MirrorSAM: Large Model Boosting Mirror Segmentation

Mirror image detection aims at identifying mirror regions in a given input image. Existing work focused on integrating semantic and structural features to exploit specific relationships between mirror and non-mirror regions. However, many networks have very rough segmentation results when dealing with mirror edges. Therefore, based on this observation, we propose to integrate shallow texture features while extracting mirror context information, which facilitates accurate segmentation of mirror edges. The article proposes a SAM empowered mirror segmentation system which includes the Global Awareness Adapter (GAA), the Side-structured Convolutional Neural Networks (CNN), and the Guided Attention Mask (GAM). The idea is fine and the experimental results show the efficacy of the proposed method.