SePiCo: Semantic-Guided Pixel Contrast for Domain Adaptive Semantic Segmentation

Domain adaptive semantic segmentation attempts to make satisfactory dense predictions on an unlabeled target domain by utilizing the supervised model trained a labeled source domain. One popular solution is self-training, which retrains with pseudo labels instances. Plenty of approaches tend alleviate noisy labels, however, they ignore intrinsic connection training data, i.e., intra-class compactness and inter-class dispersion between pixel representations across within domains. In consequence, struggle handle cross-domain variations fail build well-structured embedding space, leading less discrimination poor generalization. this work, we propose Se xmlns:xlink="http://www.w3.org/1999/xlink">mantic-Guided xmlns:xlink="http://www.w3.org/1999/xlink">Pi xmlns:xlink="http://www.w3.org/1999/xlink">xel xmlns:xlink="http://www.w3.org/1999/xlink">Co xmlns:xlink="http://www.w3.org/1999/xlink">ntrast (SePiCo) , novel one-stage adaptation framework that highlights concepts individual pixels promote learning class-discriminative class-balanced domains, eventually boosting performance self-training methods. Specifically, explore proper concepts, first investigate xmlns:xlink="http://www.w3.org/1999/xlink">centroid-aware contrast employs category centroids entire or single image guide discriminative features. Considering possible lack diversity in then blaze trail distributional perspective involve sufficient quantity instances, namely xmlns:xlink="http://www.w3.org/1999/xlink">distribution-aware approximate true distribution each from statistics data. Moreover, such optimization objective can derive closed-form upper bound implicitly involving infinite number (dis)similar pairs, making it computationally efficient. Extensive experiments show SePiCo not only helps stabilize but also yields representations, significant progress both synthetic-to-real daytime-to-nighttime scenarios. The code models are available at https://github.com/BIT-DA/SePiCo