By requesting access to this dataset, you confirm that you have read and agree to the dataset license (CC BY-NC 4.0) and the additional usage terms below.

You agree to:

Use the dataset for research, development, and safety-related purposes
NOT use the dataset for military, surveillance, or related applications
NOT use the dataset in ways that could cause harm to individuals or society
Respect all applicable laws and regulations
Use the dataset in compliance with the CC BY-NC 4.0 license (non-commercial use only)
Cite CALM and CARD in any resulting work

CALM-Scenarios

The roads that CALM drives on. It holds 159 real speed bump crossings, 114 from Germany and 45 from Italy, each reduced to what a simulator needs: the height of the road along the approach, the position, height and width of the bump, and what the car's camera detected on the way. It also holds the split into training and evaluation roads and a set of synthetic roads. The images and raw sensor data behind the crossings are in CALM-Germany and CALM-Italy, which are built on the CARD recordings.

Files

File	Content
`germany/`, `italy/`	`crossings/` holds the real crossings, one JSON file each; `perception_model/` holds the camera model of that country
`synthetic/`	Synthetic roads: single bumps and single potholes (`KEY.md` explains the names)
`split.yaml`	The training and evaluation split
`perception_model_de_it.yaml`	How often the camera detects a bump and how far off its distance estimate is, for Germany and Italy together

Real crossings

Each file covers a window around one bump, normally 5 s before and 5 s after it. It holds the road height every 5 cm, the bump itself, and the camera detections. Where the recorded road does not begin or end within 1 cm of level, the uneven part at that end is cut off, but never closer than 0.5 m to the bump. 54 of the 159 crossings are marked low_confidence because the road height measured under the wheels shows no clear bump near the position the camera flagged. The other 105 are the qualified crossings that CALM trains and evaluates on.

bumps lists only one bump per crossing, the first one the camera flagged. 15 German crossings contain 2 to 5 bumps (ingolstadt_sb20 has 5), and for these the other bumps are not listed.

Example

ingolstadt_sb01.json, shortened:

{
  "name": "ingolstadt_sb01",
  "source_sequence": "Ingol1",
  "meta": {
    "ds_m": 0.05,
    "road_length_m": 31.75,
    "camera_height_m": 1.813,
    "trim": {"offset_m": 0.0, "original_road_length_m": 31.7, "bump_margin_m": 0.5}
  },
  "road": {"ds_m": 0.05, "z": [-0.00088, -0.00108, -0.00112, "... 635 values"]},
  "bumps": [{
    "s_bump_m": 21.05, "s_start_m": 20.3, "s_end_m": 21.8,
    "peak_z_m": 0.0342, "width_m": 1.5,
    "raw_peak_z_m": 0.0342, "raw_width_m": 1.45,
    "low_confidence": false
  }],
  "perception": {"samples": [
    {"d_true": 21.05, "detected": false, "d_hat": null, "bbox": null},
    {"d_true": 20.46, "detected": true, "d_hat": 23.38, "bbox": [0.497, 0.533, 0.117, 0.007]},
    "... 51 samples"
  ]}
}

road.z is the road height in metres, one value every ds_m = 5 cm, starting at the beginning of the trimmed road. This road is 31.75 m long.
bumps lists the bump with its position along the road (s_bump_m is the centre, between s_start_m and s_end_m), its height peak_z_m (3.4 cm) and its width width_m. The raw_ values are the detected ones before they are limited to the range the road model accepts; here the width of 1.45 m became 1.5 m.
perception.samples has one entry per camera frame on the approach. d_true is the true distance to the bump, detected tells whether the camera found it, and d_hat is the distance the detector estimated, 23.4 m instead of 20.5 m in the second sample. bbox is the box around the detection in the image.
meta.trim records the trimming: offset_m is the length cut from the start, original_road_length_m the length before trimming, edge_tol_m the 1 cm tolerance for level road and bump_margin_m the 0.5 m that is always kept next to the bump. source_sequence names the CARD recording the crossing comes from.

Perception model

The three perception model files (perception_model_de_it.yaml, germany/perception_model/perception_model_de.yaml, italy/perception_model/perception_model_it.yaml) describe how the camera detector behaves at different distances from a bump. They are fitted on the labelled camera frames of the recorded crossings and let the simulator add realistic detection errors. The true distance to the bump is cut into bins of bin_width_m (1 m), and every list has one entry per bin, starting at 0 m:

Key	Content
`max_range_m`	End of the last bin in metres. Beyond it the simulated camera never detects the bump.
`n_samples`	Camera frames in the bin
`n_ranged`	Detections in the bin with a distance estimate
`p_detect`	Share of frames in the bin in which the bump was detected
`range_bias`	Mean error of the estimated distance (estimated minus true) in metres
`range_std`	Standard deviation of that error in metres

The lists stop at the first bin with fewer than 8 frames. range_bias and range_std are null for a bin with fewer than 8 distance estimates; the simulator then adds no distance error in that bin.

Split

split.yaml lists the 14 evaluation roads under eval. Every other crossing is meant for training, except augsburg_sb06 under exclude: it is the same bump as the evaluation road augsburg_sb13, driven in the opposite direction, so it is left out. Of the 105 qualified crossings, 90 are used for training.

Use

The scenarios are read by the CALM repository. calm reproduce trains and evaluates a policy on them.

License

CC BY-NC 4.0. Non-commercial use only. The crossings are derived from the CARD recordings, which are released under CC BY 4.0 for Germany and CC BY-NC 4.0 for Italy.

Citation

Please cite CALM and the source recordings.

@misc{calm_dataset,
  title  = {CALM: Comfort-Aware Longitudinal Motion Planning},
  author = {Mohammed El Said Azab Abdelazim},
  year   = {2026},
  url    = {https://huggingface.co/CALM-Dataset}
}

@inproceedings{elazab2025card,
  title     = {CARD: A Multi-Modal Automotive Dataset for Dense 3D Reconstruction in Challenging Road Topography},
  author    = {Elazab, Gasser and Neuhaus, Frank and Ko{\ss}, Tilman and Splietker, Malte and Date, Aditya and Unterreiner, Michael and Jansen, Maximilian and Hellwich, Olaf},
  booktitle = {Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)},
  year      = {2026},
  eprint    = {2605.05014},
  archivePrefix = {arXiv}
}

Downloads last month: 4,156

Total file size:

2.69 MB

Paper for CALM-Dataset/CALM-Scenarios

CARD: A Multi-Modal Automotive Dataset for Dense 3D Reconstruction in Challenging Road Topography

Paper • 2605.05014 • Published May 7