Leveraging Entropy-Driven Attention to Adapt Semantic Segmentation of Aerial Images for Autonomous Driving

Semantic segmentation of aerial images is vitally important to many aspects of autonomous driving. However, when applied to the segmentation of large-scale aerial imagery acquired from diverse geographic regions, pre-trained deep learning models often fail to produce consistently accurate predictions. To address this challenge, this paper introduces a novel unsupervised domain adaptation (UDA) method leveraging entropy-driven attention for the semantic segmentation of aerial images. The entropy-driven attention strategy contains a two-stage adversarial learning process, which utilizes the entropy map to explicitly measure the data distribution distance between two domains and then guides the model to focus on the poorly aligned features. We conducted extensive experiments on the LoveDA dataset to validate the effectiveness of our proposed UDA method. The quantitative results indicate that our approach surpasses six other state-of-the-art UDA methods, achieving superior performance.