.transform import *  # NOQA
>>> ms = {}
>>> ms['random()'] = Matrix.random()
>>> ms['eye()'] = Matrix.eye()
>>> ms['random(3)'] = Matrix.random(3)
>>> ms['random(4, 4)'] = Matrix.random(4, 4)
>>> ms['eye(3)'] = Matrix.eye(3)
>>> ms['explicit'] = Matrix(np.array([[1.618]]))
>>> for k, m in ms.items():
>>>     print('----')
>>>     print(f'{k} = {m}')
>>>     print(f'{k}.inv() = {m.inv()}')
>>>     print(f'{k}.T = {m.T}')
>>>     print(f'{k}.det() = {m.det()}')

class kwimage.Matrix(matrix)

Bases: Transform

Base class for matrix-based transform.

Example

>>> from kwimage.transform import *  # NOQA
>>> ms = {}
>>> ms['random()'] = Matrix.random()
>>> ms['eye()'] = Matrix.eye()
>>> ms['random(3)'] = Matrix.random(3)
>>> ms['random(4, 4)'] = Matrix.random(4, 4)
>>> ms['eye(3)'] = Matrix.eye(3)
>>> ms['explicit'] = Matrix(np.array([[1.618]]))
>>> for k, m in ms.items():
>>>     print('----')
>>>     print(f'{k} = {m}')
>>>     print(f'{k}.inv() = {m.inv()}')
>>>     print(f'{k}.T = {m.T}')
>>>     print(f'{k}.det() = {m.det()}')

__nice__(self)

__repr__(self)

Return repr(self).

property shape(self)

__json__(self)

classmethod coerce(cls, data=None, **kwargs)

Example

>>> Matrix.coerce({'type': 'matrix', 'matrix': [[1, 0, 0], [0, 1, 0]]})
>>> Matrix.coerce(np.eye(3))
>>> Matrix.coerce(None)

__array__(self)

Allow this object to be passed to np.asarray

References

https://numpy.org/doc/stable/user/basics.dispatch.html

__imatmul__(self, other)

__matmul__(self, other)

Example

>>> m = {}
>>> # Works, and returns a Matrix
>>> m[len(m)] = x = Matrix.random() @ np.eye(2)
>>> assert isinstance(x, Matrix)
>>> m[len(m)] = x = Matrix.random() @ None
>>> assert isinstance(x, Matrix)
>>> # Works, and returns an ndarray
>>> m[len(m)] = x = np.eye(3) @ Matrix.random(3)
>>> assert isinstance(x, np.ndarray)
>>> # These do not work
>>> # m[len(m)] = None @ Matrix.random()
>>> # m[len(m)] = np.eye(3) @ None
>>> print('m = {}'.format(ub.repr2(m)))

inv(self)

Returns the inverse of this matrix

Returns

Matrix

property T(self)

Transpose the underlying matrix

det(self)

Compute the determinant of the underlying matrix

Returns

float

classmethod eye(cls, shape=None, rng=None)

Construct an identity

classmethod random(cls, shape=None, rng=None)

class kwimage.Projective(matrix)

Bases: Linear

Currently just a stub class that may be used to implement projective / homography transforms in the future.

class kwimage.Transform

Bases: ubelt.NiceRepr

Inherit from this class and define __nice__ to “nicely” print your objects.

Defines __str__ and __repr__ in terms of __nice__ function Classes that inherit from NiceRepr should redefine __nice__. If the inheriting class has a __len__, method then the default __nice__ method will return its length.

Example

>>> import ubelt as ub
>>> class Foo(ub.NiceRepr):
...    def __nice__(self):
...        return 'info'
>>> foo = Foo()
>>> assert str(foo) == '<Foo(info)>'
>>> assert repr(foo).startswith('<Foo(info) at ')

Example

>>> import ubelt as ub
>>> class Bar(ub.NiceRepr):
...    pass
>>> bar = Bar()
>>> import pytest
>>> with pytest.warns(RuntimeWarning) as record:
>>>     assert 'object at' in str(bar)
>>>     assert 'object at' in repr(bar)

Example

>>> import ubelt as ub
>>> class Baz(ub.NiceRepr):
...    def __len__(self):
...        return 5
>>> baz = Baz()
>>> assert str(baz) == '<Baz(5)>'

Example

>>> import ubelt as ub
>>> # If your nice message has a bug, it shouldn't bring down the house
>>> class Foo(ub.NiceRepr):
...    def __nice__(self):
...        assert False
>>> foo = Foo()
>>> import pytest
>>> with pytest.warns(RuntimeWarning) as record:
>>>     print('foo = {!r}'.format(foo))
foo = <...Foo ...>

kwimage.add_homog(pts)

Add a homogenous coordinate to a point array

This is a convinience function, it is not particularly efficient.

