CARLA-MWRS

CARLA-MWRS (CARLA Multi-Weather Road Segmentation) is a fully synthetic, paired RGB--geometry benchmark for urban-road segmentation. It was prepared for IAF-Net: Illumination-Adaptive Fusion for Low-Light Urban Road Segmentation. The canonical protocol deliberately holds out a town: Town05 and Town06 are used for training, and Town04 is used only for validation. The four weather conditions are balanced within each split.

Release contents

split	towns	samples	per-weather count	archives
training	Town05 + Town06	2,400	600 × 4	training/*.tar.zst
validation	Town04 (held out)	1,200	300 × 4	validation/*.tar.zst
total	3 towns	3,600		21,600 files

Weather names are ClearDay, ClearNight, HeavyFoggyNight, and HeavyRainFoggyNight. A sample stem has the form Town06_HeavyFoggyNight_000646; the same stem occurs in all six modalities. The selection protocol and counts are recorded in dataset_manifest.json and the original reconciled manifests under metadata/.

The repository stores 12 deterministic tar.zst archives (one per split and modality) instead of tens of thousands of individual files. This is a packaging choice only; every source file is included exactly once. Archive members retain their modality directory, for example image_2/Town05_ClearNight_000224.png.

Download

python -m pip install -U huggingface_hub
hf download PeterNano/CARLA-MWRS --repo-type dataset --local-dir CARLA-MWRS

The archives are large. hf download can be resumed after an interrupted transfer. Verify SHA256SUMS.txt before extraction.

Extract one or all modalities

Install a tar implementation with Zstandard support (GNU tar 1.31+ or the zstd command-line tool), then run from the repository root:

mkdir -p training validation
for f in training/*.tar.zst; do tar --zstd -xf "$f" -C training; done
for f in validation/*.tar.zst; do tar --zstd -xf "$f" -C validation; done

Each archive extracts into its modality directory. The split lists in splits/training.txt and splits/validation.txt contain one stem per line; the TSV files additionally record town, weather, and frame index.

File formats and dimensions

The simulator materializes files at 384 × 1248 (height × width). The training pipeline resizes them to 512 × 1024 (height × width); these are different quantities and are both recorded intentionally.

modality	file	shape when loaded	dtype / encoding
RGB	image_2/*.png	384×1248×3	uint8, RGB PNG
road label	gt_image_2/*.png	384×1248	uint8, background 0, road 255
depth (integer)	depth_u16/*.png	384×1248	uint16, millimetres; values are clipped at 65.535 m
depth (metric)	depth_meters/*.npy	384×1248	little-endian float32, metres
surface normal	normal/*.npy	3×384×1248	little-endian float32, channel-first camera-frame unit vectors
calibration	calib/*.txt	text	P2 3×4 camera projection and Vehicle_pos metadata

For the canonical materialization, the integer depth is related to the metric array by

depth_u16 = clip(floor(float32(depth_meters) * float32(1000)), 0, 65535).astype(uint16)

The float32 qualification matters for bit-exact reproduction at values that lie exactly on a millimetre bin boundary.

Consequently, values at 65535 are saturated in the integer representation; use depth_meters when ranges beyond 65.535 m or the original float precision are needed. A zero depth value denotes an invalid/no-return pixel where it occurs. Normal arrays contain finite unit vectors (within floating-point tolerance); users should mask invalid pixels according to their loader.

Reproducibility and verification

The release is fixed by selection seed 42 and the reconciled Town split. The following commands validate the source tree, archive member counts, dimensions, dtype constraints, paired stems, weather counts, and recorded checksums:

python -m pip install -r requirements.txt
python scripts/validate_release.py --data-root /path/to/unpacked/CARLA-MWRS \
  --report validation_report.json
python scripts/validate_archives.py --release-root /path/to/CARLA-MWRS
sha256sum -c SHA256SUMS.txt

dataset_manifest.json gives the release-level archive sizes and hashes; SHA256SUMS.txt covers every published metadata file and archive. The reconciled source manifests are retained for provenance, while the private source directory itself is not required for a user who only downloads the archives.

Generation and limitations

The images, binary road masks, metric depth, camera-frame normals, and calibration were rendered/materialized from CARLA scenes under the four named weather presets. This v1.0.0 release is the frozen materialized output; the private rendering project, simulator executable, exact CARLA build, and trajectory-generation scripts are not redistributed here. The release manifests make the sample selection and file-level contents reproducible, but identical pixels cannot be regenerated without the corresponding upstream and rendering environment. This is not a real-world capture and does not claim to model all sensor noise, traffic behavior, or deployment conditions. Labels can contain simulator boundary and occlusion artifacts. Results on this release should be reported with the held-out-town protocol and should not be interpreted as evidence of real-world safety.

License and citation

See LICENSE_DATA.md for the CARLA/asset license boundary and the MIT terms for the helper scripts. Do not infer an unrestricted MIT license for rendered assets solely from the CARLA source-code license.

If you use this release, cite both CARLA and IAF-Net:

@misc{anonymous2026carlaMWRS,
  title        = {CARLA-MWRS: CARLA Multi-Weather Road Segmentation},
  author       = {{Anonymous Authors}},
  year         = {2026},
  howpublished = {Hugging Face Dataset: PeterNano/CARLA-MWRS},
  note         = {Version 1.0.0}
}

@inproceedings{dosovitskiy2017carla,
  title     = {CARLA: An Open Urban Driving Simulator},
  author    = {Dosovitskiy, Alexey and Ros, German and Codevilla, Felipe and Lopez, Antonio and Koltun, Vladlen},
  booktitle = {Conference on Robot Learning},
  year      = {2017}
}