Skip to content

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

DatasetPrimary adverse conditionModalitiesLabelsBest validation use
WADSFalling snow, accumulated snow, whiteout-like winter drivingLiDAR, visible/NIR/LWIR cameras, radar, GNSS/IMUDense point-wise LiDAR labels with snow classesSnow removal, snow segmentation, snow-aware mapping
CADC / CADC+Canadian winter driving, paired snow and clear sequences8 cameras, VLP-32C LiDAR, GNSS/INS3D boxes; CADC+ adds paired clear/snow evaluationSnow domain shift, de-snowing, 3D detection degradation
SemanticSTFRain, snow, light fog, dense fogLiDAR, RGB imagery, calibration/weather metadataDense point-wise semantic labelsAll-weather 3D semantic segmentation and domain generalization
REHEARSE-3DEmulated heavy rainLiDAR-256, 4D radar, rain-characteristic metadataPoint-wise rain/no-rain annotationsLiDAR point-cloud de-raining and radar-conditioned removal
RainSenseNatural rainfall with measured intensityCamera, LiDAR, 4D mmWave radar, disdrometer2D/3D target boxes by 10-second caseRain-intensity response curves and modality degradation
SemanticSpray++Wet road surface and road sprayCamera, VLP32C LiDAR, Ibeo LiDARs, Aptiv radarCamera 2D boxes, LiDAR 3D boxes/semantics, radar semanticsSpray/wet-road robustness and radar-LiDAR fusion checks
RADIATERain, fog, snow, night, clear baselinesNavtech radar, stereo camera, 32-channel LiDAR, GPS/IMU2D radar-image boxes for 8 actor classesRadar-first adverse-weather detection and fusion fallback
DSERT-RoLLClear, fog, rain, snow, varied lightingStereo RGB, event, thermal, 4D radar, dual LiDAR, GPS/IMU2D/3D boxes, track IDs, odometryCross-modal event/thermal/radar/LiDAR fusion under weather and lighting shifts
CMHT Autonomous DatasetRain, night, highway/city scenesRGB camera, IR camera, Velodyne LiDAR, mm-wave radar, GPS/IMU3D tracklets, calibration, raw ROS 2 bagsPractical radar+IR+LiDAR fusion and ROS 2 replay testing
Seeing Through Fog / DENSEFog, snow, rain, fog chamber conditionsRGB stereo, gated NIR, FIR, radar, HDL64/VLP32 LiDAR, weather station2D/3D boxes, weather/illumination/road-state tagsMultimodal fog/fusion validation and asymmetric failure studies
LIDAROCLiDAR cover contamination including dust, water, mud, oil, and other cover statesLiDAR point clouds at 5 m, 10 m, and 20 m subsetsClean/contaminated sample organization for robustness testingSensor-window contamination and dust-on-cover proxy before local airside collection

Coverage by Airside Hazard

Airside hazardStrongest public proxiesWhat to validate
Falling snowWADS, SemanticSTF, CADCSnowflake clutter removal, snowbank segmentation, detection drop under sparse returns
Accumulated snow and iceWADS, CADC/CADC+Drivable-area ambiguity, snowbank map drift, clear-vs-snow domain adaptation
Natural rainRainSense, RADIATE, SemanticSTFPoint-density loss, camera blur, radar stability, rain-rate operating limits
Heavy rain artifactsREHEARSE-3D, RainSensePoint-wise raindrop removal and radar-conditioned filtering
Wet-road spraySemanticSpray++, RADIATESpray clutter, wet-surface reflection, radar/LiDAR disagreement
Fog and steam-like aerosolSeeing Through Fog/DENSE, RADIATE, SemanticSTF, DSERT-RoLLVisibility reduction, LiDAR wobble/clutter, gated/FIR/radar/event/thermal fallback
Night and low lightRADIATE, DSERT-RoLL, CMHTRGB degradation, thermal/IR support, radar-primary fallback, over-exposure handling
Dust and sandLIDAROC for dust-on-cover, plus fog/spray/snow-dust public proxiesTreat as partial sensor-contamination and particle proxies; collect airside dust/jet-blast samples
De-icing mist and glycol spraySemanticSpray++, REHEARSE-3D, Seeing Through Fog/DENSE, LIDAROCShort-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.


  1. 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.
  2. 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.
  3. 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.
  4. 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 thisStart withThen add
LiDAR snow removalWADSSemanticSTF, CADC
LiDAR rain removalREHEARSE-3DRainSense
Weather-aware semantic segmentationSemanticSTFWADS
Snow domain adaptation or de-snowingCADC+WADS
Radar fallback in adverse weatherRADIATERainSense, SemanticSpray++
Event/thermal/radar multimodal fusionDSERT-RoLLCMHT, MUSES
ROS 2 replay for radar+IR fusionCMHTRADIATE, RainSense
Fog/steam sensor fusionSeeing Through Fog/DENSERADIATE
Spray robustnessSemanticSpray++RainSense, REHEARSE-3D

Source Notes

Public research notes collected from public sources.