Dense, multiview geometric supervision for dynamic scenes is scarce, and that scarcity restricts progress on 4D reconstruction, tracking, and geometry-aware novel-view synthesis. Syn4D supplies large-scale, fully synthetic multiview video with dense, queryable geometry so models can learn long-range spatiotemporal correspondences and geometry-consistent rendering.
What Sets It Apart
- Dense, queryable 3D tracks: every pixel can be unprojected to 3D at any time and projected into any other camera/time via an efficient barycentric-map + mesh representation, enabling complete long-range 3D tracking benchmarks.
- Multiview + geometry + humans: synchronized multi-camera RGB, per-pixel depth, camera trajectories, instance/mask metadata and SMPL‑X body pose/shape annotations in the same corpus — so tasks from pose estimation to 4D reconstruction can be trained jointly.
- Scale and diversity: ~4.7K multiview clips (≈1.4M frames) rendered in Unreal Engine using a wide catalog of environments and 1,674 animated assets plus 585 simulated humans, curated to encourage generalization.
- AI‑training license and tooling: all 3D assets are licensed for AI training, dataset packages include metadata mappings, visualizer code, and Kubric-style subsets to ease integration into training pipelines.
Who It's For and Trade-offs
Great fit if you need dense, geometry-consistent supervision for research or development in 4D reconstruction, multiview depth/camera estimation, long-range 3D tracking, or geometry-aware novel-view synthesis. Also useful for pretraining or augmenting human pose/SMPL‑X models when multiview geometry helps.
Look elsewhere if you require real-world photographic noise, raw sensor artifacts, or datasets for facial identity/facial-recognition tasks (human characters are synthetic and BEDLAM2-derived motion metadata is partially restricted). Be aware the full release is large (storage and decompression required) and synthetic-to-real domain gaps remain a consideration when transferring to real data.
