Detect Me if You Can: An Augmentation Framework for Robust Detection in Bad Visibility Conditions

Accurate detection of vehicles, pedestrians, and obstacles is essential for the decision-making capabilities of autonomous vehicles. While current methods achieve high detection accuracy in favorable conditions, their performance significantly degrades in poor visibility scenarios such as fog, rain, or snow. This degradation is partly due to the underrepresentation of these "edge-case" scenarios in existing datasets. This work introduces the AdverseAugment framework, an augmentation strategy to fine-tune detection models for increased robustness in bad visibility conditions. Our approach utilizes deep content-aware augmentations, weather overlays, and standard augmentations to cover a wide range of conditions. The AdverseAugment framework is tested for multiple detection models on a benchmark dataset with fog, rain, and snow conditions. The results indicate that our augmentation policy can enhance object detection performance, improving Average Precision by up to 5% compared to clear weather baselines and other augmentation strategies. Extensive experiments and evaluations are conducted to prove the effectiveness of the proposed framework. This work serves as a guideline for effective augmentation to increase detection robustness in bad visibility.