Fog-free training for foggy scene understanding

초록

It is essential to have semantic segmentation models that work effectively in foggy driving scenarios. This is because fog severely affects the safety of autonomous driving systems, posing significant visibility challenges and increasing the risk of accidents. Traditional methods often use complex and high-cost foggy datasets for training, which can be expensive and difficult to scale. To tackle this issue, we propose a novel fog-free method called ShiftMatch. Our method does not rely on foggy images for training. Instead, it creates virtual domain shifted images by applying simple data augmentation methods and normalization techniques. During training, we ensure that the segmentation results from both the original and the domain-shifted images are consistent. This approach prevents the model from overfitting to specific domain features, enabling it to learn domain invariant features effectively. Despite its cost-efficiency, ShiftMatch achieves state-of-the-art performance on three real foggy scene segmentation datasets. Additionally, it demonstrates superior performance in nighttime, rain, and snow driving scenarios.

키워드