v0.6.0

This minor release notably adds new high-level interfaces for black-box object detector and image similarity saliency generation. We provide some reference implementations for these interfaces, notably the D-RISE and SBSM algorithms, respectively.

Other improvements include the addition of more examples notebooks, improvements/fixes to existing implementations and a revision to the CLI object detection saliency generation tool.

See below for more details.

Updates / New Features

CI

  • Added the ATARI example notebook to the list of notebooks to run during CI.

Documentation

  • Update saliency algorithms table with perturbation-based saliency for reinforcement learning and add corresponding section to README.

  • Added a lighter color version of the logo that will appear better is both light- and dark-theme contexts. The main README file has been updated to refer to this image.

  • Added introductory sentence to the style sheet document.

  • Updated the release process to be incrementally more comprehensive and now includes the specification of a release branch with better patch release instructions. This also now includes a step to update the version referenced in the xaitk.org source.

Examples

  • Updated demo resource download links from Google Drive to data.kitware.com

  • Added example using saliency to qualitatively compare two object detection models.

  • Updated SimilarityScoring example to use new high-level image similarity saliency interface and follow new similarity interface inputs.

Interfaces

  • Added new high-level interface for black-box object detector saliency, GenerateObjectDetectorBlackboxSaliency.

  • Updated image similarity interface GenerateDescriptorSimilaritySaliency to accept multiple query images and compute a saliency map for each one, relative to the reference image.

  • Added new high-level interface for image similarity saliency, GenerateImageSimilarityBlackboxSaliency.

Implementations

  • Added three GenerateObjectDetectorBlackboxSaliency implementations: the generic PerturbationOcclusion, and two usable classes DRISEStack and RandomGridStack.

  • Updated behavior of the SlidingWindow PerturbImage implementation. For a given stride, the number of masks generated is now agnostic to the window size.

  • Updated SimilarityScoring to return [N x H x W] instead of [1 x H x W] saliency heatmaps matrix. This is inline with the similarity interface update.

  • Added two implementations of GenerateImageSimilarityBlackboxSaliency: PerturbationOcclusion and SBSMStack.

Misc.

  • Updated poetry-core build backend to version 1.0.8, which now supports pip editable installs (pip install -e .).

Utils

  • Updated COCO utility functions to use new high-level detector interface. gen_coco_sal() is now deprecated in exchange for parse_coco_dset() which parses a kwcoco.CocoDataset object into the inputs used with an implementation of GenerateObjectDetectorBlackboxSaliency.

Fixes

Dependency Versions

  • Update pinned jupyter notebook transitive dependency version due to vulnerability warning.

Examples

  • Fixed inconsistency of dependency package installs at the head of the examples/SerializedDetectionSaliency.ipynb notebook.