New study shows fog cripples drone detection, restoration not enough

A team led by Amir Pouladi at the University of Victoria has published a comprehensive study tackling a critical blind spot in drone perception: fog. Their arXiv paper (2607.05467) introduces a task-driven pipeline that links synthetic fog generation, image restoration, object detection, and tracking in one evaluation framework. Because collecting real foggy UAV footage is impractical, the team used monocular depth estimation combined with an atmospheric scattering model to inject realistic fog into clear-weather, sky-dominant outdoor images. They then compared classical, CNN-based, and transformer-based restoration methods, selecting the best performer to feed into downstream detection and tracking modules.

The results are stark: fog drives a significant increase in missed detections, degrading both detection and tracking reliability. The study tested two training regimes—clean-only and fog-inclusive—across multiple detector variants. Fog-inclusive training consistently improved robustness, making detectors far less susceptible to weather-induced failure. In contrast, test-time restoration (dehazing the image before inference) only provided a meaningful boost when the detector had never seen foggy data during training. Crucially, higher image-level restoration metrics did not correlate with better detection or tracking, underscoring the need for task-driven evaluation rather than relying solely on perceptual quality scores.