MV4DBench

A 31-item multi-view 4D animated mesh benchmark for evaluating image/video → 4D mesh reconstruction methods. Each item is rendered from 8 azimuths × N source frames and ships with a tracked ground-truth point-cloud (T=16, N=100k, xyz+normal) for chamfer distance evaluation.

Items

31 animated GLBs sourced from a curated subset of 3D animations (mixamo characters, animal/object scans, etc.). Frame counts range 11 to 757; renders are at 1024×1024, 24 fps, radius=2.2, elevation=15°.

Layout

<uid>/
├── source.glb              # original animated mesh
├── meta.json               # camera matrices, frame count, fps, etc.
├── rgba_8views/view_NN_azDDD/frame_*.png   # 8 azimuths, transparent bg
├── rgb_8views/view_NN_azDDD/frame_*.png    # 8 azimuths, white bg
├── anim_4views/view_azDDD.mp4               # 4 azimuths, MP4 videos
└── gt_surfaces.npy   # (16, 100000, 6) tracked GT point cloud, normalized to [-1,1]^3

The gt_surfaces.npy is sampled at 16 evenly-spaced source-frame indices, with barycentric tracking on the rest-pose mesh, then normalized so the entire animation fits in a [-1,1]^3 cube (matching the bpyrenderer scene-normalization used during RGBA rendering — so chamfer scores stay scale-aligned).

UID mapping

Some original UIDs included whitespace/| chars; those are sanitized to the leading hex stem here. uid_mapping.json records the original full names for traceability.

Citation

Used as held-out evaluation for the multi-view test-time-adaptation extension of ActionMesh (Meta, 2024). See accompanying paper for chamfer benchmark numbers.