0.30.8 0.30.9
roboflow/supervisionOct 6, 2026 → Oct 8, 20262 days between releases
What's New
NumPy fallback matches OpenCV behavior and validation is stricter for box and classification inputs.
Breaking Changes
- COCO, Pascal VOC, and CreateML loaders now raise errors on reversed boxes.
- `get_top_k` now raises `ValueError` for negative `k`.
New Features
- Browser-recorded WebM and variable frame rate MKV support to end of stream.
- Return empty `KeyPoints` instead of raising `IndexError` on empty pose results.
Improvements
- NumPy fallback now draws thick borders and resizes images identically to OpenCV.
- Stricter validation for `epsilon` in `approxPolyDP` (rejects NaN, infinite, or 1e30+).
- `DetectionsSmoother` ages history on frames without a tracker_id.
Bug Fixes
- Video frame generators now read to the end of the stream, fixing browser-recorded WebM support.
- `DetectionsSmoother` forgets expired tracks on frames without `tracker_id`.
Upgrade Notes
Data with negative box dimensions or invalid `k` values will now raise errors. If you rely on the NumPy fallback (no OpenCV), visual output for borders and resizing may shift slightly to match OpenCV behavior.
The Two Releases
supervision-0.30.8
Oct 6, 2026· 15 days ago
A patch release improving video quality without OpenCV, fixing YOLO pose label parsing, and correcting mAP calculations for class-agnostic scenarios.
Full release notessupervision-0.30.9
Oct 8, 2026· 13 days ago
Enhances the NumPy fallback to match OpenCV behavior, fixes video reading to the end, and enforces stricter validation on box and classification inputs.
Full release notes