Ours-V2_waymo_ArmGS

Preprocessing artifacts for the 10 selected scenes/logs used by this ArmGS workspace. This is not the entire upstream dataset and not an evaluation/trained-model repository.

Includes native three-camera decoded RGB PNGs, sky masks, camera-frame manifests, canonical CPU tensor caches, COLMAP mappings and final triangulated point clouds. Intermediate COLMAP databases/images are intentionally omitted. These caches do not include the full original Waymo parquet dataset.

Layout and use

Files retain their ArmGS-workspace-relative paths (data/... and configs/...). Download to a separate directory and inspect before copying into a matching ArmGS checkout. These are preprocessing artifact files, not a Hugging Face datasets.load_dataset table.

Portability: JSON mappings and Torch caches may contain absolute source-workspace paths. They are preserved as provenance, not silently rewritten. Rebuild canonical manifests using the matching ArmGS preparation command, or carefully rebase path fields for your local dataset/workspace before training. Do not assume this snapshot runs without the original dataset or without path adjustment.

Torch cache safety: .pt files contain prepared data, not trained model checkpoints. Only load trusted artifacts; when inspecting, use CPU torch.load(..., weights_only=True). SHA256 checks detect accidental changes and do not establish trust in an unknown publisher.

Scope: Training checkpoints, optimizer states, training logs, W&B data, credentials, quarantined/obsolete variants and visualization renders are excluded. Decoded dataset RGB PNGs and segmentation-mask PNGs, where included, are training inputs rather than visualization renders.

Upstream data: No new license to upstream data is granted by this snapshot. Users must obtain required source data through the original dataset providers and comply with their terms.

See UPLOAD_FILE_MANIFEST.json for byte sizes and SHA256 digests of all data/configuration files.