Tags: Kitware/SMQTK-Detection
Tags
v0.22.0 ======= Added new configuration option for handling malformed bboxes within ``CenterNetVisdrone``. Updates / New Features ---------------------- - Previously, malformed bboxes (``max_vertex < min_vertex``) returned by the internal model architecture of ``CenterNetVisdrone`` caused an exception to be raised by ``AxisAlignedBoundingBox`` during bbox construction. Now, ``reorder_malformed_bboxes`` is exposed as a configuration parameter to more appropriately handle behavior when malformed bboxes are encountered. If ``True``, bbox vertices will be sorted such that (``max_vertex >= min_vertex``), otherwise, (new default) the offending detection will be dropped from the list of detections completely. This new default is similar to previous behavior, but avoids hitting the exception such that the remaining well-formed detections can be returned. Debug logging may be enabled to have more visibility into when malformed bboxes are encountered. Fixes -----
v0.21.0 ======= ``CenterNetVisdrone`` torch fix for MPS devices, GitHub CI updates. Updates / New Features ---------------------- CI/CD * Updated Github Workflows from ``cache@v2`` to ``cache@v4``. * To resolve issues with public forks lacking access to the Codecov token, we've included it directly in ``codecov.yml``. The file also documents the rationale for this security exception. Fixes ----- * Fixed a bug in ``CenterNetVisdrone`` that prevented running on MPS device.
v0.20.0 ======= This minor release updates the mimumum supported python to `python = "^3.8"`, addresses dependency vulnerabilities, and updates typing to conform with current mypy and pytest standards. Updates / New Features ---------------------- Python * New minimum supported python changed to `python = "^3.8"`. Dependencies * Updated python minimum requirement to 3.8 (up from 3.6). This involved a number of updates and bifurcations of abstract requirements, an update to pinned versions for development/CI, and expansion of CI to cover python versions 3.10 and 3.11 (latest current release). Fixes ----- Docs * Fix erroneous references to previous monorepo. * Fixed ``sphinx_server.py`` to reference correct directories. Dependency versions * Updated the locked versions of dependencies to reflect new minimum support `python = "^3.8"`.
v0.19.0 ======= This minor release includes refinements to the `DetectImageObjects` plugins, including a change to return semantic class labels instead of just integers, reducing the need for the caller to know intrinsics about the plugin/model in order to use it. See below for additional updates and fixes. Updates / New Features ---------------------- CI * Updated CI unittests workflow to include codecov reporting. Reduced CodeCov report submission by skipping this step on scheduled runs. Documentation * Updated CONTRIBUTING.md to reference smqtk-core's CONTRIBUTING.md file. Detect Image Objects * Updated the `ResNetFRCNN` to return as its class labels the label strings instead of integers, reducing the burden of users from having to repeatedly find and allocate the appropriate int-to-label map. * Update `CenterNetVisdrone` to lazy load its model on first inference. * Update `CenterNetVisdrone` to return VisDrone class labels instead of integers. Unit Tests * Added unit tests for the `CenterNetVisdrone` plugin. Utilities * Remove duplicate implementation of `AxisAlignedBoundingBox`, using the tested version that comes from the ``smqtk_image_io`` package. Fixes ----- Detect Image Object * Fixed batched operation memory usage in `ResNetFRCNN` by loading only current batch into computation device memory. Previously all images were loaded at once. * Fixed device mapping when loading certain background architectures for `CenterNetVisdrone`. Dependency Versions * Updated the developer dependency and locked version of ipython to address a security vulnerability. * Removed `jedi = "^0.17.2"` requirement since recent `ipython = "^7.17.3"` update appropriately addresses the dependency.
v0.18.1 ======= This patch release addresses an issue with the centernet plugin where image preprocessing was not happening on an appropriate per-batch cadence which led to more RAM being consumed than intended for large input sequences. Fixes ----- CI * Also run CI unittests for PRs targeting branches that match the `release*` glob. Implementations * Fix ``CenterNetVisdrone`` batched operation and class labels