Bare-hand detector (YOLOX-M, 960)
This YOLOX-M detector finds bare hands in egocentric video and labels each one as the anatomical left or right hand.
OpenGraph Labs uses it as the 2D hand front end in its NVIDIA V2D Challenge Track 3 pipeline (og_v2d.hands.front_end).
| Architecture | YOLOX-M: depth 0.67, width 0.75 (Megvii YOLOX at 419778480ab6) |
| Classes | 0 left hand, 1 right hand (anatomical) |
| Input | Aspect-preserving resize to 960 x 960 with pad value 114. BGR as decoded by OpenCV, no mean/std normalisation (YOLOX ValTransform(legacy=False)) |
| Post-processing | score = objectness x class confidence. Keep scores ≥ 0.75, then per-class NMS at IoU 0.65 and at most 2 boxes per class |
| File | yolox_m_barehand_960.pth: 202,748,467 bytes, sha256 d73d06cdaec1fba24292619e1dbc3efca980f1eff676acd1c9d7998708166705 |
Training data
We trained the detector on OpenGraph Labs' internal egocentric bare-hand images. That dataset is not released.
Evaluation
These scores come from our internal held-out sets at score threshold 0.75:
| Split | Images | AP50 | Precision | Recall | Left/right swaps |
|---|---|---|---|---|---|
| Validation | 790 | 0.947 | 0.976 | 0.923 | 2 |
| Locked test | 240 | 0.777 | 0.870 | 0.800 | 3 |
Usage
For inference, use og_v2d.hands.front_end in the og-v2d-track3 repository, or load the checkpoint with a YOLOX-M exp
built with num_classes=2 and test_size=(960, 960).
License
Apache-2.0.