FOCUS Foot Refinement ONNX
Export of Logimade/FOCUS's existing YOLO26n-pose checkpoint. No new training is claimed. The upstream checkpoint is pinned and its Git object is verified before export. The model weights retain Ultralytics AGPL-3.0 terms; the FOCUS README's code license does not replace the pretrained model license. This is separate from the Apache-2.0 Wholebody-Refinement-onnx repository.
Eight points: left big toe, small toe, heel, ankle; right big toe, small toe, heel, ankle. COCO-WholeBody indices: 17,18,19,15,20,21,22,16. These are 2D estimated landmarks. Five individual toes, toe articulation, depth and true visibility are not measured.
Input: RGB float32 NCHW 1x3x640x640, letterbox padding 114, divide by 255. Output coordinates are already decoded in input-image pixels. Undo letterbox and add any crop origin; do not apply a second YOLO head decoder. See foot-models.json for actual measured output shapes and export settings. The one-to-many head needs external NMS when retaining multiple detections. Argmax selects one foot pair.
Source and reproducible Actions evaluation:
- https://github.com/Logimade/FOCUS
- https://github.com/DaisukeDaisuke/Wholebody-test
- https://www.ultralytics.com/license
The small COCO-WholeBody diagnostic evaluation is not an independent generalization benchmark. Images and annotations are not distributed in this model repository.