SeeAlso:

cv2.convertPointsToHomogeneous

Example

>>> pts = np.random.rand(10, 2)
>>> add_homog(pts)

Benchmark:

>>> import timerit
>>> ti = timerit.Timerit(1000, bestof=10, verbose=2)
>>> pts = np.random.rand(1000, 2)
>>> for timer in ti.reset('kwimage'):
>>>     with timer:
>>>         kwimage.add_homog(pts)
>>> for timer in ti.reset('cv2'):
>>>     with timer:
>>>         cv2.convertPointsToHomogeneous(pts)
>>> # cv2 is 4x faster, but has more restrictive inputs

kwimage.remove_homog(pts, mode='divide')

Remove homogenous coordinate to a point array.

This is a convinience function, it is not particularly efficient.

SeeAlso:

cv2.convertPointsFromHomogeneous

Example

>>> homog_pts = np.random.rand(10, 3)
>>> remove_homog(homog_pts, 'divide')
>>> remove_homog(homog_pts, 'drop')

kwimage.subpixel_accum(dst, src, index, interp_axes=None)

Add the source values array into the destination array at a particular subpixel index.

Parameters

• dst (ArrayLike) – destination accumulation array

• src (ArrayLike) – source array containing values to add

• index (Tuple[slice]) – subpixel slice into dst that corresponds with src

• interp_axes (tuple) – specify which axes should be spatially interpolated

Notes

Inputs:

+—+—+—+—+—+ dst.shape = (5,)

+—+—+ src.shape = (2,) |=======| index = 1.5:3.5

Subpixel shift the source by -0.5. When the index is non-integral, pad the aligned src with an extra value to ensure all dst pixels that would be influenced by the smaller subpixel shape are influenced by the aligned src. Note that we are not scaling.

+—+—+—+ aligned_src.shape = (3,) |===========| aligned_index = 1:4

Example

>>> dst = np.zeros(5)
>>> src = np.ones(2)
>>> index = [slice(1.5, 3.5)]
>>> subpixel_accum(dst, src, index)
>>> print(ub.repr2(dst, precision=2, with_dtype=0))
np.array([0. , 0.5, 1. , 0.5, 0. ])

Example

>>> dst = np.zeros((6, 6))
>>> src = np.ones((3, 3))
>>> index = (slice(1.5, 4.5), slice(1, 4))
>>> subpixel_accum(dst, src, index)
>>> print(ub.repr2(dst, precision=2, with_dtype=0))
np.array([[0. , 0. , 0. , 0. , 0. , 0. ],
          [0. , 0.5, 0.5, 0.5, 0. , 0. ],
          [0. , 1. , 1. , 1. , 0. , 0. ],
          [0. , 1. , 1. , 1. , 0. , 0. ],
          [0. , 0.5, 0.5, 0.5, 0. , 0. ],
          [0. , 0. , 0. , 0. , 0. , 0. ]])
>>> # xdoctest: +REQUIRES(module:torch)
>>> dst = torch.zeros((1, 3, 6, 6))
>>> src = torch.ones((1, 3, 3, 3))
>>> index = (slice(None), slice(None), slice(1.5, 4.5), slice(1.25, 4.25))
>>> subpixel_accum(dst, src, index)
>>> print(ub.repr2(dst.numpy()[0, 0], precision=2, with_dtype=0))
np.array([[0.  , 0.  , 0.  , 0.  , 0.  , 0.  ],
          [0.  , 0.38, 0.5 , 0.5 , 0.12, 0.  ],
          [0.  , 0.75, 1.  , 1.  , 0.25, 0.  ],
          [0.  , 0.75, 1.  , 1.  , 0.25, 0.  ],
          [0.  , 0.38, 0.5 , 0.5 , 0.12, 0.  ],
          [0.  , 0.  , 0.  , 0.  , 0.  , 0.  ]])

Doctest:

