BASNet: Improving semantic segmentation via boundary-assistant symmetrical network

Recently, boundary information has gained more attention in improving the performance of semantic segmentation. This paper presents a novel symmetrical network, called BASNet, which contains four components: the pre-trained ResNet-101 backbone, semantic segmentation branch (SSB), boundary detection branch (BDB), and aggregation module (AM). More specifically, our BDB only focuses on processing boundary-related information using a series of spatial attention blocks (SABs). On the other hand, a set of global attention blocks (GABS) are used in SSB to further capture more accurate object boundary information and semantic information. Finally, the outputs of SSB and BDB are fed into AM, which merges the features from SSB and BDB to boost performance. The exhaustive experimental results show that our method not only predicts the boundaries of objects more accurately, but also improves the performance of semantic segmentation.