🌟 Summary
YOLO26-pose training now assigns keypoints using pose accuracy, with a feature-pyramid scale gate to help match objects to suitable training anchors.
📊 Key Changes
- OKS-only matching: For ground truths with visible keypoints, assignment now uses keypoint similarity (OKS) instead of blending it with box overlap.
- Scale-aware candidate selection: An FCOS-style gate limits candidate anchors to pyramid levels suited to each object’s size. The finest and coarsest levels remain available for objects outside the usual size range.
- Fallbacks preserved: If the gate would leave a ground truth with no candidates, the ungated candidates are retained. Ground truths without visible keypoints continue to use box overlap.
- Training-only scope: The change applies to YOLO26-pose training by default. It adds no configuration options or dependencies and does not change inference, export, or the model architecture.
🎯 Purpose & Impact
- 🎯 Encourages training assignments to favor anchors with more accurate keypoints.
- 📏 Helps pure keypoint-based matching account for object size across feature-pyramid levels.
- 🛠️ Users get the updated behavior automatically when training YOLO26-pose; no migration steps are needed.
What's Changed
- ultralytics 8.4.187 OKS-only label assignment with a pyramid-level gate for YOLO26-pose by @Bovey0809 in #26843
Full Changelog: v8.4.186...v8.4.187