Context-Aware Mixup for Domain Adaptive Semantic Segmentation

Unsupervised domain adaptation (UDA) aims to adapt a model of the labeled source an unlabeled target domain. Existing UDA-based semantic segmentation approaches always reduce shifts in pixel level, feature and output level. However, almost all them largely neglect contextual dependency, which is generally shared across different domains, leading less-desired performance. In this paper, we propose novel Context-Aware Mixup (CAMix) framework for adaptive segmentation, exploits important clue context-dependency as explicit prior knowledge fully end-to-end trainable manner enhancing adaptability toward Firstly, present mask generation strategy by leveraging accumulated spatial distributions relationships. The generated critical work will guide context-aware mixup on three levels. Besides, provided context knowledge, introduce significance-reweighted consistency loss penalize inconsistency between mixed student prediction teacher prediction, alleviates negative transfer adaptation, e.g., early performance degradation. Extensive experiments analysis demonstrate effectiveness our method against state-of-the-art widely-used UDA benchmarks.