>>> # TODO: move to a unit test file
>>> subpixel_accum(np.zeros(5), np.ones(2), [slice(1.5, 3.5)]).tolist()
[0.0, 0.5, 1.0, 0.5, 0.0]
>>> subpixel_accum(np.zeros(5), np.ones(2), [slice(0, 2)]).tolist()
[1.0, 1.0, 0.0, 0.0, 0.0]
>>> subpixel_accum(np.zeros(5), np.ones(3), [slice(.5, 3.5)]).tolist()
[0.5, 1.0, 1.0, 0.5, 0.0]
>>> subpixel_accum(np.zeros(5), np.ones(3), [slice(-1, 2)]).tolist()
[1.0, 1.0, 0.0, 0.0, 0.0]
>>> subpixel_accum(np.zeros(5), np.ones(3), [slice(-1.5, 1.5)]).tolist()
[1.0, 0.5, 0.0, 0.0, 0.0]
>>> subpixel_accum(np.zeros(5), np.ones(3), [slice(10, 13)]).tolist()
[0.0, 0.0, 0.0, 0.0, 0.0]
>>> subpixel_accum(np.zeros(5), np.ones(3), [slice(3.25, 6.25)]).tolist()
[0.0, 0.0, 0.0, 0.75, 1.0]
>>> subpixel_accum(np.zeros(5), np.ones(3), [slice(4.9, 7.9)]).tolist()
[0.0, 0.0, 0.0, 0.0, 0.099...]
>>> subpixel_accum(np.zeros(5), np.ones(9), [slice(-1.5, 7.5)]).tolist()
[1.0, 1.0, 1.0, 1.0, 1.0]
>>> subpixel_accum(np.zeros(5), np.ones(9), [slice(2.625, 11.625)]).tolist()
[0.0, 0.0, 0.375, 1.0, 1.0]
>>> subpixel_accum(np.zeros(5), 1, [slice(2.625, 11.625)]).tolist()
[0.0, 0.0, 0.375, 1.0, 1.0]

kwimage.subpixel_align(dst, src, index, interp_axes=None)

Returns an aligned version of the source tensor and destination index.

Used as the backend to implement other subpixel functions like:

subpixel_accum, subpixel_maximum.

kwimage.subpixel_getvalue(img, pts, coord_axes=None, interp='bilinear', bordermode='edge')

Get values at subpixel locations

Parameters

• img (ArrayLike) – image to sample from

• pts (ArrayLike) – subpixel rc-coordinates to sample

• coord_axes (Sequence, default=None) – axes to perform interpolation on, if not specified the first d axes are interpolated, where d=pts.shape[-1]. IE: this indicates which axes each coordinate dimension corresponds to.

• interp (str) – interpolation mode

• bordermode (str) – how locations outside the image are handled

Example

>>> from kwimage.util_warp import *  # NOQA
>>> img = np.arange(3 * 3).reshape(3, 3)
>>> pts = np.array([[1, 1], [1.5, 1.5], [1.9, 1.1]])
>>> subpixel_getvalue(img, pts)
array([4. , 6. , 6.8])
>>> subpixel_getvalue(img, pts, coord_axes=(1, 0))
array([4. , 6. , 5.2])
>>> # xdoctest: +REQUIRES(module:torch)
>>> img = torch.Tensor(img)
>>> pts = torch.Tensor(pts)
>>> subpixel_getvalue(img, pts)
tensor([4.0000, 6.0000, 6.8000])
>>> subpixel_getvalue(img.numpy(), pts.numpy(), interp='nearest')
array([4., 8., 7.], dtype=float32)
>>> subpixel_getvalue(img.numpy(), pts.numpy(), interp='nearest', coord_axes=[1, 0])
array([4., 8., 5.], dtype=float32)
>>> subpixel_getvalue(img, pts, interp='nearest')
tensor([4., 8., 7.])

References

stackoverflow.com/uestions/12729228/simple-binlin-interp-images-numpy

SeeAlso:

cv2.getRectSubPix(image, patchSize, center[, patch[, patchType]])

kwimage.subpixel_maximum(dst, src, index, interp_axes=None)

Take the max of the source values array into and the destination array at a particular subpixel index. Modifies the destination array.

Parameters

• dst (ArrayLike) – destination array to index into

• src (ArrayLike) – source array that agrees with the index

• index (Tuple[slice]) – subpixel slice into dst that corresponds with src

• interp_axes (tuple) – specify which axes should be spatially interpolated

Example

>>> dst = np.array([0, 1.0, 1.0, 1.0, 0])
>>> src = np.array([2.0, 2.0])
>>> index = [slice(1.6, 3.6)]
>>> subpixel_maximum(dst, src, index)
>>> print(ub.repr2(dst, precision=2, with_dtype=0))
np.array([0. , 1. , 2. , 1.2, 0. ])

Example

>>> # xdoctest: +REQUIRES(module:torch)
>>> dst = torch.zeros((1, 3, 5, 5)) + .5
>>> src = torch.ones((1, 3, 3, 3))
>>> index = (slice(None), slice(None), slice(1.4, 4.4), slice(1.25, 4.25))
>>> subpixel_maximum(dst, src, index)
>>> print(ub.repr2(dst.numpy()[0, 0], precision=2, with_dtype=0))
np.array([[0.5 , 0.5 , 0.5 , 0.5 , 0.5 ],
          [0.5 , 0.5 , 0.6 , 0.6 , 0.5 ],
          [0.5 , 0.75, 1.  , 1.  , 0.5 ],
          [0.5 , 0.75, 1.  , 1.  , 0.5 ],
          [0.5 , 0.5 , 0.5 , 0.5 , 0.5 ]])

kwimage.subpixel_minimum(dst, src, index, interp_axes=None)

Take the min of the source values array into and the destination array at a particular subpixel index. Modifies the destination array.

