Class-level Aware Network for Human Parsing
Jiayi Yin, Weibin Liu, Weiwei Xing, Yuan Xiao · 2021
Having shown great performance in human parsing, convolutional neural networks(CNNs) come with much computation budget. In this paper, a novel class-level aware network(CANet), which employs an asymmetric encoder-decoder architecture, is presented to achieve reliable human parsing results in a memory friendly way. To achieve the trade-off between speed and accuracy in human parsing, we design group-split-bottleneck(GS-bt) block, where group convolution and channel split are utilized in the residual block. In decoder network, the attention pyramid pooling module(APPM) is proposed to recovering the details of human parsing. Moreover, a multi-class classification branch is developed to extract class-level information and revise human parsing results. Compared to current models, our model has less parameters and experiments demonstrate that the proposed CANet can reach state-of-the-art results on PASCAL-Person-Part dataset.