BRG: Bidirectional Regularization Guidance for Unsupervised Domain Adaptation Semantic Segmentation

Unsupervised Domain Adaptation Semantic Segmentation (UDASS) aims to harness labeled data from the source domain alongside unlabeled data from the target domain to effectively segment target domain images. While self-training has achieved tremendous success in UDASS, most existing methods heavily rely on source data and lack of sufficient learning from target data, resulting in a performance decline. To address this issue, this paper introduces a Bidirectional Regularization Guidance (BRG) method, which combines a Target Perturbation Consistency (TPC) module and a thing-class ImageNet Feature Distance Reweighting (FDR) module to provide effective regularization guidance. Specifically, the TPC module is introduced to perturb the target stream by masking at the input level and adding feature noise at the feature level. This module employs a target perturbation consistency loss to penalize inconsistencies between perturbed student predictions and teacher model predictions, thereby facilitating improved learning of contextual information by the student model from the target images. To further enhance the model’s generalization ability to the target domain, We propose an FDR module that employs a transferability map to calculate a reweighted graph, helping to effectively align source features with ImageNet features by reweighting the feature distances, especially those with lower transferability Through rigorous experimentation in standard UDASS settings—training with synthetic labeled and real unlabeled data—BRG outperforms baseline model on GTAV→Cityscapes and SYNTHIA→Cityscapes.