task_categories:
- autonomous-driving
- computer-vision
language:
- en
pretty_name: LO-World nuScenes BEVLayout (2 Hz)
size_categories:
- 10K<n<100K
LO-World nuScenes BEVLayout (2 Hz)
This dataset contains the complete 2 Hz train/validation BEVLayout arrays used by LO-World. It contains 34,149 per-sample NumPy files: 28,130 train and 6,019 validation samples. Each file is named <sample_token>.npy and stores a binary uint8 array with shape (18, 200, 200) (720,128 bytes uncompressed).
Layout on the Hub
2hz/train.tar.zst # flat <sample_token>.npy files
2hz/val.tar.zst # flat <sample_token>.npy files
manifests/train.jsonl
manifests/val.jsonl
SHA256SUMS
Extract both archives into the repository's data/bev_layout/ directory (for example, tar --zstd -xf 2hz/train.tar.zst -C data/bev_layout). LO-World looks up files by nuScenes sample token.
Coverage and integrity
The files cover all 34,149 nuScenes trainval keyframes: 100% of the 850 official train/validation scenes, with no missing or extra sample tokens. Every array was read and checked for shape, dtype, and binary values. The split manifests record token, scene, timestamp, split, and per-file SHA256. SHA256SUMS provides the same file checksums in standard checksum format.
These are 2 Hz keyframe layouts, not dense 12 Hz layouts. They do not contain the intermediate frames in 12 Hz metadata. The selected 12 Hz / 81-frame condition packs are included under 12hz_train700_81f/ and 12hz_val150_81f/. Together they contain 850 selected scenes × 81 frames. Each clip combines its current 2 Hz keyframe with 80 future Layout frames. These selected windows do not cover every 12 Hz timestamp in nuScenes.
Data origin and terms
These arrays are derived from nuScenes and are provided for non-commercial research use. Users must obtain their own nuScenes access and comply with the current nuScenes terms of use and dataset information. This repository does not include nuScenes sensor data or metadata, and does not grant rights to the underlying nuScenes dataset.
Please cite both LO-World and nuScenes. The official nuScenes information page describes the dataset as available strictly for non-commercial purposes. No broader license for nuScenes content is granted by this repository.
Citation
Please cite both LO-World and nuScenes when using these layouts. nuScenes citation:
@article{nuscenes2019,
title={nuScenes: A multimodal dataset for autonomous driving},
author={Caesar, Holger and Bankiti, Varun and Lang, Alex H. and others},
journal={arXiv preprint arXiv:1903.11027},
year={2019}
}