Kavin60606/hand-stress-suite-data

Egocentric Hand Stress Suite

Curation + world-space benchmark of egocentric hand pipelines under two stress axes: A1 object-induced occlusion and A2 bimanual overlap. Sources: HOT3D (Aria), ARCTIC (ego), H2O (reserve). Frozen set v1.1 = 100 clips (50 stress, 50 generic; HOT3D 59, ARCTIC 41). Pipelines round 1: HaWoR, MINT, Dyn-HaMR.

Repos

• Kavin60606/hand-stress-suite (this, public): code, docs, results, index CSVs, frozen set.
• Kavin60606/hand-stress-suite-data (private): exported clips + GT at repo root (<clip_id>/{video_pinhole.mp4,video_fisheye.mp4,intrinsics.json,gt.npz,meta.json}; HaWoR raw outputs under <clip_id>/video_pinhole/), adapter outputs (preds/<pipe>/<clip>/pred.npz), Dyn-HaMR raw fits (<date>/<clip>-all-shot-0-*/smooth_fit/), full per-frame indexes (index_full/, parquet), overlay renders (verify_overlays/). Private because clips are derived from HOT3D / ARCTIC (their licences forbid redistribution).

Layout

Path	What
docs/	Curation Axes v0.2, Benchmark Reference v0.3, Benchmark Results v0.2 (PDF) + their ReportLab builders
docs/frozen_set_v1.1.csv	the 100-clip frozen set (set, axis, subtype, source, sequence, window, split, thresholds)
curation/code/	GT-rendered occlusion scorers (score_hot3d_gpu.py, score_arctic.py, score_h2o.py, gpu_score.py nvdiffrast), rules.py (single source of subtype rules), mine_events.py, mine_generic.py, freeze.py, export_clips.py, review dashboard/
curation/index/	candidate / event / generic CSVs per source, review.sqlite (keep/drop decisions), frozen csvs
benchmark/code/	eval/metrics.py + eval/run_eval.py (shared metric layer), adapters/{mint,hawor,dynhamr}_adapter.py (→ pred.npz contract), env build scripts, per-pipeline runners (run_*_all.sh), sweep_dynhamr.sh
benchmark/patches/	git diff + SHA of upstream HaWoR / Dyn-HaMR / MINT checkouts (focal + intrinsics env overrides, GPU pin removal, --no_vis, BMC assets off)
benchmark/envs/	pip freeze of /venv/hawor, /venv/dynhamr, /venv/mint-inference
benchmark/logs/	eval / adapter / runner logs
results/	<pipe>_per_clip.csv (one row per clip per hand, all metrics) + <pipe>_summary.json

Conventions

• MANO-21 joints = smplx 16 + tips [744,320,443,554,671]; pipelines emit OpenPose order → adapters apply inverse of [0,13,14,15,16,1,2,3,17,4,5,6,18,10,11,12,19,7,8,9,20].
• HOT3D input: upright pinhole f=520 @ 1408², pose = T_world_cam_fisheye · Rz(−90°). ARCTIC: cv2.undistort at 0.5× (1400×1000).
• Metrics: PA-MPJPE, MPJPE-cam (wrist-relative), WA/W-MPJPE-100 (WHAM alignment), RTE, ATE / arc ratio, accel error, coverage / false presence, MPJPE-p, Action-MPJPE (Macrodata), L/R swap, event vs context, failure (W-MPJPE > 100 mm or coverage < 0.5). Pose metrics gated on GT-visible frames.

Reproduce

1. Build envs per benchmark/code/build_*.sh (torch 1.13 + cu117 for HaWoR / Dyn-HaMR; create_env.sh for MINT); apply benchmark/patches/*.diff on the listed SHAs.
2. Download the <clip_id>/ folders from the data repo into /workspace/bench/clips/.
3. run_mint_all.sh, run_hawor_all.sh <gpu> <shard> <nshards>, run_dynhamr_all2.sh <gpu> 0 1 (claim-runner; check runtime.json exit code).
4. python adapters/<pipe>_adapter.py (CPU MANO) → python eval/run_eval.py --pipe <pipe> → results/.
5. python docs/build_results.py <ver> → PDF.