CO3Dv2 indexed training data

Upload in progress. Consult dataset-status.json for completion and coverage. The target is every annotated frame in all 51 categories of the official co3dv2_231130 release. Partial uploads must not be interpreted as the full dataset.

Each all/<category>/part-NNNN/ prefix is directly readable with HFByteSource and SceneIndex. Each part has HDF5 payloads, NPZ indices, verified completion records, a packing plan, and a manifest published only after artifact verification. The final catalog.json lists every part only after all categories are complete.

RGB is resized without upsampling to a maximum height of 480 (even dimensions), encoded with H.264 CRF 18 and closed GOP 8. Cameras use OpenCV c2w and normalized intrinsics. Depth is float16 with a validity mask and Zstandard compression; depth and camera translations retain CO3D reconstruction units, not guaranteed metres. These are lossy training derivatives.

The all selection preserves all annotated frames, labels every frame as a training view, and does not implement an official benchmark train/val/test protocol. Original official split lists are retained in provenance/. Foreground probability masks, point clouds, and optical flow are not packed.

Source: Meta's CO3D release. Original source archives are downloaded and verified against the official SHA-256 list. Source archive hashes, original split metadata, category licenses, converter revision, and per-scene selected asset digests accompany the derivatives. See each category's original license under provenance/. Changes are the conversion, resizing, video/depth encoding, and camera convention conversion above.

Converter revision: f369c7fe6d545640646839e7ee9bd62704b505f0.

Citation: Reizenstein et al., Common Objects in 3D: Large-Scale Learning and Evaluation of Real-life 3D Category Reconstruction, ICCV 2021.