Source-Free Domain Adaptation for Semantic Segmentation with Region-Aware Contrastive Alignment

Image semantic segmentation often requires a substantial amount of annotated data. Traditional unsupervised domain adaptation (UDA) methods tackle the challenge of unannotated data in the target domain by transferring knowledge from a source domain with annotations. However, most existing UDA techniques depend on accessing source domain data during adaptation, which poses a risk of data leakage. To address this, we adopt a source-free domain adaptation (SFDA) approach. In this method, adaptation is carried out using only models pre-trained on the source domain and unlabeled target domain data. This approach extends UDA techniques that leverage self-training and incorporate contrastive learning between image pairs. First, target domain data is augmented to generate challenging variants, facilitating region-aware contrastive learning between original and augmented image pairs. Second, a self-training regularization strategy transfers pre-trained knowledge from the source domain to the target domain. Finally, to reduce reliance on source domain data during adaptation, the source domain's feature distribution is approximated by sharing a domain-independent embedding space. We conducted extensive experiments on benchmark datasets and analyzed the results to confirm the effectiveness of our method.