Parameters

• dst (ArrayLike) – destination array to index into

• src (ArrayLike) – source array that agrees with the index

• index (Tuple[slice]) – subpixel slice into dst that corresponds with src

• interp_axes (tuple) – specify which axes should be spatially interpolated

Example

>>> dst = np.array([0, 1.0, 1.0, 1.0, 0])
>>> src = np.array([2.0, 2.0])
>>> index = [slice(1.6, 3.6)]
>>> subpixel_minimum(dst, src, index)
>>> print(ub.repr2(dst, precision=2, with_dtype=0))
np.array([0. , 0.8, 1. , 1. , 0. ])

Example

>>> # xdoctest: +REQUIRES(module:torch)
>>> dst = torch.zeros((1, 3, 5, 5)) + .5
>>> src = torch.ones((1, 3, 3, 3))
>>> index = (slice(None), slice(None), slice(1.4, 4.4), slice(1.25, 4.25))
>>> subpixel_minimum(dst, src, index)
>>> print(ub.repr2(dst.numpy()[0, 0], precision=2, with_dtype=0))
np.array([[0.5 , 0.5 , 0.5 , 0.5 , 0.5 ],
          [0.5 , 0.45, 0.5 , 0.5 , 0.15],
          [0.5 , 0.5 , 0.5 , 0.5 , 0.25],
          [0.5 , 0.5 , 0.5 , 0.5 , 0.25],
          [0.5 , 0.3 , 0.4 , 0.4 , 0.1 ]])

kwimage.subpixel_set(dst, src, index, interp_axes=None)

Add the source values array into the destination array at a particular subpixel index.

Parameters

• dst (ArrayLike) – destination accumulation array

• src (ArrayLike) – source array containing values to add

• index (Tuple[slice]) – subpixel slice into dst that corresponds with src

• interp_axes (tuple) – specify which axes should be spatially interpolated

Todo

• [ ]: allow index to be a sequence indices

Example

>>> import kwimage
>>> dst = np.zeros(5) + .1
>>> src = np.ones(2)
>>> index = [slice(1.5, 3.5)]
>>> kwimage.util_warp.subpixel_set(dst, src, index)
>>> print(ub.repr2(dst, precision=2, with_dtype=0))
np.array([0.1, 0.5, 1. , 0.5, 0.1])

kwimage.subpixel_setvalue(img, pts, value, coord_axes=None, interp='bilinear', bordermode='edge')

Set values at subpixel locations

Parameters

• img (ArrayLike) – image to set values in

• pts (ArrayLike) – subpixel rc-coordinates to set

• value (ArrayLike) – value to place in the image

• coord_axes (Sequence, default=None) – axes to perform interpolation on, if not specified the first d axes are interpolated, where d=pts.shape[-1]. IE: this indicates which axes each coordinate dimension corresponds to.

• interp (str) – interpolation mode

• bordermode (str) – how locations outside the image are handled

Example

>>> from kwimage.util_warp import *  # NOQA
>>> img = np.arange(3 * 3).reshape(3, 3).astype(float)
>>> pts = np.array([[1, 1], [1.5, 1.5], [1.9, 1.1]])
>>> interp = 'bilinear'
>>> value = 0
>>> print('img = {!r}'.format(img))
>>> pts = np.array([[1.5, 1.5]])
>>> img2 = subpixel_setvalue(img.copy(), pts, value)
>>> print('img2 = {!r}'.format(img2))
>>> pts = np.array([[1.0, 1.0]])
>>> img2 = subpixel_setvalue(img.copy(), pts, value)
>>> print('img2 = {!r}'.format(img2))
>>> pts = np.array([[1.1, 1.9]])
>>> img2 = subpixel_setvalue(img.copy(), pts, value)
>>> print('img2 = {!r}'.format(img2))
>>> img2 = subpixel_setvalue(img.copy(), pts, value, coord_axes=[1, 0])
>>> print('img2 = {!r}'.format(img2))

kwimage.subpixel_slice(inputs, index)

Take a subpixel slice from a larger image. The returned output is left-aligned with the requested slice.

