1(School of Information Engineering,Inner Mongolia University of Science & Technology,Baotou 014010,China)2(Institute of Information Engineering,Inner Mongolia University of Technology,Hohhot 010051,China)3(Institute of Information Science and Engineering,Yanshan University,Qinhuangdao 066004,China)
Abstract:Instance segmentation is a challenging task that requires both instance-level and pixel-level prediction,and is widely used in autonomous driving,video analysis,and scene understanding.In recent years,instance segmentation methods based on deep learning have developed rapidly,such as the powerful instance segmentation benchmark Mask R-CNN,which is expanded by the two-stage detector Faster R-CNN,focusing on network accuracy rather than speed,and once became the benchmark for instance segmentation.The instance segmentation algorithm YOLACT extended by the single-stage detector with high-speed detection fills the gap of the real-time instance segmentation model,and has high research and application value.This paper first classifies the instance segmentation algorithm,then analyzes some representative algorithms and their improved algorithms in depth,and explains the advantages and disadvantages of the related algorithms.Finally,the future development of the instance segmentation method is prospected.