FogAdapt: Self-supervised domain adaptation for semantic segmentation of foggy images

This paper presents FogAdapt, a novel approach for domain adaptation of semantic segmentation dense foggy scenes. Although significant research has been directed to reduce the shift in segmentation, scenes with adverse weather conditions remains an open question. Large variations visibility scene due conditions, such as fog, smog, and haze, exacerbate shift, thus making unsupervised scenarios challenging. We propose self-entropy multi-scale information augmented self-supervised method (FogAdapt) minimize segmentation. Supported by empirical evidence that increase fog density results high probabilities, we introduce based loss function guide method. Furthermore, inferences obtained at different image scales are combined weighted uncertainty generate scale-invariant pseudo-labels target domain. These robust scale variations. evaluate proposed model on real clear-weather synthetic non-foggy images scenarios. Our experiments demonstrate FogAdapt significantly outperforms current state-of-the-art images. Specifically, considering standard settings compared (SOTA) methods, gains 3.8% Foggy Zurich, 6.0% Driving-dense, 3.6% Driving mIoU when adapted from Cityscapes Zurich.