Parameters

• inputs (ArrayLike) – data

• index (Tuple[slice]) – a slice to subpixel accuracy

Example

>>> # xdoctest: +REQUIRES(module:torch)
>>> import kwimage
>>> import torch
>>> # say we have a (576, 576) input space
>>> # and a (9, 9) output space downsampled by 64x
>>> ospc_feats = np.tile(np.arange(9 * 9).reshape(1, 9, 9), (1024, 1, 1))
>>> inputs = torch.from_numpy(ospc_feats)
>>> # We detected a box in the input space
>>> ispc_bbox = kwimage.Boxes([[64,  65, 100, 120]], 'ltrb')
>>> # Get coordinates in the output space
>>> ospc_bbox = ispc_bbox.scale(1 / 64)
>>> tl_x, tl_y, br_x, br_y = ospc_bbox.data[0]
>>> # Convert the box to a slice
>>> index = [slice(None), slice(tl_y, br_y), slice(tl_x, br_x)]
>>> # Note: I'm not 100% sure this work right with non-intergral slices
>>> outputs = kwimage.subpixel_slice(inputs, index)

Example

>>> inputs = np.arange(5 * 5 * 3).reshape(5, 5, 3)
>>> index = [slice(0, 3), slice(0, 3)]
>>> outputs = subpixel_slice(inputs, index)
>>> index = [slice(0.5, 3.5), slice(-0.5, 2.5)]
>>> outputs = subpixel_slice(inputs, index)

>>> inputs = np.arange(5 * 5).reshape(1, 5, 5).astype(float)
>>> index = [slice(None), slice(3, 6), slice(3, 6)]
>>> outputs = subpixel_slice(inputs, index)
>>> print(outputs)
[[[18. 19.  0.]
  [23. 24.  0.]
  [ 0.  0.  0.]]]
>>> index = [slice(None), slice(3.5, 6.5), slice(2.5, 5.5)]
>>> outputs = subpixel_slice(inputs, index)
>>> print(outputs)
[[[20.   21.   10.75]
  [11.25 11.75  6.  ]
  [ 0.    0.    0.  ]]]

kwimage.subpixel_translate(inputs, shift, interp_axes=None, output_shape=None)

Translates an image by a subpixel shift value using bilinear interpolation

Parameters

• inputs (ArrayLike) – data to translate

• shift (Sequence) – amount to translate each dimension specified by interp_axes. Note: if inputs contains more than one “image” then all “images” are translated by the same amount. This function contains no mechanism for translating each image differently. Note that by default this is a y,x shift for 2 dimensions.

• interp_axes (Sequence, default=None) – axes to perform interpolation on, if not specified the final n axes are interpolated, where n=len(shift)

• output_shape (tuple, default=None) – if specified the output is returned with this shape, otherwise

Notes

This function powers most other functions in this file. Speedups here can go a long way.

Example

>>> inputs = np.arange(5) + 1
>>> print(inputs.tolist())
[1, 2, 3, 4, 5]
>>> outputs = subpixel_translate(inputs, 1.5)
>>> print(outputs.tolist())
[0.0, 0.5, 1.5, 2.5, 3.5]

Example

>>> # xdoctest: +REQUIRES(module:torch)
>>> inputs = torch.arange(9).view(1, 1, 3, 3).float()
>>> print(inputs.long())
tensor([[[[0, 1, 2],
          [3, 4, 5],
          [6, 7, 8]]]])
>>> outputs = subpixel_translate(inputs, (-.4, .5), output_shape=(1, 1, 2, 5))
>>> print(outputs)
tensor([[[[0.6000, 1.7000, 2.7000, 1.6000, 0.0000],
          [2.1000, 4.7000, 5.7000, 3.1000, 0.0000]]]])

Ignore:

>>> inputs = np.arange(5)
>>> shift = -.6
>>> interp_axes = None
>>> subpixel_translate(inputs, -.6)
>>> subpixel_translate(inputs[None, None, None, :], -.6)
>>> inputs = np.arange(25).reshape(5, 5)
>>> shift = (-1.6, 2.3)
>>> interp_axes = (0, 1)
>>> subpixel_translate(inputs, shift, interp_axes, output_shape=(9, 9))
>>> subpixel_translate(inputs, shift, interp_axes, output_shape=(3, 4))

kwimage.warp_image(inputs, mat, **kw)

kwimage.warp_points(matrix, pts, homog_mode='divide')

Warp ND points / coordinates using a transformation matrix.

Homogoenous coordinates are added on the fly if needed. Works with both numpy and torch.

Parameters

• matrix (ArrayLike) – [D1 x D2] transformation matrix. if using homogenous coordinates D2=D + 1, otherwise D2=D. if using homogenous coordinates and the matrix represents an Affine transformation, then either D1=D or D1=D2, i.e. the last row of zeros and a one is optional.

• pts (ArrayLike) – [N1 x … x D] points (usually x, y). If points are already in homogenous space, then the output will be returned in homogenous space. D is the dimensionality of the points. The leading axis may take any shape, but usually, shape will be [N x D] where N is the number of points.

• homog_mode (str, default=’divide’) – what to do for homogenous coordinates. Can either divide, keep, or drop.

