Can frozen pretrained human mesh recovery produce 3D body pose features that support Yoga-82 classification and an interactive single-person demo?
- RGB classifier.
- 2D-keypoint classifier.
- SMPL rotation and normalized 3D-joint MLP.
- 3D-joint graph classifier.
Inspect 100--300 stratified Yoga-82 images before classifier training. Record:
- person-detection failure;
- left/right limb swap;
- inverted-body failure;
- self-contact topology error;
- severe mesh/image misalignment;
- low-confidence or ambiguous reconstruction.
Do not treat downstream classification failures as evidence about 3D pose if the upstream reconstruction has not passed this audit.
- Remove global translation and normalize body scale.
- Preserve orientation relative to gravity; do not rotate inverted poses to an upright canonical pose.
- Exclude body shape from the primary classifier to reduce identity leakage.
- Ignore face, jaw, expression, and detailed fingers in the first baseline.
- Report Top-1, Top-5, macro-F1, per-class recall, and hierarchical accuracy.
- Compare 3D features against both RGB and 2D-keypoint baselines.
- Report end-to-end P50/P95 latency and FPS on the target GPU.
- Report failure coverage and abstention, not accuracy only on successful cases.