Internal-external boundary attention fusion for glass surface segmentation

Detecting transparent objects and mirrors in an image is a highly challenging task because their glass surfaces contain the visual appearance of other transmitted or reflected objects. In this work, we highlight the importance of exploiting the transition region between glass and non-glass surfaces, which is called boundary. In other words, we exploit the visual appearances of the transition region to facilitate the task of glass surface segmentation. Specifically, we divide the transition region into internal boundary and external boundary and propose an internal-external boundary attention module (IEBAM) to separately learn their visual characteristics. The processed features are then integrated dynamically via a fusion boundary attention module (FBAM). Extensible results on six benchmark datasets show that our proposed IEBAM and FBAM are effective in improving the performance of glass surface segmentation.