CaScenes

A multi-modal autonomous-driving dataset collected in the CARLA simulator, designed for BEV perception and SLAM research. CaScenes provides time-synchronized camera, LiDAR, radar, GNSS, and IMU streams together with BEV map ground truth, across multiple towns and weather conditions.

Highlights

• 48,453 keyframes across 96 segments in 3 weather conditions (Sunny, Night, Rainy) in CARLA Town 01.
• 6 surround-view cameras + 1 top LiDAR + 5 radars + GNSS + IMU + BEV map ground truth per frame.
• Ego pose (ego2global) and full sensor-to-ego calibration matrices included.
• Split into train (66 segments, 38,768 frames) and test (30 segments, 9,685 frames).

Splits

Split	Scenarios	Segments	Frames
train	Town01_Sunny / Town01_Night / Town01_Rainy (50 vehicles each)	22 + 22 + 22 = 66	38,768
test	Town01_Sunny / Town01_Night / Town01_Rainy (50 vehicles each)	10 + 10 + 10 = 30	9,685

Sensors

Each segment contains synchronized per-frame data under the following modalities:

Modality	Folder	Format	Notes
Front camera	CAM_FRONT/<ts>.png	PNG	800×450 (see camera_intrinsics)
Front-left camera	CAM_FRONT_LEFT/<ts>.png	PNG
Front-right camera	CAM_FRONT_RIGHT/<ts>.png	PNG
Back camera	CAM_BACK/<ts>.png	PNG
Back-left camera	CAM_BACK_LEFT/<ts>.png	PNG
Back-right camera	CAM_BACK_RIGHT/<ts>.png	PNG
Top LiDAR	LIDAR_TOP/<ts>.pcd	PCD
Radar (×5)	RADAR_{FRONT,FRONT_LEFT,FRONT_RIGHT,BACK_LEFT,BACK_RIGHT}/<ts>.*
GNSS	GNSS/<ts>.*
IMU	IMU/<ts>.*
BEV map (semantic)	BEV_MAP/<ts>.npy	NumPy
BEV map (color)	BEV_MAP_COLOR/<ts>.*
BEV view (rendered)	BEV_VIEW/<ts>.*
Ego pose	VEHICLE_TRANSFORM/<ts>.*		Also embedded in data.json as ego2global

Each segment also contains a data.json index listing every frame with full calibration and file paths.

data.json schema

{
  "infos": [
    {
      "weather": "Sunny",
      "timestamp": 7285258,
      "ego2global":  [[...4×4 row-major...]],   // CARLA world -> ego
      "lidar2ego":   [[...4×4...]],
      "lidar2global":[[...4×4...]],
      "lidar_path":  "CaScenes/datasets/test/.../LIDAR_TOP/26805.pcd",
      "bev_map":     "CaScenes/datasets/test/.../BEV_MAP/26805.npy",
      "cams": {
        "CAM_FRONT": {
          "data_path":         "CaScenes/datasets/test/.../CAM_FRONT/26805.png",
          "lidar2camera":      [[...4×4...]],
          "camera_intrinsics": [[...3×4 / 4×4...]]
        },
        "...": {}
      },
      "radars":  { "...": {} },
      "sweeps":  [ /* nearby unkeyed frames for temporal context */ ]
    }
  ]
}

All file paths inside data.json are relative to the parent of the CaScenes/ directory — i.e., extract the tarballs from a workspace root and reference files via the paths in data.json directly.

Download & extract

The dataset is distributed as plain (uncompressed) tar files. The train split is split into ~40 GB chunks (train.tar.part-aa, train.tar.part-ab, ...) to stay below HF's per-file LFS limit; concatenate them with cat before extracting. The test split fits in a single test.tar.

# from huggingface_hub (recommended)
pip install -U huggingface_hub
hf download Zixia3/CaScenes \
    --repo-type dataset \
    --local-dir ./CaScenes_release \
    --include "*.tar" "train.tar.part-*" "SHA256SUMS"

cd CaScenes_release
sha256sum -c SHA256SUMS    # verify integrity

# extract into a workspace where `CaScenes/datasets/...` should live
mkdir -p /path/to/workspace && cd /path/to/workspace

# train (split): cat parts back together and pipe into tar
cat /path/to/CaScenes_release/train.tar.part-* | tar -xf -

# test (single tarball)
tar -xf /path/to/CaScenes_release/test.tar

# resulting layout:
#   /path/to/workspace/CaScenes/datasets/train/Town01_Sunny_50_vehicles/segment_0/...
#   /path/to/workspace/CaScenes/datasets/test/Town01_Sunny_50_vehicles/segment_4a/...

A minimal Python loader that does download + checksum + extract in one shot is provided in download.py.

License

Released under Creative Commons Attribution 4.0 International (CC BY 4.0). You are free to use, share, and adapt the data, including for commercial purposes, provided you give appropriate credit.

Credits

CaScenes is collected using the CARLA open-source autonomous-driving simulator (MIT-licensed). If you use CaScenes, please also credit CARLA:

Dosovitskiy et al. CARLA: An Open Urban Driving Simulator. CoRL 2017.

Citation

If you use CaScenes in your research, please cite:

@misc{cascenes2026,
  title  = {CaScenes: A Multi-Modal CARLA Dataset for BEV Perception and SLAM},
  author = {Xia, Zixia and others},
  year   = {2026},
  url    = {https://huggingface.co/datasets/Zixia3/CaScenes}
}

Companion code

Methods built on CaScenes live at: https://github.com/ZixiaXia/SmartFusion-SLAM

Changelog

• v1.0 (2026-05-10) — Initial public release.