A Novel Human Parsing Method Enhanced by Cross-Refinement
Proceedings of 2020 the 10th International Workshop on Computer Science and Engineering · 2020
The existing human parsing methods lead to some semantic errors and incomplete semantic recognition. To solve this two problems, we design a novel approach, called Cross Refinement Network (CRN). It is built based on the Part Grouping Network (PGN). The CRN remains the same as the PGN in the backbone network and the final refinement block, and uses the ResNet-101 and the Refinement branch separately, just like the PGN. Compared to the PGN, CRN innovatively utilizes the atrous convolution to map the last three feature modules of the backbone network into the feature space, and proposes the cross-refinement module, which consists of the semantic segmentation refinement branch and the edge detection refinement branch, to cross-merge the output feature maps. Experiments are performed based on Crowd Instance-level Human Parsing(CIHP), Pascal-Person-Part (PPP), and LIP datasets separately. The results show that our method outperforms the typical methods on the task of human parsing.