在無監督領域自適應中，使來源域上訓練的模型能夠將其所獲得的知識轉移到目標域並展現出良好性能，而無需目標域的標註。目前，既有的方法主要著重於減少不同域之間特徵、像素和預測的差異。然而，圖像內部的上下文相關性等領域內知識仍然未得到充分探索。針對語義分割任務，本文提出了一種基於極化注意力網路的無監督領域自適應模型PANDA，主要透過極化注意力捕捉圖像的局部和全局結構，同時結合通道和空間的資訊提高對特徵的感知能力，以改善模型的多層級特徵融合。該方法在兩個廣泛使用的無監督領域自適應情境中取得了進步，PANDA使得最先進的性能在GTA→Cityscapes提高了0.2 mIoU，而在SYNTHIA→Cityscapes提高了1.4 mIoU，分別達到76.1和68.7 mIoU。實驗成果表明了模型捕捉局部細節和全局特徵的有效性，為無監督領域自適應問題提供了一種新的解決方案。;In Unsupervised Domain Adaptation (UDA), the goal is to enable models trained on the source domain to transfer their acquired knowledge to the target domain and exhibit robust performance without annotations from the target domain. Presently, existing methods primarily focus on reducing differences in features, pixels, and predictions between different domains. However, domain-internal knowledge, such as contextual relevance within images, remains underexplored. For semantic segmentation, we propose a novel UDA method PANDA that primarily leverages Polarized Self-Attention (PSA) to capture both local and global structures of the images, while integrating channel and spatial information to enhance feature perception and improve multi-level feature fusion within the model. The proposed method demonstrates advancements in two widely used UDA scenarios. Specifically, PANDA improves the state-of-the-art performance by 0.2 mIoU on GTA→Cityscapes and 1.4 mIoU on SYNTHIA→Cityscapes, resulting in 76.1 mIoU and 68.7 mIoU, respectively. Experimental results illustrate the effectiveness of PANDA in capturing both local details and global features, offering a solution for UDA problems.
