WADS: Winter Adverse Driving Dataset
Last updated: 2026-05-09
WADS is a Michigan Tech winter-driving dataset focused on severe snow, whiteout-like visibility loss, and LiDAR snow artifacts. It is one of the most useful public datasets for validating falling-snow removal because its released labeled subset includes dense point-wise LiDAR annotations with explicit active-snow and accumulated-snow classes.
Related pages: dataset index, LiDAR artifact removal techniques, radar-LiDAR fusion in adverse weather, production perception systems
What It Measures
WADS measures how severe winter weather changes autonomous-driving sensor data. The useful signal for removal validation is the separation between:
- active falling snow, which appears as transient LiDAR clutter and near-field false returns;
- accumulated snow, which changes ground/curb/roadside geometry and can break ground removal or localization;
- normal semantic road-scene classes, which must be preserved after filtering.
The Michigan Tech publication record describes three winter seasons of collection, more than 36 TB of adverse winter data, and a labeled LiDAR subset of around 1000 sequential scenes totaling more than 7 GB and 3.6 billion labeled points. The WADS sequence-26 data page describes the released sequence as SemanticKITTI-format labels with all SemanticKITTI classes plus falling and accumulated snow.
Sensors And Modalities
| Modality | Notes |
|---|---|
| High-resolution LiDAR | Primary source for dense point labels and snow-removal validation |
| Side-mounted LiDARs | Part of the sensor-pod concept used in the collection platform |
| Visible camera | Forward-facing visual context in winter weather |
| Near-infrared camera | Helps compare visible and NIR degradation in snow |
| Long-wave infrared camera | Useful for thermal contrast under low-visibility winter conditions |
| Forward radar | Weather-resilient comparison channel |
| GNSS/IMU | Pose information for sequence aggregation and mapping studies |
Labels And Tasks
| Label type | Use |
|---|---|
| Dense point-wise LiDAR semantic labels | Train/evaluate snow segmentation and snow removal |
| Active-snow class | Direct target for falling snow clutter filtering |
| Accumulated-snow class | Detect snowbanks, snow-covered drivable surfaces, and changing roadside geometry |
| SemanticKITTI-style classes | Check that removal preserves roads, vehicles, vegetation, buildings, and pedestrians |
| Sequential scans with pose | Evaluate temporal filters, mapping, and scan aggregation |
WADS supports semantic segmentation, panoptic-style scene parsing, snow/noise removal, localization and mapping under snow, and object-detection robustness after filtering.
Weather And Environment
WADS was collected in Michigan's Upper Peninsula, a snow-belt region that provides frequent severe winter events. The paper record emphasizes moderate-to-severe winter weather, heavy snowfall, occasional whiteout conditions, rural and semi-rural settings, and winter behaviors such as snow-covered sidewalks, altered pedestrian paths, and snowbanks near drivable surfaces.
This matters for airside work because airport aprons have similar open exposure: blowing snow, plowed snow ridges, low-contrast pavement markings, and large unobstructed surfaces where wind moves snow across the LiDAR path.
Benchmark Use For Perception And Removal
Use WADS as the first public validation set for snow removal:
- Train or tune filters on active-snow labels.
- Measure false-removal of non-snow classes before and after filtering.
- Evaluate semantic segmentation on raw point clouds and filtered point clouds.
- Check temporal consistency across sequential scans.
- Evaluate accumulated-snow handling separately from falling-snow removal.
For removal algorithms, the most important metric is not just how many snow points are removed. A production filter must preserve small obstacles, curbs, cones, pedestrians, and vehicle edges while removing transient snow returns.
Strengths
- Direct point-wise labels for falling snow, which many adverse-weather datasets lack.
- Explicit accumulated-snow class for snowbank and road-edge ambiguity.
- Severe winter collection rather than light cosmetic snowfall.
- Sequential scans with pose support temporal and mapping experiments.
- Multimodal collection allows radar/camera context even when LiDAR is the labeled focus.
Limitations
- The labeled public subset is much smaller than the full raw collection.
- It is snow-focused; it does not validate rain, fog, spray, dust, or steam directly.
- Object diversity is lower than urban AV datasets because severe winter driving reduces traffic and pedestrian frequency.
- Rural/semi-rural scenes transfer imperfectly to dense apron operations with aircraft, GSE, cones, dollies, and jet bridges.
- License and access terms should be checked per bundle; the sequence-26 page lists a CC BY 4.0 license.
Airside Transfer
WADS is the strongest public proxy for airside snow operations. It should be used to validate:
- falling-snow clutter removal near the sensor;
- snowbank and plowed-edge segmentation;
- preservation of small vertical obstacles after filtering;
- performance of LiDAR-only perception before radar fallback;
- snow-aware map matching where accumulated snow changes expected geometry.
It does not cover de-icing spray or jet-blast steam. For those, use WADS only as a particle-clutter pretest and then require apron recordings around de-icing areas and aircraft exhaust.