Ours_Argoverse2_ADGS

Training-ready preprocessing artifacts for the AD-GS Argoverse 2 experiment. This release contains only scenes that passed the repository's full validate_splatad_scene.py --dataset av2 validation path.

Included release

Scene	Status	Cameras	Camera split	LiDAR split
a7bcdabb-f9b7-3c16-806d-3ddf1c2d49a2	strict validation passed	7 / 2,234 frames	1,120 train / 1,114 val	156 train / 156 val

The split is the SplatAD-style LINSPACE 50% protocol used by this experiment. The seven cameras are ring_front_center, ring_front_left, ring_front_right, ring_rear_left, ring_rear_right, ring_side_left, and ring_side_right.

Each released scene contains:

• image/: processed camera images.
• depth/: monocular depth priors.
• semantic/: semantic priors.
• sky/: sky masks.
• flow/: optical-flow priors for training observations.
• meta.npz: normalized camera, LiDAR, pose, and split metadata.
• colmap.ply: COLMAP point cloud.
• points3d.ply: segmented AD-GS initialization point cloud.

The unused intermediate backup points3d.unsegmented.ply, active staging directories, training checkpoints, logs, W&B data, and incomplete scenes are not included.

Integrity

release_manifest.json records the exact release inventory and validation summary. DATA_SHA256SUMS contains a SHA-256 digest for every data file.

Layout

scenes/
└── a7bcdabb-f9b7-3c16-806d-3ddf1c2d49a2/
    ├── image/
    ├── depth/
    ├── semantic/
    ├── sky/
    ├── flow/
    ├── meta.npz
    ├── colmap.ply
    └── points3d.ply

License and attribution

Argoverse 2 is distributed under the Creative Commons Attribution-NonCommercial-ShareAlike 4.0 International license. This release contains source-derived and transformed AV2 Sensor Dataset material and is distributed under the same license. See AV2_LICENSE_NOTICE.md and the official Argoverse 2 page.

Citation

@inproceedings{Argoverse2,
  author    = {Benjamin Wilson and William Qi and Tanmay Agarwal and John Lambert and Jagjeet Singh and Siddhesh Khandelwal and Bowen Pan and Ratnesh Kumar and Andrew Hartnett and Jhony Kaesemodel Pontes and Deva Ramanan and Peter Carr and James Hays},
  title     = {Argoverse 2: Next Generation Datasets for Self-driving Perception and Forecasting},
  booktitle = {NeurIPS Datasets and Benchmarks},
  year      = {2021}
}

AD-GS experiment code and preprocessing scripts: Dororo99/AD-GS.