VBLC: Visibility Boosting and Logit-Constraint Learning for Domain Adaptive Semantic Segmentation under Adverse Conditions

Generalizing models trained on normal visual conditions to target domains under adverse is demanding in the practical systems. One prevalent solution bridge domain gap between clear- and adverse-condition images make satisfactory prediction target. However, previous methods often reckon additional reference of same scenes taken from conditions, which are quite tough collect reality. Furthermore, most them mainly focus individual condition such as nighttime or foggy, weakening model versatility when encountering other weathers. To overcome above limitations, we propose a novel framework, Visibility Boosting Logit-Constraint learning (VBLC), tailored for superior normal-toadverse adaptation. VBLC explores potential getting rid resolving mixture simultaneously. In detail, first visibility boost module dynamically improve via certain priors image level. Then, figure out overconfident drawback conventional cross-entropy loss self-training method devise logit-constraint learning, enforces constraint logit outputs during training mitigate this pain point. best our knowledge, new perspective tackling challenging task. Extensive experiments two normal-to-adverse adaptation benchmarks, i.e., Cityscapes ACDC FoggyCityscapes + RainCityscapes, verify effectiveness VBLC, where it establishes state art. Code available at https://github.com/BIT-DA/VBLC.