Spatial-Semantic Fusion Network for Semantic Segmentation in Real-time
Fang Yu, Xuehe Zhang, He Zhang, Liu Gangfeng, Li Changle, Zhao Jie · 2019
In this paper, a real-time semantics segmentation algorithm for mobile terminals such as robots is proposed, which captures large-scale and small-scale features respectively by using convolution and dilation convolution to obtain spatial information. It can get semantic information by using decode-encode structure. It contains two branches, one extracts semantic information, the other extracts spatial information, and finally fuses the two branches together through self-attention method. Our algorithm achieves 71.3% mIoU on the Pascal VOC 2012 data sets, and it takes only 16ms to infer a 480*640 picture in GTX1080Ti. It can be applied to most mobile robot platforms.