New Release: Ultralytics v8.4.187

Discussion
, ,
1

:rocket: Ultralytics v8.4.187: Smarter YOLO26-Pose Training

Ultralytics YOLO v8.4.187 improves YOLO26-pose training with keypoint-focused matching and scale-aware anchor selection. The update applies automatically—no new settings, dependencies, or migration steps required.

:bullseye: Key Improvements

  • Keypoint-first matching: For objects with visible keypoints, training assignment now uses only Object Keypoint Similarity (OKS), rather than blending it with bounding-box overlap. This favors anchors with more accurate keypoints.
  • Scale-aware anchor selection: An FCOS-style gate limits candidate anchors to feature-pyramid levels suited to each object’s size. The finest and coarsest levels remain available for objects outside the usual size range.

:shield: Fallbacks Preserved

  • If scale filtering would leave an object without candidates, the original, ungated candidates are retained.
  • Objects without visible keypoints continue to use bounding-box overlap for matching.

:gear: What This Means for You

The updated behavior is enabled by default for YOLO26-pose training. Inference, export, and the model architecture remain unchanged.

Thanks to @Bovey0809 for implementing OKS-only assignment and the pyramid-level gate in PR #26843, making this training update possible.

:package: Try It Out

Upgrade your installation:

pip install -U ultralytics

Explore the v8.4.187 release notes for the announcement, or review the full changelog to inspect the changes.

:speech_balloon: Training a YOLO26-pose model? Try the update and share your results, observations, or questions below—we’d love to hear your feedback!