nuplan-mini-sample / README.md
runfuture's picture
Upload README.md with huggingface_hub
e366775 verified

NuPlan mini sample (VADv2-ready)

This dataset contains a single NuPlan mini log converted into the VADv2 training/evaluation format.

Contents

nuplan_data/
  maps/                              # Full NuPlan map bundle (gpkg, npy, json)
  nuplan-v1.1/
    sensor_blobs/2021.05.12.22.00.38_veh-35_01008_01518/
        CAM_* folders + MergedPointCloud/   # Cameras and fused lidar frames
    splits/
        mini/2021.05.12.22.00.38_veh-35_01008_01518.db   # Mini split SQLite DB
converted_data/
  2021.05.12.22.00.38_veh-35_01008_01518.pkl            # Raw converter output
  us-nv-las-vegas-strip_map.json                         # Converted map JSON
annotations/
  2021.05.12.22.00.38_veh-35_01008_01518_infos.pkl      # VADv2-ready info dict
  us-nv-las-vegas-strip_map.json                         # Copy of map JSON

The annotations/*.pkl file can be plugged directly into adzoo/vad/configs/VAD/VADv2_voca4096_mini_config.py after setting data_root to data/nuplan/dataset.

Provenance & Conversion

  • Source: motional-nuplan/public/nuplan-v1.1/ on AWS S3 (mini split).
  • Conversion: Nuplan-VAD-DataGen executed in a Python 3.10 environment.
  • Only the log 2021.05.12.22.00.38_veh-35_01008_01518 is fully materialised (cameras + lidar).

Notes

  • Map annotations (annotations/us-nv-las-vegas-strip_map.json) are provided in both the original converter output and the annotations directory.
  • The maps/ directory includes the full NuPlan map bundle so that map APIs work without additional downloads.
  • The camera folders (8x CAM_*) are included. Due to the Hub's per-directory limit (10k files), the MergedPointCloud lidar frames are not part of this snapshot; the paths remain in the annotations so you can drop the original files into nuplan_data/nuplan-v1.1/sensor_blobs/2021.05.12.22.00.38_veh-35_01008_01518/MergedPointCloud/ if you need them.
  • The mini split SQLite DB is the only one required for this log; additional DB files from the mini bundle are not included to keep the footprint small (~37 GB total).

Usage

  1. Download the dataset and place it under your project root.
  2. Point data_root / ann_file_* in the VADv2 config to the paths above.
  3. Ensure the environment uses NumPy ≥ 2.0 to read the pickled infos (or add a compatibility shim).
python adzoo/vad/test.py \
  adzoo/vad/configs/VAD/VADv2_voca4096_mini_config.py \
  ckpts/VADv2_checkpoint.pth \
  --launcher none --eval bbox --json_dir test/mini_eval

License

  • NuPlan data is subject to the Motional NuPlan license.
  • Converted artifacts are provided for research purposes only.