Domain Adaptation for Image Segmentation with Category-Guide Classifier

Unsupervised domain adaptation gains remarkable progress in real visual tasks by leveraging the learned knowledge from labeled source domain to solve a similar task from unlabeled target domain by adopting pre-trained large models. Fine-tuning is the most popular approach to adapt source model to the target domain when facing data inaccessibility or it is expensive to match features because of data or model size. In this situation, self-supervised learning is introduced to domain adaptation to fine-tune the source model to performed on target domain. However, many existing methods rarely take imbalanced category distributions into account. In this paper, we designs a category-guide classifier (CGC) domain adaptation model for image segmentation, which leverages regional frequency from different categories of an image to improve the recognition ability of global prediction by combining the outputs of multiple classifiers. The experiments on real-world datasets can indicate the superiority of the propose CGC model.