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

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