DAMSeg: degradation-aware modulation network for robust semantic segmentation of road scenes under adverse conditions

Abstract

Semantic segmentation models for autonomous driving suffer significant performance degradation under adverse weather and lighting conditions such as fog, rain, snow, and nighttime illumination. Existing approaches include restoration before segmentation and robustness-oriented adaptation or generalization; comparatively few use an image-derived degradation estimate as an explicit conditioning input to the segmentation network. We propose DAMSeg, a degradation-aware modulation framework that treats degradation characteristics as conditioning signals within a unified end-to-end model. DAMSeg comprises an implicit degradation encoder that extracts a compact conditioning vector directly from the input image, a conditional feature modulation module that applies stage-specific channel transformations with spatially adaptive fusion, and degradation contrastive regularization (DCR), which structures the projection space used during training. Scene-level condition annotations are used to construct DCR pairs when they are available; the network itself receives only the RGB image and requires no condition label at inference. Experiments on ACDC, Foggy Cityscapes, and BDD100K show that DAMSeg achieves 75.9% mIoU on the ACDC test set with a MiT-B2 backbone, improving the SegFormer-B2 baseline by 5.7 percentage points with gains across all four adverse conditions. On the evaluated dataset transfers, the performance benefit persists without target-domain fine-tuning.