Evaluation and analysis of accurate human body image parsing methods in semantic boundary regions

GONG Qiqi; ZHAO Yao

doi:10.11860/j.issn.1673-0291.20230115

Abstract：Accurate human body image parsing is a fine-grained image segmentation task aimed atpixel-level classification of human body images. Due to its wide range of applications,accurate human body image parsing has garnered significant attention from researchers in the past decade,leading to rapid advancements in related technologies. This paper focuses on evaluating the predictive performance of existing human body image parsing models in semantic boundary regions. Firstly,we provide a comprehensive overview of existing human parsing datasets,comparing their scale and annotation categories. Secondly,we classify existing methods based on their theoretical foundations,encompassing general semantic segmentation approaches,methods guided by auxiliary information,techniques for enhancing high-resolution features,and label denoising methodologies. Thirdly,addressing the limitations of current evaluation metrics in assessing predictive performance in boundary regions,we introduce a new evaluation metric,namely mean Boundary Intersection over Union (mBIoU),and employ it to evaluate existing models,quantitatively comparing their performance dif-ferences. Finally,we offer insights into future research directions. Our findings indicate that mBIoU,compared to existing metrics,better distinguishes models' predictive performance in semantic bound-ary regions. The proposed mBIoU effectively evaluates the capability of accurate human body image parsing models to address the specific challenges inherent in human parsing tasks,thereby facilitating algorithm development and performance evaluation.