Retrns:

new_pts (ArrayLike): the points after being transformed by the matrix

Example

>>> from kwimage.util_warp import *  # NOQA
>>> # --- with numpy
>>> rng = np.random.RandomState(0)
>>> pts = rng.rand(10, 2)
>>> matrix = rng.rand(2, 2)
>>> warp_points(matrix, pts)
>>> # --- with torch
>>> # xdoctest: +REQUIRES(module:torch)
>>> pts = torch.Tensor(pts)
>>> matrix = torch.Tensor(matrix)
>>> warp_points(matrix, pts)

Example

>>> from kwimage.util_warp import *  # NOQA
>>> # --- with numpy
>>> pts = np.ones((10, 2))
>>> matrix = np.diag([2, 3, 1])
>>> ra = warp_points(matrix, pts)
>>> # xdoctest: +REQUIRES(module:torch)
>>> rb = warp_points(torch.Tensor(matrix), torch.Tensor(pts))
>>> assert np.allclose(ra, rb.numpy())

Example

>>> from kwimage.util_warp import *  # NOQA
>>> # test different cases
>>> rng = np.random.RandomState(0)
>>> # Test 3x3 style projective matrices
>>> pts = rng.rand(1000, 2)
>>> matrix = rng.rand(3, 3)
>>> ra33 = warp_points(matrix, pts)
>>> # xdoctest: +REQUIRES(module:torch)
>>> rb33 = warp_points(torch.Tensor(matrix), torch.Tensor(pts))
>>> assert np.allclose(ra33, rb33.numpy())
>>> # Test opencv style affine matrices
>>> pts = rng.rand(10, 2)
>>> matrix = rng.rand(2, 3)
>>> ra23 = warp_points(matrix, pts)
>>> rb23 = warp_points(torch.Tensor(matrix), torch.Tensor(pts))
>>> assert np.allclose(ra33, rb33.numpy())

kwimage.warp_tensor(inputs, mat, output_dims, mode='bilinear', padding_mode='zeros', isinv=False, ishomog=None, align_corners=False, new_mode=False)

A pytorch implementation of warp affine that works similarly to cv2.warpAffine / cv2.warpPerspective.

It is possible to use 3x3 transforms to warp 2D image data. It is also possible to use 4x4 transforms to warp 3D volumetric data.

Parameters

• inputs (Tensor[…, *DIMS]) – tensor to warp. Up to 3 (determined by output_dims) of the trailing space-time dimensions are warped. Best practice is to use inputs with the shape in [B, C, *DIMS].

• mat (Tensor) – either a 3x3 / 4x4 single transformation matrix to apply to all inputs or Bx3x3 or Bx4x4 tensor that specifies a transformation matrix for each batch item.

• output_dims (Tuple[int]*) –

  The output space-time dimensions. This can either be in the form

  (W,), (H, W), or (D, H, W).

• mode (str) – Can be bilinear or nearest. See torch.nn.functional.grid_sample

• padding_mode (str) – Can be zeros, border, or reflection. See torch.nn.functional.grid_sample.

• isinv (bool, default=False) – Set to true if mat is the inverse transform

• ishomog (bool, default=None) – Set to True if the matrix is non-affine

• align_corners (bool, default=False) – Note the default of False does not work correctly with grid_sample in torch <= 1.2, but using align_corners=True isnt typically what you want either. We will be stuck with buggy functionality until torch 1.3 is released.

  However, using align_corners=0 does seem to reasonably correspond with opencv behavior.

Notes

Also, it may be possible to speed up the code with F.affine_grid

KNOWN ISSUE: There appears to some difference with cv2.warpAffine when

rotation or shear are non-zero. I’m not sure what the cause is. It may just be floating point issues, but Im’ not sure.

Todo

• [ ] FIXME: see example in Mask.scale where this algo breaks when

the matrix is 2x3 - [ ] Make this algo work when matrix ix 2x2

References

https://discuss.pytorch.org/t/affine-transformation-matrix-paramters-conversion/19522 https://github.com/pytorch/pytorch/issues/15386

