T3A — Pose / keypoints — camera & studio shift¶
Paper evidence: main.pdf · Block findings
Lemma: D3 · Stack: pytorch
Nuisance key: augmentation
Production change: Camera, lighting, viewpoint change; same keypoint indices.
Notebook (Run All, built-in demo): t03a-pose-gradient.ipynb
pip install matching-pmh torch
# Open the notebook and Run All
What this task achieved (headline)¶
E1_aniso subspace PMH on COCO pose: 54.49% clean PCK@0.05 (+22.4 pp vs baseline 32.07%).
| baseline | E1_aniso |
|---|---|
| 32.07% PCK | 54.49% PCK |
| — | 35.21% @ occ 0.40 |
Subtasks (paper)¶
Calibrate occlusion subspace W¶
Gradient-SVD on COCO pose features.
Train baseline / E1 / E1_aniso / VAT¶
E1_aniso +22.4 pp PCK vs baseline.
Robustness + embedding eval¶
Occlusion stress + drift metrics.
Run with matching-pmh¶
from pmh import PMHTrainer, evaluate_robust_fit
# nuisance="augmentation"
Do not use PMH when¶
Different skeleton or keypoint definitions at deploy.
Replace demo data with yours¶
Swap demo loaders for your train_loader, source_batches, target_batches, and deploy holdout. Hook the backbone before your task head.