Clothing Parsing Based on Context Prior and Flow Alignment Pyramid

Clothing parsing plays an important role in the fashion industry, and it is also a significant branch of semantic segmentation. The advanced related work explores various methods to efficiently aggregate contextual information, but they do not clearly distinguish the difference of different contextual relationships. In this paper, we propose a novel network based on Context Prior and Flow Alignment Pyramid for clothing parsing. Specifically, we effectively establish an original context map to exploit intra-class and inter-class contextual information. We devise a new loss function to supervise the learning of the original context map mentioned above by using a standard context map. Considering the sizes of different classes have a large degree of variation in clothing images, we fuse semantic information from multiple levels of the network to avoid small-sized items from disappearing during network propagation. To circumvent the offset that occurs in fusion process, the Flow Alignment Module (FAM) is designed to make the output contains more spatial detail information. Experiments show that our method achieves promising performance on the CFPD dataset (51.71% of mIoU and 93.23% of PA) compared with other state-of-the-art methods.