Compact interactive dual-branch network for real-time semantic segmentation

Abstract The traditional complete dual-branch structure is effective for semantic segmentation tasks. However, it redundant in some sense. Moreover, the simple additive fusion of features from two branches may not achieve satisfactory performance. To alleviate these problems, this paper we propose an efficient compact interactive network (CIDNet) real-time segmentation. Specifically, first build a by constructing detail branch and branch. Furthermore, detail-semantic module to fuse several specific stages backbone with corresponding resolution Finally, contextual attention deeply extracted predict final result. Extensive experiments on Cityscapes CamVid dataset demonstrate that proposed CIDNet trade-off between accuracy inference speed, outperforms 20 representative methods.