ADCD: Redefining Autonomous Driving Adverse Weather Perception Benchmarks with 600,000 Synthetic Images

Adverse Driving Conditions Dataset (ADCD) is a dataset containing 600,000 synthetic images covering 12 adverse weather conditions. Developed by the University of Dayton research team, it uses generative AI to convert sunny driving images into scenes like rain, snow, fog, and night. This dataset provides the first large-scale standardized benchmark for evaluating autonomous driving perception systems under extreme weather. The corresponding paper was published in Machine Vision and Applications in 2025, and the dataset is available on GitHub.

Autonomous driving has advanced, but adverse weather remains a key bottleneck for safe deployment. Real-world bad weather data is scarce and hard to label (unpredictable, high cost, complex annotation). Existing datasets like KITTI and nuScenes focus on sunny weather, making it difficult to evaluate model performance in extreme environments. ADCD was created to solve this data scarcity via generative AI.

Data Sources

Integrates 5 public datasets: Udacity (24,007 images), ApolloScape (7,040), IDD (5,713), A2D2 (12,469), DDD (755).

Weather Synthesis

12 effects (8 single, 4 mixed):

• Single: Crack, Flare, Haze, Raindrop, Snow (InstructPix2Pix), Sunset (InstructPix2Pix), Night (CycleGAN-Turbo), Rain (CycleGAN-Turbo)
• Mixed: Haze+Raindrop, Haze+Night, Rain+Raindrop, Crack+Flare Each effect generates 50k images (total 600k).

Label Consistency

Reuses original sunny image labels (YOLO format, 6 categories: Car, Truck, Bicycle, Motorcycle, Person, Traffic light) since weather conversion doesn’t change object positions/categories.

Models Tested

YOLOv5-YOLOv11, DETR, R-CNN, Faster R-CNN, RetinaNet, SSD (pre-trained weights, no fine-tuning).

Metrics

COCO protocol: IoU ≥0.5, confidence ≥0.25, AP/mAP (full-point interpolation).

Key Findings

1. All models show significant performance drops (some mAP >50%).
2. Night/fog cause worse drops than light rain.
3. YOLO balances speed/accuracy; DETR has better generalization in some conditions.
4. Small/distant objects (e.g., traffic lights) are harder to detect.

Academic

• Paper: Towards safer roads... (Machine Vision and Applications 2025, Vol.36 No.4).
• Funded by NSF (Grant 2025234).
• Bibtex citation available.

Practical

• Researchers: Standard benchmark, data augmentation, domain adaptation.
• Industry: Safety testing, defect discovery, compliance.
• Society: Improve weather-related driving safety, reduce accidents.

Limitations

1. Synthetic vs real weather domain gap.
2. Limited dynamic weather simulation.
3. No lidar/radar data.

Future

• Add video sequences.
• Expand to multi-sensor data.
• Better physical simulation.
• Build end-to-end perception solutions.