Completed: 2026-08-17
- Frozen dataset:
yoga82_reachable_clean_v1.csv - Rows: 11,760
- Detector: Faster R-CNN ResNet-50 FPN v2, score threshold 0.55
- HMR: official frozen HMR2 checkpoint at epoch 35 / step 1,000,000
- Body model: licensed SMPL Neutral
| Status | Count |
|---|---|
| Detected | 10,623 |
| No person | 1,137 |
| Decode error | 0 |
OpenCV initially rejected 47 files whose names ended in .jpg but whose
content was GIF. The production loader now falls back to the first Pillow
frame; all 47 were retried, producing 27 detections and 20 ordinary no-person
rejections.
| Measure | Result |
|---|---|
| Detected rows submitted to HMR | 10,623 |
| Successful HMR caches | 10,623 |
| HMR failures | 0 |
| Successful train | 7,622 |
| Successful test | 3,001 |
| Classifier dimensions | 216 |
| P50 HMR forward | 0.0230 s |
| P95 HMR forward | 0.0237 s |
| Feature NPZ files | 10,623 |
| Metadata JSON files | 10,623 |
| Vertices saved | 0 |
| Total HMR output size | 96 MB |
Each NPZ contains finite SMPL global orientation and body pose, predicted 3D joints, camera parameters, detector bbox, and the gravity-preserving 216-dimensional classifier feature vector. Full vertices and renders were omitted because they are not classification inputs.
Both successful train and test subsets retain complete 6 / 20 / 82 hierarchy coverage. Independent validation opened and checked every NPZ rather than trusting only the run summary.
- Detection manifest and summary:
/root/autodl-tmp/3Dyujia/outputs/detection/yoga82_clean_v1/ - HMR feature manifest, NPZ, metadata, run summary and validation:
/root/autodl-tmp/3Dyujia/outputs/hmr2/yoga82_clean_v1/
The authoritative classifier input is:
/root/autodl-tmp/3Dyujia/outputs/hmr2/yoga82_clean_v1/feature_manifest.csv
Numerical cache validity does not establish correct 3D reconstruction. The existing 164-image visual audit remains the evidence for gross mesh quality, and the accepted/rejected populations must be reported separately in later classification experiments.