Attention-guided trilateral network for real-time semantic segmentation

Siming Jia, Yongsheng Dong, Lintao Zheng, Chongchong Mao, Lin Wang, Guoyong Wang

Abstract: The dual-branch structure is a common network structure in real-time semantic segmentation field. However, the gap between the information of two branch feature maps is generally large. This may cause problems of feature confusion and information loss during fusion, and thus lead to unsatisfactory results. This work aims to minimize this gap and enhance fusion efficacy. To this end, we propose an Attention-Guided Trilateral Network (AGTNet) for real-time semantic segmentation. Specifically, AGTNet is a triple-branch structure consisting of a semantic branch, a detail branch and an alignment branch. We propose a Local Alignment Module (LAM) to construct the alignment branch. It is used to capture information from both the semantic branch and the detail branch, thus reducing the information gap in the final fusion. We further propose a Simple Parallel Dilation Residual Module (SPDRM) to enable detail branch to capture detail information at different scales. Finally, we propose a Trilateral Attention Fusion Module (TAFM) to fuse the feature information of the three branches. Extensive experiments show that our proposed AGTNet achieves good results in terms of segmentation accuracy and inference speed.