ActionBench: 128 Paired Video and Animated 3D Point-Cloud Samples

Actionbench is a benchmark of 128 paired video and animated point-cloud samples created by Meta (Facebook) for evaluating 3D mesh generation. Each record contains 16 RGBA frames with alpha masks and a corresponding animated point cloud consisting of 16 keyframes sampled from mesh surfaces.

Use Cases

• Benchmarking video-to-3D generation by comparing output to the (T, V, 6) point cloud data
• Training motion-aware encoders using the 16 RGBA frames
• Validating surface reconstruction using the sampled mesh vertices

Strengths

• 128 paired video-3D samples
• Includes alpha masks for all 16 frames
• 16-keyframe temporal resolution for 3D point clouds

Limitations

• Small sample size of 128 records
• Synthetic data may lack real-world sensor noise
• Limited to 16 frames per sequence

Provenance

Source

Meta (Facebook)

Collection Method

Synthetic generation and sampling from animated 3D meshes

Freshness

Last updated March 2026.