Gong Ke ; Liang Xiaodan ; Li Yicheng ; Chen Yimin ; Yang Ming ; Lin Liang

Abstract

Instance-level human parsing towards real-world human analysis scenarios isstill under-explored due to the absence of sufficient data resources andtechnical difficulty in parsing multiple instances in a single pass. Severalrelated works all follow the "parsing-by-detection" pipeline that heavilyrelies on separately trained detection models to localize instances and thenperforms human parsing for each instance sequentially. Nonetheless, twodiscrepant optimization targets of detection and parsing lead to suboptimalrepresentation learning and error accumulation for final results. In this work,we make the first attempt to explore a detection-free Part Grouping Network(PGN) for efficiently parsing multiple people in an image in a single pass. OurPGN reformulates instance-level human parsing as two twinned sub-tasks that canbe jointly learned and mutually refined via a unified network: 1) semantic partsegmentation for assigning each pixel as a human part (e.g., face, arms); 2)instance-aware edge detection to group semantic parts into distinct personinstances. Thus the shared intermediate representation would be endowed withcapabilities in both characterizing fine-grained parts and inferring instancebelongings of each part. Finally, a simple instance partition process isemployed to get final results during inference. We conducted experiments onPASCAL-Person-Part dataset and our PGN outperforms all state-of-the-artmethods. Furthermore, we show its superiority on a newly collected multi-personparsing dataset (CIHP) including 38,280 diverse images, which is the largestdataset so far and can facilitate more advanced human analysis. The CIHPbenchmark and our source code are available at http://sysu-hcp.net/lip/.