AD-Assurance-Lab/winter-driving-dataset

AD Winter Driving Dataset: MI-Snow1000 Lane Detection and MCity-WSPI Physics

This is the official Hugging Face repository hosting the image binary packages and pre-trained benchmarks for the AD Winter Driving Dataset, created by the AD Assurance Lab in partnership with Western Michigan University and MCity. It has two modules:

• MI-Snow1000 is a lane detection dataset for snow-covered roads: 4,888 dashcam frames (1280x720) from over 1,000 miles of winter driving in Michigan, Colorado and Wyoming, each annotated with the left and right boundaries of the ego lane in TuSimple format. Where snow buries the paint, the boundaries follow the drivable corridor shown by tire tracks and plow lines. It includes a sequence-disjoint train, validation and test split (3,105, 887 and 337 frames), per-frame snow, glare and curvature tags, and trained SCNN, PolyLaneNet, LaneATT and CLRNet checkpoints in PyTorch and ONNX. License CC BY 4.0.
• MCity-WSPI pairs front-camera frames with vehicle dynamics, IMU, GPS and road friction telemetry recorded on snow and ice at Mcity.

For the annotations, splits, synchronized CSV metadata, loaders, and developer documentation, please visit the main GitHub repository: 👉 GitHub: AD-Assurance-Lab/dataset--winter-driving--main

---

MI-Snow1000 in brief

Frames	4,888 annotated dashcam frames, 1280x720 JPEG
Labels	Left and right boundary of the ego lane, TuSimple format (56 rows, 160 to 710; -2 where a boundary is absent)
Benchmark split	metadata/sequence_disjoint/ on GitHub: 3,105 training, 887 validation, 337 test frames, no drive shared between partitions
Condition tags	metadata/condition_tags.json on GitHub: snow severity, glare, curvature, single boundary
Models	models/ here: SCNN, PolyLaneNet, LaneATT, CLRNet at 320x180, trained on the benchmark training partition (seed 0)
License	CC BY 4.0

Frames with no lane boundaries are intentional. Where the annotators judged that no boundary could be placed with confidence, for example on a fully snow-covered road with no tracks, the frame was kept with no lanes labeled, so that a detector learns to output nothing rather than guess. These are 535 of the 4,888 frames (11%). Treat them as frames whose correct output is no lane, not as missing labels.

To ask for an image to be removed, for example one that shows your face or license plate, open an issue on the GitHub repository with the frame name.

---

Dataset Structure

This repository is structured to separate tabular metadata, pre-trained models, and large image binaries:

winter-driving-dataset/
├── metadata/                 # Annotations, splits and synchronized CSVs (on GitHub)
├── models/                   # Pre-trained lane detection checkpoints
│   ├── onnx/                 # ONNX float32/float16 format checkpoints
│   └── pytorch/              # PyTorch .pt / .pth checkpoints
└── images/
    ├── mcity_wspi/           # Individual zip packages for each dynamics sequence
    │   ├── jan27-downtown-1_images.zip
    │   ├── feb23-straight-10_images.zip
    │   └── ...
    └── reva_perception/
        └── non_labeled_set/  # Unlabeled MI-Snow1000 drive frames, one zip per session
└── raw/
    └── mcity_wspi/           # Raw recordings the synchronized CSVs were made from
        ├── csv/<day>/        # Per-topic ROS message exports, 62 runs over four days
        ├── jan20/            # Shakedown day: OxTS and vehicle CSVs, video, topic list
        ├── bags/             # The one surviving ROS 2 bag (jan20 highwaysouth1, 29 s)
        └── SHA256SUMS        # Hashes of all 712 raw files

---

How to Download and Load the Dataset

Instead of downloading this entire dataset manually, we provide dedicated Python tools inside the GitHub Repository.

1. Download Specific Run Images

To download and automatically unzip a specific run package, use the download script:

python3 tools/download_wspi.py --run jan27-downtown-1_images --dest_dir ./data

2. Download Pre-trained Lane Models

python3 tools/download_models.py --model all --dest_dir ./models

---

Aligned Physics Variables

Each run contains synchronized camera frames at 25Hz aligned with:

• Vehicle Dynamics: CAN bus steering angle, throttle/brake inputs, brake torque, and wheel speeds (fl, fr, rl, rr in rad/s).
• Inertial Response: High-grade 6-DOF IMU rotational rates and linear accelerations (including vertical ride response).
• GPS/RTK: Precise geographical coordinate trajectories.
• Mobile Weather (MARWIS): Real-time optical road condition classifications and pavement friction estimations.

---

Raw MCity-WSPI Recordings

The MCity-WSPI module (called Winter Wonderset in the SAE WCX paper describing it) was recorded on a Ford Mustang Mach-E at Mcity on 27, 28 and 29 January and 23 February 2026. The synchronized CSVs on GitHub are derived products. The recordings they were made from are under raw/mcity_wspi/:

Item	Content	Size
csv/<day>/	One CSV per ROS topic per run (oxts_imu, oxts_odometry, oxts_velocity, oxts_nav_sat_fix, vehicle_brake_info, vehicle_brake_report, vehicle_throttle_info, vehicle_throttle_report, vehicle_steering_report, vehicle_wheel_speeds, vehicle_wheel_positions) for 62 runs (jan27: 15, jan28: 13, jan29: 10, feb23: 24), plus the MARWIS log for jan27 and jan28. Each row is one message with its receive timestamp in nanoseconds, the header stamp and every field flattened	285 MB, 688 files
jan20/	The 20 January shakedown drives downtown1 and highwaysouth1: OxTS CSVs (imu, imu_bias, lever_arm, nav_sat_fix, nav_sat_ref, ncom, odometry), vehicle brake, wheel speed and position CSVs, a 29 s video, the bag metadata and topic list	22 MB
bags/	wmu-jan20-highwaysouth1_0.db3, the only ROS 2 bag that survived: 29 s, 68,319 messages, 50 topics including /arenacam6/images, /oxts/*, /vehicle/* (Dataspeed ds_dbw_msgs) and /can_bus_dbw/can_rx	3.5 GB

The bags for the January 27 to 29 and February 23 runs were lost; only these CSV exports and the camera frames survive. Raw run names carry a wmu- prefix (January) or a feb23-2026_ prefix with underscores; the public names drop them, so raw wmu-jan27-test2 is public jan27-test-2. The camera frames in images/mcity_wspi/ are the raw frames renamed frame_NNNNN.png in timestamp order; each frame's capture time is the timestamp_ns column of its row in the run's _sync.csv. See docs/raw-data.md on GitHub for the full map, and tools/wspi_sync.py to rebuild the synchronized CSVs from these exports.

Download one day's exports, or everything raw:

hf download AD-Assurance-Lab/winter-driving-dataset --repo-type dataset --include "raw/mcity_wspi/csv/jan27/*" --local-dir ./data
hf download AD-Assurance-Lab/winter-driving-dataset --repo-type dataset --include "raw/mcity_wspi/*" --local-dir ./data

---

Citation

If you use MI-Snow1000, please cite:

@misc{tye2026misnow1000,
  title={MI-Snow1000: A Comprehensive Dataset and Benchmark for Lane Detection in Adverse Winter Conditions},
  author={Tye, Eugene and Fanas Rojas, Johan and Masterson, Alex and Kadav, Parth and Hong, Guan Yue and Asher, Zachary D.},
  year={2026},
  note={Manuscript under review}
}