Weather Robustness Datasets for Perception and Artifact Removal
Last updated: 2026-05-23
This index summarizes adverse-weather driving datasets that are useful for validating perception degradation, LiDAR artifact removal, and sensor-fusion fallback behavior. The emphasis is not only algorithm selection, but also whether the validation data can expose failures caused by snow, rain, fog, wet-road spray, steam-like aerosol, dust-like obscurants, and asymmetric sensor degradation.
Related research pages: LiDAR artifact removal techniques, radar-LiDAR fusion in adverse weather, Airport-FOD3S synthetic FOD data engine, production perception systems
Dataset Coverage Matrix
| Dataset | Primary adverse condition | Modalities | Labels | Best validation use |
|---|---|---|---|---|
| WADS | Falling snow, accumulated snow, whiteout-like winter driving | LiDAR, visible/NIR/LWIR cameras, radar, GNSS/IMU | Dense point-wise LiDAR labels with snow classes | Snow removal, snow segmentation, snow-aware mapping |
| CADC / CADC+ | Canadian winter driving, paired snow and clear sequences | 8 cameras, VLP-32C LiDAR, GNSS/INS | 3D boxes; CADC+ adds paired clear/snow evaluation | Snow domain shift, de-snowing, 3D detection degradation |
| SemanticSTF | Rain, snow, light fog, dense fog | LiDAR, RGB imagery, calibration/weather metadata | Dense point-wise semantic labels | All-weather 3D semantic segmentation and domain generalization |
| REHEARSE-3D | Emulated heavy rain | LiDAR-256, 4D radar, rain-characteristic metadata | Point-wise rain/no-rain annotations | LiDAR point-cloud de-raining and radar-conditioned removal |
| RainSense | Natural rainfall with measured intensity | Camera, LiDAR, 4D mmWave radar, disdrometer | 2D/3D target boxes by 10-second case | Rain-intensity response curves and modality degradation |
| SemanticSpray++ | Wet road surface and road spray | Camera, VLP32C LiDAR, Ibeo LiDARs, Aptiv radar | Camera 2D boxes, LiDAR 3D boxes/semantics, radar semantics | Spray/wet-road robustness and radar-LiDAR fusion checks |
| RADIATE | Rain, fog, snow, night, clear baselines | Navtech radar, stereo camera, 32-channel LiDAR, GPS/IMU | 2D radar-image boxes for 8 actor classes | Radar-first adverse-weather detection and fusion fallback |
| DSERT-RoLL | Clear, fog, rain, snow, varied lighting | Stereo RGB, event, thermal, 4D radar, dual LiDAR, GPS/IMU | 2D/3D boxes, track IDs, odometry | Cross-modal event/thermal/radar/LiDAR fusion under weather and lighting shifts |
| CMHT Autonomous Dataset | Rain, night, highway/city scenes | RGB camera, IR camera, Velodyne LiDAR, mm-wave radar, GPS/IMU | 3D tracklets, calibration, raw ROS 2 bags | Practical radar+IR+LiDAR fusion and ROS 2 replay testing |
| Seeing Through Fog / DENSE | Fog, snow, rain, fog chamber conditions | RGB stereo, gated NIR, FIR, radar, HDL64/VLP32 LiDAR, weather station | 2D/3D boxes, weather/illumination/road-state tags | Multimodal fog/fusion validation and asymmetric failure studies |
| LIDAROC | LiDAR cover contamination including dust, water, mud, oil, and other cover states | LiDAR point clouds at 5 m, 10 m, and 20 m subsets | Clean/contaminated sample organization for robustness testing | Sensor-window contamination and dust-on-cover proxy before local airside collection |
Coverage by Airside Hazard
| Airside hazard | Strongest public proxies | What to validate |
|---|---|---|
| Falling snow | WADS, SemanticSTF, CADC | Snowflake clutter removal, snowbank segmentation, detection drop under sparse returns |
| Accumulated snow and ice | WADS, CADC/CADC+ | Drivable-area ambiguity, snowbank map drift, clear-vs-snow domain adaptation |
| Natural rain | RainSense, RADIATE, SemanticSTF | Point-density loss, camera blur, radar stability, rain-rate operating limits |
| Heavy rain artifacts | REHEARSE-3D, RainSense | Point-wise raindrop removal and radar-conditioned filtering |
| Wet-road spray | SemanticSpray++, RADIATE | Spray clutter, wet-surface reflection, radar/LiDAR disagreement |
| Fog and steam-like aerosol | Seeing Through Fog/DENSE, RADIATE, SemanticSTF, DSERT-RoLL | Visibility reduction, LiDAR wobble/clutter, gated/FIR/radar/event/thermal fallback |
| Night and low light | RADIATE, DSERT-RoLL, CMHT | RGB degradation, thermal/IR support, radar-primary fallback, over-exposure handling |
| Dust and sand | LIDAROC for dust-on-cover, plus fog/spray/snow-dust public proxies | Treat as partial sensor-contamination and particle proxies; collect airside dust/jet-blast samples |
| De-icing mist and glycol spray | SemanticSpray++, REHEARSE-3D, Seeing Through Fog/DENSE, LIDAROC | Short-duration LiDAR occlusion, radar-primary fallback, sensor-cleaning trigger thresholds, cover-contamination response |
The key gap is dust/steam/de-icing fluid realism. Existing public data provides useful particle, aerosol, wet-surface, and sensor-cover contamination proxies, but an airside validation program still needs local recordings around jet blast, de-icing trucks, apron dust, rubber residue, glycol film, wet-apron multipath, and sensor-window contamination.
