(A) subdomain-specific knowledge distillation for unsupervised domain adaptation in adverse weather conditions

Abstract

Semantic segmentation is a branch of scene understanding that can be used to perceive urban driving scenes in applications such as self-driving cars. However, existing models are trained on clear weather images, and they suffer from performance degradation when weather, season, and brightness changes. To address this issue, unsupervised domain adaptation methods have been proposed to adapt models trained on clear weather to adverse weather without ground-truth. The previous methods either utilizes reference images captured in clear weather from the same locations as the target images to address the domain gap or employs a visual boosting module for enhancement. However, all of these methods assume that the target domain follows a single distribution. In this paper, we show that the distributions in adverse weather conditions are significantly different. Therefore, it is necessary to adapt the models separately, rather than treating them as a single domain. To this end, we propose to train two Subdomain-Specialized Teachers, one for fog,rain and snow and another for night respectively. The knowledge of the two different subdomains is distilled to the student in an online manner with symmetric cross-entropy. This allows the student to have complementary knowledge from each weather domain. Our approach demonstrates the effectiveness of subdomain-specific methods by significantly reducing the performance gap between fog, rain, and snow conditions and night conditions.