Example

>>> # Create a relatively simple affine matrix
>>> # xdoctest: +REQUIRES(module:torch)
>>> import skimage
>>> mat = torch.FloatTensor(skimage.transform.AffineTransform(
>>>     translation=[1, -1], scale=[.532, 2],
>>>     rotation=0, shear=0,
>>> ).params)
>>> # Create inputs and an output dimension
>>> input_shape = [1, 1, 4, 5]
>>> inputs = torch.arange(int(np.prod(input_shape))).reshape(*input_shape).float()
>>> output_dims = (11, 7)
>>> # Warp with our code
>>> result1 = warp_tensor(inputs, mat, output_dims=output_dims, align_corners=0)
>>> print('result1 =\n{}'.format(ub.repr2(result1.cpu().numpy()[0, 0], precision=2)))
>>> # Warp with opencv
>>> import cv2
>>> cv2_M = mat.cpu().numpy()[0:2]
>>> src = inputs[0, 0].cpu().numpy()
>>> dsize = tuple(output_dims[::-1])
>>> result2 = cv2.warpAffine(src, cv2_M, dsize=dsize, flags=cv2.INTER_LINEAR)
>>> print('result2 =\n{}'.format(ub.repr2(result2, precision=2)))
>>> # Ensure the results are the same (up to floating point errors)
>>> assert np.all(np.isclose(result1[0, 0].cpu().numpy(), result2, atol=1e-2, rtol=1e-2))

Example

>>> # Create a relatively simple affine matrix
>>> # xdoctest: +REQUIRES(module:torch)
>>> import skimage
>>> mat = torch.FloatTensor(skimage.transform.AffineTransform(
>>>     rotation=0.01, shear=0.1).params)
>>> # Create inputs and an output dimension
>>> input_shape = [1, 1, 4, 5]
>>> inputs = torch.arange(int(np.prod(input_shape))).reshape(*input_shape).float()
>>> output_dims = (11, 7)
>>> # Warp with our code
>>> result1 = warp_tensor(inputs, mat, output_dims=output_dims)
>>> print('result1 =\n{}'.format(ub.repr2(result1.cpu().numpy()[0, 0], precision=2, supress_small=True)))
>>> print('result1.shape = {}'.format(result1.shape))
>>> # Warp with opencv
>>> import cv2
>>> cv2_M = mat.cpu().numpy()[0:2]
>>> src = inputs[0, 0].cpu().numpy()
>>> dsize = tuple(output_dims[::-1])
>>> result2 = cv2.warpAffine(src, cv2_M, dsize=dsize, flags=cv2.INTER_LINEAR)
>>> print('result2 =\n{}'.format(ub.repr2(result2, precision=2)))
>>> print('result2.shape = {}'.format(result2.shape))
>>> # Ensure the results are the same (up to floating point errors)
>>> # NOTE: The floating point errors seem to be significant for rotation / shear
>>> assert np.all(np.isclose(result1[0, 0].cpu().numpy(), result2, atol=1, rtol=1e-2))

Example

>>> # Create a random affine matrix
>>> # xdoctest: +REQUIRES(module:torch)
>>> import skimage
>>> rng = np.random.RandomState(0)
>>> mat = torch.FloatTensor(skimage.transform.AffineTransform(
>>>     translation=rng.randn(2), scale=1 + rng.randn(2),
>>>     rotation=rng.randn() / 10., shear=rng.randn() / 10.,
>>> ).params)
>>> # Create inputs and an output dimension
>>> input_shape = [1, 1, 5, 7]
>>> inputs = torch.arange(int(np.prod(input_shape))).reshape(*input_shape).float()
>>> output_dims = (3, 11)
>>> # Warp with our code
>>> result1 = warp_tensor(inputs, mat, output_dims=output_dims, align_corners=0)
>>> print('result1 =\n{}'.format(ub.repr2(result1.cpu().numpy()[0, 0], precision=2)))
>>> # Warp with opencv
>>> import cv2
>>> cv2_M = mat.cpu().numpy()[0:2]
>>> src = inputs[0, 0].cpu().numpy()
>>> dsize = tuple(output_dims[::-1])
>>> result2 = cv2.warpAffine(src, cv2_M, dsize=dsize, flags=cv2.INTER_LINEAR)
>>> print('result2 =\n{}'.format(ub.repr2(result2, precision=2)))
>>> # Ensure the results are the same (up to floating point errors)
>>> # NOTE: The errors seem to be significant for rotation / shear
>>> assert np.all(np.isclose(result1[0, 0].cpu().numpy(), result2, atol=1, rtol=1e-2))

Example

>>> # Test 3D warping with identity
>>> # xdoctest: +REQUIRES(module:torch)
>>> mat = torch.eye(4)
>>> input_dims = [2, 3, 3]
>>> output_dims = (2, 3, 3)
>>> input_shape = [1, 1] + input_dims
>>> inputs = torch.arange(int(np.prod(input_shape))).reshape(*input_shape).float()
>>> result = warp_tensor(inputs, mat, output_dims=output_dims)
>>> print('result =\n{}'.format(ub.repr2(result.cpu().numpy()[0, 0], precision=2)))
>>> assert torch.all(inputs == result)

