SAM3D canonical-init-frame fix: 12 clips refreshed (sanity 5 + batch 7) (6/12)

2026-05-15: state machine v2 — STATIC / GRASPED_L/R/BOTH / MOVING + L/R-aware grasp (34/105)

Upload 112 clip(s) (4.99 GB) — --oo8_XIuOM_799.5_809.8, --oo8_XIuOM_900.3_917.4, -0RheyDV3a0_474.8_487.3, ... (+109 more) [2026-05-12T20:49:27+00:00] (34/112)

v1.11: pure canonical_scale (hf_reference simplified). No ICP / Umeyama / sanity refinement. Mesh scale = SAM3D canonical_scale directly; R = canonical_rotation_quat; per-frame translation = obs centroid + Gaussian smooth.

v1.10: restore original 5-stage tracking (estimate_anchor_pose FGR+ICP+Umeyama + track_6dof flow propagation). 40/225 oids get full 6DoF tracking; rest fall back to centroid translation + identity rotation. Scale closer to A100 reference (cereal box 0.219 vs reference 0.263).

v1.9: scale_correction is now ABSOLUTE (PLY units→world m, matches A100 reference convention). Robust min-enclose algorithm: per-axis [2,98]% pre-filter + median over axes of per-axis 90% enclose (depth-fringe-resistant)

v1.8: tracking uses SAM3D canonical_rotation_quat (per-clip initial orientation) + min-scale-95% enclose algorithm; object_pose.bin now reads pose_track_info.T_seq (full 6DoF tracked pose, not stale PCA OBB)

v1.7: SAM3D refill round 2 (226 PLYs, +25 from v1.6) — decoder cleanup frees 1.7GB GPU; native-res bbox crop (SAM3D internally resizes to 518)

v1.3: SAM3D meshes (164 PLYs, ~383 MB) + viser scenes with mesh layer enabled (Plan B mask scoring: area * (1-hand_iou) * static_ratio)

v1.1: 105 clips, SAM3-quality-sorted index (best→worst), WiLoR-native skeleton overlay

WiLoR-native joints_2d for skeleton overlay (88 clips re-exported, MoGe-2-independent 2D projection bypasses focal estimation errors)

Fix trim consistency: scene.json fps/duration use trim window (was 2x bug), video_source pinned to trim, frontend trim review pre-loads saved values

Retry 2 clips: silver-jar→silver-can prompts, pat-turkey unified bowl-of-turkey compound

Step 1 full re-run @ 15fps with compound+fallback prompts and containment merge (55 clips)

Re-export 21 clips: scene.json objects/[] now populated from sam3_obj_data when no SAM3D mesh; silver-can clip uses 'silver can' instead of 'can of olives' (SAM3 hits 4/4 vs 3/4)

Replace samples/ with 21-clip Step 1 test batch (no SAM3D, no 6DoF)

Refresh scene.json: surface Action100M action_detailed / summary / video_title; sync index.json

Refresh dataset: 48 trim-reviewed clips, manual SAM3 prompts, full descriptions in scene.json

Initial deploy