Domain Adaptive Semantic Segmentation

Domain adaptive semantic segmentation aims to train models that accurately segment images from a target domain using only labeled data from a different, source domain. Current research focuses on improving the robustness and efficiency of these models, often employing self-training, contrastive learning, and techniques like mixup and prototype-based methods to bridge the domain gap. These advancements leverage various architectures, including transformers and convolutional neural networks, to address challenges such as noisy pseudo-labels, class imbalance, and inconsistent taxonomies between domains. The resulting improvements have significant implications for applications like autonomous driving and remote sensing, where labeled data for every scenario is scarce or expensive to obtain.