Example

>>> # Test 3D warping with scaling
>>> # xdoctest: +REQUIRES(module:torch)
>>> mat = torch.FloatTensor([
>>>     [0.8,   0,   0, 0],
>>>     [  0, 1.0,   0, 0],
>>>     [  0,   0, 1.2, 0],
>>>     [  0,   0,   0, 1],
>>> ])
>>> input_dims = [2, 3, 3]
>>> output_dims = (2, 3, 3)
>>> input_shape = [1, 1] + input_dims
>>> inputs = torch.arange(int(np.prod(input_shape))).reshape(*input_shape).float()
>>> result = warp_tensor(inputs, mat, output_dims=output_dims, align_corners=0)
>>> print('result =\n{}'.format(ub.repr2(result.cpu().numpy()[0, 0], precision=2)))
result =
np.array([[[ 0.  ,  1.25,  1.  ],
           [ 3.  ,  4.25,  2.5 ],
           [ 6.  ,  7.25,  4.  ]],
          ...
          [[ 7.5 ,  8.75,  4.75],
           [10.5 , 11.75,  6.25],
           [13.5 , 14.75,  7.75]]], dtype=np.float32)

Example

>>> # xdoctest: +REQUIRES(module:torch)
>>> mat = torch.eye(3)
>>> input_dims = [5, 7]
>>> output_dims = (11, 7)
>>> for n_prefix_dims in [0, 1, 2, 3, 4, 5]:
>>>      input_shape = [2] * n_prefix_dims + input_dims
>>>      inputs = torch.arange(int(np.prod(input_shape))).reshape(*input_shape).float()
>>>      result = warp_tensor(inputs, mat, output_dims=output_dims)
>>>      #print('result =\n{}'.format(ub.repr2(result.cpu().numpy(), precision=2)))
>>>      print(result.shape)

Example

>>> # xdoctest: +REQUIRES(module:torch)
>>> mat = torch.eye(4)
>>> input_dims = [5, 5, 5]
>>> output_dims = (6, 6, 6)
>>> for n_prefix_dims in [0, 1, 2, 3, 4, 5]:
>>>      input_shape = [2] * n_prefix_dims + input_dims
>>>      inputs = torch.arange(int(np.prod(input_shape))).reshape(*input_shape).float()
>>>      result = warp_tensor(inputs, mat, output_dims=output_dims)
>>>      #print('result =\n{}'.format(ub.repr2(result.cpu().numpy(), precision=2)))
>>>      print(result.shape)

Ignore:

import xdev globals().update(xdev.get_func_kwargs(warp_tensor)) >>> # xdoctest: +REQUIRES(module:torch) >>> import cv2 >>> inputs = torch.arange(9).view(1, 1, 3, 3).float() + 2 >>> input_dims = inputs.shape[2:] >>> #output_dims = (6, 6) >>> def fmt(a): >>> return ub.repr2(a.numpy(), precision=2) >>> s = 2.5 >>> output_dims = tuple(np.round((np.array(input_dims) * s)).astype(int).tolist()) >>> mat = torch.FloatTensor([[s, 0, 0], [0, s, 0], [0, 0, 1]]) >>> inv = mat.inverse() >>> warp_tensor(inputs, mat, output_dims) >>> print(‘## INPUTS’) >>> print(fmt(inputs)) >>> print(’nalign_corners=True’) >>> print(’—-‘) >>> print(‘## warp_tensor, align_corners=True’) >>> print(fmt(warp_tensor(inputs, inv, output_dims, isinv=True, align_corners=True))) >>> print(‘## interpolate, align_corners=True’) >>> print(fmt(F.interpolate(inputs, output_dims, mode=’bilinear’, align_corners=True))) >>> print(’nalign_corners=False’) >>> print(’—-‘) >>> print(‘## warp_tensor, align_corners=False, new_mode=False’) >>> print(fmt(warp_tensor(inputs, inv, output_dims, isinv=True, align_corners=False))) >>> print(‘## warp_tensor, align_corners=False, new_mode=True’) >>> print(fmt(warp_tensor(inputs, inv, output_dims, isinv=True, align_corners=False, new_mode=True))) >>> print(‘## interpolate, align_corners=False’) >>> print(fmt(F.interpolate(inputs, output_dims, mode=’bilinear’, align_corners=False))) >>> print(‘## interpolate (scale), align_corners=False’) >>> print(ub.repr2(F.interpolate(inputs, scale_factor=s, mode=’bilinear’, align_corners=False).numpy(), precision=2)) >>> cv2_M = mat.cpu().numpy()[0:2] >>> src = inputs[0, 0].cpu().numpy() >>> dsize = tuple(output_dims[::-1]) >>> print(’nOpen CV warp Result’) >>> result2 = (cv2.warpAffine(src, cv2_M, dsize=dsize, flags=cv2.INTER_LINEAR)) >>> print(‘result2 =n{}’.format(ub.repr2(result2, precision=2)))