Attention Mechanism-Enhanced for Unsupervised Domain Adaptation Semantic Segmentation of Stitching Image

Abstract

Stitching image semantic segmentation is a critical technology for high-precision environmental perception in autonomous driving, which integrates multi-view image information to expand the perceptual field. However, the domain shift between annotated source data (e.g., nuimages datasets) and unannotated target data (e.g., stitching image scenes) severely degrades segmentation performance. To address this issue, we propose an Attention Mechanism-Enhanced Unsupervised Domain Adaptation (AME-UDA) framework. The framework consists of three core modules: a Cross-Domain Feature Alignment Module (CD-FAM) based on adversarial training, a Channel-Spatial Attention Module (CS-AM) for discriminative feature extraction, and a Consistency Regularization Module (CRM) to refine segmentation results. Extensive experiments are conducted on three public datasets (Cityscapes, GTAV, and SYNTHIA) and a self-built stitching image dataset. The results show that our method achieves a mean Intersection over Union (mIoU) of 78.3% on the Cityscapes test set when transferring from GTAV, outperforming state-of-the-art methods by 3.2%–5.7%. Ablation studies verify the effectiveness of each module, and visualization results demonstrate that the proposed framework can effectively align cross-domain features and improve segmentation accuracy for edge and small objects in stitching images. The code and dataset are available at https://github.com/xinghu/AME-UDA.