Dual Context Based Network for Clothing Parsing
Shaoping Ye, Shaoyu Wang, Jingyi Fan, An Cheng Xu, Xiao Ma, Xiujin Shi · 2022
As a branch of semantic segmentation, clothing parsing is of great research value and can be applied in many practical aspects. In this paper, we propose a novel clothing parsing network, which aggregates dual context information to augment pixel representation. Specifically, we extend the FCN with Attention Class Feature Module (ACFM) and Pixel Correlation Module (PCM) to capture class-level context and global context respectively. ACFM is used to capture class-level context through a coarse-to-fine segmentation structure, which describes the average features of each class on the image. The attention mechanism is utilized in ACFM so that the network only needs to focus on the categories which present in the image, and obtains attention class feature which aggregates class-level context. Furthermore, we use PCM to refine attention class feature and obtain an augmented pixel representation that aggregates dual context information by computing pixel correlation matrix which describes the similarity of any two pixels and contains rich global context information. The experiments on the CFPD dataset show that our method achieves 93.03% of PA and 50.95% of mIoU which is a promising result compared with other state-of-the-art methods.