📖 Overview

ActionBench is a benchmark dataset of 128 paired video ↔ animated point-cloud samples for evaluating animated 3D mesh generation from video. The dataset consists of synthetic scenes of animated objects from ObjaverseXL, rendered using Blender 3.5.1.

Each sample contains:

• Video: 16 RGBA frames with alpha mask

• Camera (camera.json): Camera parameters using Blender convention (X_cam = X @ R^T + T, camera looks along -Z). See projection.py for how to project the point cloud onto the image plane.

• Animated Point Cloud: Surface points sampled on the animated object with shape (T, V, 6) where:

  • T=16: number of keyframes
  • V=100_000: number of vertices (points randomly sampled on the mesh surface)
  • 6: position (x, y, z) + normal (nx, ny, nz) for each point

  Note: The point cloud is tracked: each point index corresponds to the same surface point deformed across timesteps, providing dense correspondences over time.

  The animation lie in normalized space [-1., 1.]^3.

📊 Evaluation

To evaluate on ActionBench, produce a list of animated meshes saved as .glb files.

Each subdirectory must be named with the corresponding uid from ActionBench:

predictions/
├── <uid_1>/
│   ├── mesh_00.glb
│   ├── mesh_01.glb
│   └── ...
├── <uid_2>/
│   ├── mesh_00.glb
│   └── ...
└── ...

Download Actionbench dataset, then run the evaluation script in ActionMesh:

python actionbench/evaluate.py \
    --pred_root predictions/ \
    --gt_root data/actionbench/data/ \
    --output_csv results.csv \
    --device cuda

Note: Evaluation requires the same dependencies as ActionMesh plus PyTorch3D.

Metrics are described in the ActionMesh paper:

• CD-3D: Chamfer Distance 3D — measures geometric accuracy per frame
• CD-4D: Chamfer Distance 4D — measures spatio-temporal consistency
• CD-M: Motion Chamfer Distance — measures motion fidelity

🏛️ License

See the LICENSE file for details about the license under which this dataset is made available.

📚 Citation

If you use ActionBench, please cite the following paper:

@inproceedings{ActionMesh2026,
  author = {Remy Sabathier and David Novotny and Niloy Mitra and Tom Monnier},
  title = {ActionMesh: Animated 3D Mesh Generation with Temporal 3D Diffusion},
  year = {2026},
}