Canadian Adverse Driving Conditions Dataset

Matthew Pitropov, Danson Evan Garcia, Jason Rebello, Michael Smart, Carlos Wang, Krzysztof Czarnecki, Steven Lake Waslander

Abstract: The Canadian Adverse Driving Conditions (CADC) dataset was collected with the Autonomoose autonomous vehicle platform, based on a modified Lincoln MKZ. The dataset, collected during winter within the Region of Waterloo, Canada, is the first autonomous vehicle dataset that focuses on adverse driving conditions specifically. It contains 7,000 frames collected through a variety of winter weather conditions of annotated data from 8 cameras (Ximea MQ013CG-E2), Lidar (VLP-32C) and a GNSS+INS system (Novatel OEM638). The sensors are time synchronized and calibrated with the intrinsic and extrinsic calibrations included in the dataset. Lidar frame annotations that represent ground truth for 3D object detection and tracking have been provided by Scale AI.

Introduction: The Challenge of Adverse Weather

Most autonomous vehicle (AV) datasets—like the seminal KITTI or Waymo Open Dataset—were collected in regions with temperate climates (e.g., California or Germany) under clear skies. However, for AVs to be truly viable, they must navigate adverse weather. This has led to a "fair-weather bias" in self-driving research that the Canadian Adverse Driving Conditions (CADC) dataset aims to rectify.

The Autonomoose platform operating in snowy conditions.

Snowfall presents a unique "double whammy" for perception systems. First, there is sensor degradation: falling snow creates "noise" or backscatter in LiDAR returns and obscures camera lenses. Second, there is environmental change: snow accumulation changes the appearance of the road, hides lane markings, and alters the geometry of the world, making localization and object detection significantly harder. The authors address the lack of high-quality, annotated data for these specific conditions, which has historically prevented the research community from developing robust solutions for winter driving.

The Autonomoose Platform and Methodology

The researchers utilized the "Autonomoose" platform, a modified Lincoln MKZ equipped with a sophisticated sensor suite designed for 360° coverage and high precision.

Sensor layout and coordinate frames on the Autonomoose vehicle.

Sensor Suite

  • Vision: 8 Ximea cameras providing a full 360° field of view.
  • LiDAR: A Velodyne VLP-32C (32-beam) sensor.
  • Localization: A high-precision Novatel GNSS+INS system and Xsens IMUs.

Technical Innovations

To ensure data consistency, all 8 cameras are hardware-triggered at 10Hz via a custom-designed signal distribution board. This ensures images are captured at the exact moment the LiDAR sensor passes its "phase lock" angle. Furthermore, the team implemented a post-processing step for LiDAR motion correction. Because a LiDAR scan takes time to rotate, the vehicle's motion can distort the point cloud; the team used GPS/IMU data to transform points back to a single timestamp, making the world appear stationary.

To categorize the severity of the weather, the researchers used a "Dynamic Radius Outlier Removal" (DROR) algorithm to filter snow noise from the LiDAR. By counting the number of points removed, they quantitatively categorized drives into four levels: Light, Medium, Heavy, and Extreme snowfall.

Dataset Composition and Results

The CADC dataset introduces 75 driving sequences covering over 20km of driving in the Waterloo region of Ontario, Canada.

  • Annotations: 7,000 frames feature 3D cuboid annotations provided by Scale AI. Unlike the Oxford RoboCar dataset (which has only one snowy drive and no 3D labels), CADC provides full 3D labels specifically for snowy conditions.
  • Scale: The dataset includes 281,941 total instances of "Car," 62,851 "Pedestrians," and 20,411 "Trucks."
  • Diversity: The data covers a wide range of "states," including parked, stopped, and moving vehicles, which is crucial for training tracking algorithms.
  • Accessibility: The authors released a Python-based development kit (cadc_devkit) that allows researchers to easily project 3D annotations onto 2D images and visualize vehicle paths.

Key Insights and Innovations

  • The "Snowplow" Attribute: Recognizing that winter driving involves unique actors, the authors included specific attributes for vehicles like "Snowplow Truck," which have different shapes and behaviors than standard trucks.
  • LiDAR as a Weather Gauge: The researchers found they could use the LiDAR sensor itself to measure snowfall intensity. By identifying "noise" points that don't belong to solid objects, they created a repeatable metric for weather severity.
  • Trinocular Configuration: The three forward-facing cameras are aligned to allow for easy image rectification, supporting researchers working on stereo vision depth estimation.

Limitations and Future Work

While the dataset is a significant step forward, the authors acknowledge several limitations. The data was collected over only three days, representing a limited temporal window. There is also a class imbalance; the "Car" label currently covers everything from small sedans to large SUVs and vans. The authors plan to release 2D annotations to help subdivide this class.

Future updates are expected to include more detailed labels for truncation and occlusion, which are particularly prevalent when snow piles up on the side of the road. Additionally, while the dataset is a great resource, the paper does not yet establish a formal leaderboard or baseline performance metrics for standard algorithms (like PointPillars or YOLO) on this specific adverse-weather data.

Summary updated

Abstract

The Canadian Adverse Driving Conditions (CADC) dataset was collected with the Autonomoose autonomous vehicle platform, based on a modified Lincoln MKZ. The dataset, collected during winter within the Region of Waterloo, Canada, is the first autonomous vehicle dataset that focuses on adverse driving conditions specifically. It contains 7,000 frames collected through a variety of winter weather conditions of annotated data from 8 cameras (Ximea MQ013CG-E2), Lidar (VLP-32C) and a GNSS+INS system (Novatel OEM638). The sensors are time synchronized and calibrated with the intrinsic and extrinsic calibrations included in the dataset. Lidar frame annotations that represent ground truth for 3D object detection and tracking have been provided by Scale AI.
BibTeX
@article{pitropov2021canadian,
  author = {Matthew Pitropov and Danson Evan Garcia and Jason Rebello and Michael Smart and Carlos Wang and Krzysztof Czarnecki and Steven Lake Waslander},
  title = {{Canadian Adverse Driving Conditions Dataset}},
  journal = {Int. J. Robotics Res.},
  year = {2021},
  url = {https://openreview.net/forum?id=Ap0URx4XnmBN},
}
Figure