Dual-level Consistency Learning for Unsupervised Domain Adaptive Night-time Semantic Segmentation

Feifei Ding, Jianjun Li, Wanyong Tian

Abstract: In autonomous driving, it is crucial to train a segmentation model that can perform well in poor illumination. Due to the lack of pixel-level annotated nighttime images and a large domain discrepancy between day-night image pairs, it could be tough to achieve encouraging performance for the nighttime model. In this paper, we propose a novel Dual-level Consistency Learning (DCL) method which fully utilizes these day-night correspondences across two domains. Specifically, our DCL regularizes the style and semantic difference consistency of day-night images at different feature levels. We conduct style difference consistency at the shallow feature level and facilitate contrastive semantic difference regression consistency at the deep feature level. These designs enable the proposed DCL to acquire pairwise pixel-level pseudo supervision, reducing the dependence on annotated nighttime images. Empirically, we conduct extensive experiments on standard benchmarks such as Dark Zurich and ACDC. The encouraging results show the effectiveness of our DCL method.