Evidence boundary: public road-weather and sensor-contamination datasets are screening and training proxies. They may justify model selection, stress-test design, degraded-mode thresholds, and sensor-cleaning triggers, but they do not close an airside safety claim for de-icing mist, glycol film, jet-blast dust/FOD entrainment, steam, heat shimmer, apron floodlighting, wet-apron multipath, standing-water reflection, or retroreflector bloom without direct target-airside validation.
Recommended Validation Stack
- Point-level removal first: use WADS for falling/accumulated snow, REHEARSE-3D for rain-point removal, and SemanticSTF for all-weather semantic segmentation stress tests.
- Object-level degradation next: use CADC/CADC+ for snow-vs-clear 3D detection, RainSense for measured rain-rate curves, RADIATE for radar-first adverse-weather detection, and SemanticSpray++ for wet-road spray.
- Fusion robustness last: use Seeing Through Fog/DENSE and RADIATE to validate that radar, gated NIR, FIR, camera, and LiDAR degrade asymmetrically rather than assuming one weather scalar applies to every sensor.
- Airside transfer gate: after public-dataset screening, require a proprietary airside set with aircraft, GSE, cones, baggage carts, jet bridges, reflective markings, de-icing mist, dust, heated exhaust plumes, wet-apron multipath, and do-not-delete hazard labels before production claims. LIDAROC is a LiDAR cover-contamination proxy, not proof that free-air dust, glycol mist, or wet-apron multipath is handled.
Practical Selection Guidance
| If the model does this | Start with | Then add |
|---|---|---|
| LiDAR snow removal | WADS | SemanticSTF, CADC |
| LiDAR rain removal | REHEARSE-3D | RainSense |
| Weather-aware semantic segmentation | SemanticSTF | WADS |
| Snow domain adaptation or de-snowing | CADC+ | WADS |
| Radar fallback in adverse weather | RADIATE | RainSense, SemanticSpray++ |
| Event/thermal/radar multimodal fusion | DSERT-RoLL | CMHT, MUSES |
| ROS 2 replay for radar+IR fusion | CMHT | RADIATE, RainSense |
| Fog/steam sensor fusion | Seeing Through Fog/DENSE | RADIATE |
| Spray robustness | SemanticSpray++ | RainSense, REHEARSE-3D |
Source Notes
- WADS source records: Michigan Tech dataset page and Michigan Tech publication record
- CADC/CADC+: CADC arXiv paper, CADC+ project page, CADC+ arXiv paper
- SemanticSTF: GitHub, Hugging Face, CVPR 2023 arXiv paper
- REHEARSE-3D: arXiv paper, Sensors article
- RainSense: SAE paper record, GitHub release repository
- SemanticSpray++: project page, arXiv paper
- RADIATE: project page, dataset documentation, arXiv paper
- DSERT-RoLL: project page, arXiv paper, GitHub repository, Hugging Face dataset card
- CMHT: ScienceDirect article, PMC full text, data DOI, MacDrive data link
- Seeing Through Fog/DENSE: GitHub, Princeton dataset page, DENSE dataset page
- LIDAROC: 20m Zenodo dataset, 10m Zenodo dataset, 5m Zenodo dataset, IEEE Sensors Letters paper record