Files
AV-Live/data_only_viz/training
L'électron rare beb94d2a4c feat(data-only-viz): action-head v3 hands+lips
Extends the action-head feature pipeline from v2 (302-D) to v3 (428-D).

- Replace placeholder SMPLX_FINGERTIP_VERTS with canonical vertex IDs
  from smplx.vertex_ids (lthumb/lindex/lmiddle/lring/lpinky, mirrored R)
- Add HANDS_KP_* constants (21 kp/hand, 42 total, 126-D flat block)
- FEATURE_DIM: 302 -> 428; hands_kp block inserted at [288:414]
- FeatureExtractor.from_buffer gains hands_kp param (42, 3),
  zero-padded when absent
- ActionHead.step gains hands_kp param, threads to from_buffer
- _read_sources returns 5-tuples with hands_kp42x3 per person
- MediaPipe FaceMesh inner-lip (idx 13/14) used for mouth_open;
  fallback to SMPL-X v3d lip vertices when face not available
- _build_hands_map and _build_face_mouth_map helpers added
- dataset.py: RawFrame/WindowRow/DatasetRow gain hands_kp fields
- train_action_head.py: reads hands_kp_stack per step, zeros if absent
- extract_j3d_offline.py: writes zero-filled hands_kp in jsonl output
- Tests: FEATURE_DIM 302->428, param bound 80k->100k, +4 new tests
2026-05-13 23:26:14 +02:00
..