GCAENet: global-class context with advanced edge network for single human parsing

Xiukun Zhang, Weibin Liu, Weiwei Xing, Xiang Wei

Abstract: In this paper, we propose an effective single human parsing framework, called global-class context with advanced edge network (GCAENet), which explores the human parsing task in terms of both contextual information and edge information. Since rich contextual information is crucial for pixel-level classification tasks, e.g., human parsing, some researches of human parsing have adopted atrous spatial pyramid pooling and pyramid pooling module to exploit context information. However, these methods focus only on global contextual information and not adequate attention to class contextual information. Hence, we propose an integrated approach, where a global-class context module is introduced to join the global context and the class context. Furthermore, for the problems of the boundary confusion between adjacent parts and intra-class semantic inconsistency in parsing results, we propose advanced edge module based on the edge perceiving module from Context Embedding with Edge Perceiving framework to attain more refined edge prediction map and provide guidance edge information for parsing task. In addition, we also utilize cross-entropy and Lovasz-Softmax double loss as parsing supervise. Experimental results demonstrate the proposed GCAENet achieves state-of-the-art accuracy on LIP and ATR datasets.