DAFormer
DAFormer presents an advanced network architecture for improving unsupervised domain adaptation by overcoming limitations of previous models. It combines a Transformer encoder and a multi-level context-aware feature fusion decoder, enhanced by training strategies such as Rare Class Sampling and Learning Rate Warmup. These features result in a notable performance boost, with improvements of 10.8 mIoU for GTA to Cityscapes and 5.4 mIoU for Synthia to Cityscapes. Furthermore, DAFormer effectively handles generalization tasks without requiring access to the target domain, achieving state-of-the-art performance with a 6.5 mIoU enhancement. This design supports better learning of complex classes, making DAFormer a crucial tool in semantic image segmentation.
