Boundary-Aware Multi-modal Mirror Segmentation
Haonan Tang, Shuhan Chen, Lifeng Zhang · 2024
Mirror segmentation is a challenging task in the field of computer vision, with the difficulty being the system's struggle to distinguish between mirror reflections and actual objects in the background of an image. Traditional methods typically rely on the contextual relationships in the image and the reflective properties of mirrors for identification. However, these approaches are susceptible to the variability of mirror placement and the randomness of objects in the scene, consequently limiting the system's recognition capabilities. Moreover, due to the high similarity between the boundaries and the background, these methods often face issues of unclear or erroneous segmentation boundaries. Research has found that mirrors exhibit more pronounced features in depth images. Based on this discovery, we propose a boundary-aware multi-modal mirror segmentation method. This method involves the use of a multi-modal feature fusion module to integrate cross-modal feature. A boundary guidance module is also employed, utilizing boundary and high-frequency information to optimize the segmentation results. Compared to traditional image segmentation methods, our multi-modal approach demonstrates significant advantages in handling mirror segmentation tasks in complex scenes, particularly in terms of boundary clarity and segmentation accuracy.