Real-Time Semantic Segmentation via an Efficient Multi-Column Network

Existing semantic segmentation networks based on the multi-column structure can hardly satisfy the efficiency and precision requirements simultaneously due to their shallow spatial branches. In this paper, we propose a new efficient multi-column network termed as LadderNet to address this problem. Our LadderNet includes two branches where the spatial branch generates high-resolution output feature map and the context branch encodes accurate semantic information. In particular, we first propose a channel attention fusion block and a global context module to enhance the information encoding ability of the context branch. Subsequently, a new branch fusion method, i.e., fusing some middle feature maps of the context branch into the spatial branch, is developed to improve the depth of the spatial branch. Meanwhile, we design a feature fusing module to enhance the fusion quality of these two branches, leading to a more efficient network. We compare our model with other state-of-the-arts on PASCAL VOC 2012 and Cityscapes benchmarks. Experimental results demonstrate that, compared with other state-of-the-art methods, our LadderNet can achieve average 1.25% mIoU improvement with comparable or less computation.