AsyncBEV asynchronous data supplement

Supplementary sweep maps and flow labels for AsyncBEV: Cross-modal Flow Alignment in Asynchronous 3D Object Detection (ICLR 2026).

Code and setup instructions · Paper · Project page

This supplement is used with the CMT implementation of AsyncBEV. Obtain the original nuScenes v1.0-trainval dataset separately from nuScenes, and prepare the CMT/MMDetection3D info files as described in the code repository. This repository does not include the original sensor samples, sweeps, maps, model checkpoints, or data-generation scripts.

Release contents

The source inventory contains 593,493 pickle files totaling 539,568,241,773 bytes before compression:

Content	Source files	Uncompressed bytes
Camera flow labels	228,304	206,758,945,911
LiDAR flow labels	365,185	330,447,153,615
Sweep maps and history indices	4	2,362,142,247

The flow files are packaged into 512 independent .tar.zst archives: 256 each for camera and LiDAR, grouped by the first two hexadecimal characters of the sample token. Each archive preserves paths rooted at async_supplement/. The four index files are provided directly under async_supplement/:

• cams_sample_sweeps_map_ego_sample_w_ego_pose.pkl
• cams_sample_sweeps_map_ego_sample_w_ego_pose_timestamp.pkl
• lidar_sample_sweeps_map_w_extrin_ego_sample_ref_coord.pkl
• token_history_dict.pkl

manifest.json records payload sizes, source file counts, and SHA-256 checksums. SHA256SUMS covers every archive and index file. The final manifest and checksum file are published only after all payload files have been uploaded and verified.

Download and extract

Install the Hugging Face CLI in your download environment, and ensure GNU tar and zstd are available. The extraction requires approximately 540 GB in addition to the compressed download size. Consult the manifest for the actual compressed sizes.

python -m pip install -U huggingface_hub
hf download ShimingWang/AsyncBEV --repo-type dataset --local-dir asyncbev_download
cd asyncbev_download
sha256sum -c SHA256SUMS

# Set this to the directory containing your nuScenes samples/, sweeps/, and maps/.
NUSCENES_ROOT=/path/to/nuscenes
mkdir -p "$NUSCENES_ROOT/async_supplement"
cp -n async_supplement/*.pkl "$NUSCENES_ROOT/async_supplement/"
for archive in archives/async_cam_flow/*.tar.zst archives/async_lidar_flow/*.tar.zst; do
  tar --zstd --keep-old-files -xf "$archive" -C "$NUSCENES_ROOT"
done

Extract into a new supplement directory. The commands preserve existing files and report collisions rather than replacing an existing installation. The resulting flow paths are:

async_supplement/async_cam_flow/<sample_token>/<async_data_token>.pkl
async_supplement/async_lidar_flow/<sample_token>/<async_data_token>.pkl

For reproducible downloads, use hf download ... --revision <commit-sha> with the dataset revision you intend to use.

Compatibility and validation scope

Use the matching CMT/MMDetection3D environment and info files described in the code repository. The loaders consume stored paths from the sweep maps; placing the supplement under a different directory does not rewrite those paths. Check the configured data paths before running. The released asynchronous test pipelines require flow labels during evaluation as well as training.

This release preserves the serialized source files without converting or deserializing them. Packaging and upload checks establish file inventory and byte integrity; they do not establish coverage of every configuration or reproduce the paper's experiments. The repository contains Python pickle files; load them only in an environment where you trust their source.

Citation

@inproceedings{wang2026asyncbev,
  title={AsyncBEV: Cross-modal Flow Alignment in Asynchronous 3D Object Detection},
  author={Wang, Shiming and Caesar, Holger and Nan, Liangliang and Kooij, Julian F. P.},
  booktitle={International Conference on Learning Representations},
  year={2026}
}