跳转至

D3D-HOI One-Stop Evaluation State

Project/legacy operations page。 下文 Full20 review 是 2026-07 development artifact，不是当前 formal evaluator、paper-frozen metric 或 target-local Reconstruction evidence。正式协议见 Evaluation protocol，结果边界见 Results。

Objective

Evaluate final HOI-align result.pt files with object and HOI metrics, render hoi.mp4, build JSON/CSV/HTML review outputs, and publish the timestamped version through Tailscale.

Historical state and current target audit

  • A read-only 2026-07 checkout recorded a Full20 review under version d3dhoi-default-v1.
  • The historical relative paths were output/runs/d3dhoi_default/eval/hoi_align_review/20260714_183710_d3dhoi-default-v1 and output/public_shares/full_hoi_review_20260714_183710_d3dhoi-default-v1; neither exists in the current target.
  • Current target output/runs/d3dhoi_default/ has no result.pt, review JSON/CSV or Full20 MP4 evidence.
  • The old review tunnel is retired. The current research-site URL serves documentation only, not this bundle.
  • The historical report claimed 20 cases with object/HOI metrics and MP4s, but it needs read-only source provenance and hashes before reuse; it is not a target formal result.
  • The review title format is Full20 HOI Review · <timestamp> · <version>.
  • Canonical videos live beside each result at recon/output/hoi.mp4 and are reused when result.pt is unchanged.
  • The RGB renderer batches the three views per frame and runs two frames concurrently on separate CUDA streams.

Render benchmark

Case: b001-0010_cad12561, 120 frames, 2880x1280 output at 30 FPS.

  • Original serial renderer: 63.37 s.
  • CUDA-stream chunk 2: 48.38 s and 50.39 s; mean 49.39 s.
  • Speedup: 1.28x; elapsed time reduced by 22.1%.
  • Chunk 4/8/16/32: 49.13/49.61/49.98/51.27 s.
  • Default frame chunk: 2.

Validation

  • Real-frame human RGB max difference: 1.788e-7; alpha and depth differences: 0.
  • Real-frame object RGB max difference: 5.960e-8; alpha and depth differences: 0.
  • The old report recorded 20 JSON rows, 20 CSV rows and 20 MP4s; these files are target-missing today.
  • Historical HTTP/test checks do not validate the current public docs or current evaluator.

Next command

Only after a new formal manifest produces target-local result.pt artifacts may a fresh review be run. The old command below is illustrative project history, not authorization to run a large experiment:

/DATA/intern/hoi4d/miniconda3/envs/arthoi4d/bin/python \
  scripts/experiments/evaluate_d3dhoi_hoi_results.py \
  --cases-root output/runs/d3dhoi_default \
  --version d3dhoi-default-v1 \
  --gpus 0,1,2,3,4,5,6,7