Multi-class human body parsing with edge-enhancement network

Abstract

Single human parsing aims at partitioning an image into semantically consistent regions belonging to the body parts or clothing items, which has gained remarkable improvement owing to a wide range of proposed methodologies. From the perspective of the loss design, besides the parsing loss of the final output, most existing studies target on exploiting multiple other losses to enhance parsing results, which is hard to make the model reach balanced condition by adjusting their ratios and may weaken the potential of some losses. In this work, we propose an edge enhancement module to emphasize the potential of edge loss and boundary information. At the same time, local and global information will be explored for complex multi-class human body parsing problem by densely connected atrous spatial pyramid pooling. This scheme results in a simple yet powerful Edge-Enhancement Network (EEN). Extensive experiments demonstrate that EEN achieves 56.55% mIoU on LIP dataset and 62.60% mIoU on CIHP dataset, which outperform the state-of-the-arts by 3.45% and 4.02%, respectively. The code of EEN is available at https://github.com/huangxi6/EEN.