EgoHand4D · Benchmark

Benchmark Gallery — every clip, all methods in one frame

One video per clip, everything side by side: row 1 = OURS (current champion : raw · MANO-AR · 3D orbit + ego-cam) · row 2 = Dyn-HaMR (corrected-K hardened fork, same shared renderer) · row 3 = MediaPipe (industry 2D baseline — 2D keypoints only, it has no 3D). Cells regenerate with the new champion on every model iteration (`bench_compare.py --render`). Clips whose baselines are still computing show the champion 4-panel as a labelled fallback. ■ left ■ right · ← back to the report

Iteration overview — every experiment, promoted or rejected (click a row to watch that model)

trainedmodelstepschangelab mean5wild-vs-MP