bigwolfe/gaussian-jepa-3d-data

Gaussian JEPA 3D scene data

Object-local 3D Gaussian fits to 24-frame, 768 × 768 Kubric MOVi-E videos. Each clip has camera and object poses. Aligned files provide projected 2D Gaussian tokens and world-space 3D targets.

The current archive has 301 complete scene, token, and target triples. It also has 92 token and target pairs without saved scene fits. Seven clips have .failed markers. The 1,087 data files total 69,225,016,637 bytes. manifest.tsv gives each file's size and SHA-256 digest. A full triple and its comparison image appear in sample/.

This data is too small and too narrow to settle persistent 3D object tracking. On one 12-clip MOVi validation probe, there was one heavy occlusion followed by a reveal and no full occlusion. The MOVi camera also follows an almost straight path, which makes moving-object depth ambiguous. We generated 804 additional 2D Gaussian trajectory clips across three motion sets; those are not extra fitted 3D scenes and are not in this archive.

Load the included clip with python load_clip.py sample clip00000. Use python verify_manifest.py . --sample to check its hashes, or omit --sample to hash the full archive. These files use PyTorch serialization; load only files from a trusted source.

The data is CC BY 4.0. The original utility code here is Apache 2.0. Google Scanned Objects supplies the scanned objects under CC BY 4.0. Poly Haven supplies the HDRI assets under CC0. We rendered new clips and fitted Gaussians with Kubric and gsplat. See ATTRIBUTION.md for source and modification details.