FogDrive dataset benchmarks self-driving perception under graded fog

FogDrive, created by Vansh Panwar, is a rigorously calibrated multi-modal synthetic dataset for autonomous driving perception under adverse weather. Built with the CARLA simulator, it comprises 660 scenes (133k fully annotated frames, 50:50 day/night) from four synchronized cameras (RGB, depth, semantic segmentation), a LiDAR and semantic-LiDAR pair, and front radar. Fog is physically modeled using the Koschmieder model on camera channels and the Beer-Lambert law on LiDAR channels, at three calibrated visibility densities: 160m, 100m, and 50m. Each scene is provided in four matched variants (clean plus three graded fog levels) with cross-calibrated 2D and 3D bounding boxes. A quality audit over 8k images validates annotations at 95.1% precision and >99% recall for vehicles within 40m.

The paper establishes baselines using state-of-the-art architectures: TransFusion, BEVFusion, and YOLOv8-m across 3D multi-modal fusion and 2D image restoration paradigms. Key findings include that mixing multi-density fog during training tightens 3D bounding-box geometry without added data-scaling cost, while in 2D pipelines, image-quality metrics (PSNR, SSIM) are poor predictors of downstream detection performance. This challenges common assumptions in restoration-based approaches. FogDrive will be fully open-sourced along with its data-generation framework, providing the community with a standard benchmark for robust multi-modal perception under graded fog.