The kwimage module
Read the docs
The main webpage for this project is: https://gitlab.kitware.com/computer-vision/kwimage
The kwimage module handles low-level image operations at a high level.
The core kwimage is a functional library with image-related helper functions that are either unimplemented in or more have a more general interface then their opencv counterparts.
The kwimage module builds on kwarray and provides tools commonly needed when addressing computer vision problems. This includes functions for reading images, resizing, image warp transformations, run-length-encoding, and non-maximum-suppression.
The kwimage module is also the current home of my annotation data structures, which provide efficient ways to interoperate between different common annotation formats (e.g. different bounding box / polygon / point formats). These data structures have both a .draw and .draw_on method for overlaying visualizations on matplotlib axes or numpy image matrices respectively.
Read the docs at: http://kwimage.readthedocs.io/en/main/
The top-level API is:
from .algo import (available_nms_impls, daq_spatial_nms, non_max_supression,) from .im_alphablend import (ensure_alpha_channel, overlay_alpha_images, overlay_alpha_layers,) from .im_color import (Color,) from .im_core import (atleast_3channels, ensure_float01, ensure_uint255, exactly_1channel, find_robust_normalizers, make_channels_comparable, normalize, normalize_intensity, num_channels, padded_slice,) from .im_cv2 import (connected_components, convert_colorspace, gaussian_blur, gaussian_patch, imcrop, imresize, imscale, morphology, warp_affine, warp_image, warp_projective,) from .im_demodata import (checkerboard, grab_test_image, grab_test_image_fpath,) from .im_draw import (draw_boxes_on_image, draw_clf_on_image, draw_header_text, draw_line_segments_on_image, draw_text_on_image, draw_vector_field, fill_nans_with_checkers, make_heatmask, make_orimask, make_vector_field, nodata_checkerboard,) from .im_filter import (fourier_mask, radial_fourier_mask,) from .im_io import (imread, imwrite, load_image_shape,) from .im_runlen import (decode_run_length, encode_run_length, rle_translate,) from .im_stack import (stack_images, stack_images_grid,) from .structs import (Box, Boxes, Coords, Detections, Heatmap, Mask, MaskList, MultiPolygon, Points, PointsList, Polygon, PolygonList, Segmentation, SegmentationList, smooth_prob,) from .transform import (Affine, Linear, Matrix, Projective, Transform,) from .util_warp import (add_homog, remove_homog, subpixel_accum, subpixel_align, subpixel_getvalue, subpixel_maximum, subpixel_minimum, subpixel_set, subpixel_setvalue, subpixel_slice, subpixel_translate, warp_points, warp_tensor,)
NOTE: THE KWIMAGE STRUCTS MAY? EVENTUALLY MOVE TO THE KWANNOT REPO (But this transition might take awhile)
The most notable feature of the kwimage module are the kwimage.structs objects. This includes the primitive Boxes, Mask, and Coords objects, The semi-primitive Points, Polygon structures, and the composite Heatmap and Detections structures (note: Heatmap is just a composite of array-like structures).
The primitive and semi-primitive objects store and manipulate annotation geometry, and the composite structures combine primitives into a single object that jointly manipulates the primitives using warp operations.
The Detections structure is a meta-structure that associates the other more primitive components, and allows a developer to compose them into something that represents objects of interest. The details of this composition are left up to the end-application.
The Detections object can also be “rasterized” and converted into a Heatmap object, which represents the same information, but is in a form that is more suitable for use when training convolutional neural networks. Likewise, the output of neural networks can be directly encoded in a kwimage.Heatmap object. The Heatmap.detect method can then be used to convert the dense heatmap representation into a spare Detections representation that is more suitable for use in an object-detection system. We note that the detect function is not a special detection algorithm. The detection algorithm (which is outside the scope of kwimage) produces the heatmap, and the detect method effectively “inverts” the rasterize procedure of Detections by finding peaks in the heatmap, and running non-maximum suppression.
This module contains data structures for three image annotation primitives:
Boxes # technically this could be made out of Coords, probably not for efficiency and decoupling
Mask # likewise this could be renamed to Raster
These primative structures are used to define these metadata-containing composites:
All of these structures have a self.data attribute that holds a pointer to the underlying data representation.
Some of these structures have a self.format attribute describing the underlying data representation.
Most of the compositie strucutres also have a self.meta attribute, which holds user-level metadata (e.g. info about the classes).
There are a few small quirks with installing kwimage. There is an issue with the opencv python bindings such that we could rely on either the opencv-python or opencv-python-headless package. If you have either of these module already installed you can simply pip install kwimage without encountering any issues related to this. But if you do not already have a module that provides import cv2 installed, then you should install kwimage with one of the following “extra install” tags:
# We recommend using the headless version pip install kwimage[headless] # OR # If other parts of your system depend on the opencv qt libs # (NOT RECOMMENDED: this can conflict with pyqt5) pip install kwimage[graphics]
Some features also require the kwimage_ext package to be installed, which contains binary extensions that used to be distributed with this package in older versions. These extension can be obtained by explicitly pip install kwimage_ext or via pip install kwimage[optional] (which also brings in other optional libraries). You can disable loading of c-extensions at runtime by setting the environment variable: KWIMAGE_DISABLE_C_EXTENSIONS=1.
A Note on GDAL
The kwimage library can use GDAL library for certain tasks (e.g. IO of geotiffs). GDAL can be a pain to install without relying on conda. Kitware also has a pypi index that hosts GDAL wheels for linux systems:
pip install --find-links https://girder.github.io/large_image_wheels GDAL
Release history Release notifications | RSS feed
Download the file for your platform. If you're not sure which to choose, learn more about installing packages.