Chu Jiaming ; Jin Lei ; Xing Junliang ; Zhao Jian

Abstract

Multi-human parsing is an image segmentation task necessitating bothinstance-level and fine-grained category-level information. However, priorresearch has typically processed these two types of information throughseparate branches and distinct output formats, leading to inefficient andredundant frameworks. This paper introduces UniParser, which integratesinstance-level and category-level representations in three key aspects: 1) wepropose a unified correlation representation learning approach, allowing ournetwork to learn instance and category features within the cosine space; 2) weunify the form of outputs of each modules as pixel-level segmentation resultswhile supervising instance and category features using a homogeneous labelaccompanied by an auxiliary loss; and 3) we design a joint optimizationprocedure to fuse instance and category representations. By virtual of unifyinginstance-level and category-level output, UniParser circumvents manuallydesigned post-processing techniques and surpasses state-of-the-art methods,achieving 49.3% AP on MHPv2.0 and 60.4% AP on CIHP. We will release our sourcecode, pretrained models, and online demos to facilitate future studies.