Feature Fusion and Graph Convolution Correction Network for Human Parsing

Abstract

AbstractIn human parsing research, using a multitasking approach i.e., parsing task assisted by other tasks is an effective method to improve parsing accuracy. However, such methods do not enable the effective fusion of different types of features generated by the two tasks, so a feature fusion and graph convolution correction network (FGNet) for human parsing is proposed in the paper. The proposed network follows a multitasking architecture while a non-local feature fusion module is designed inspired by the idea of attention mechanism to keep the consistency of edge features and parsing features. Moreover, by taking the pixels in images as graph nodes, the graph convolution correction module is proposed to learn the relationship between nodes to further correct and refine the wrong annotations in the parsing results obtained from the non-local feature fusion module. The final experimental results show that the method we proposed performs well in both quantitative and qualitative aspects.KeywordsDeep learningHuman parsingAttention mechanismGraph convolution