loading page

Advancements in Instance-Level Human Parsing: Integrating Visual Saliency with Multi-Task Learning for Complex Environments
  • Xu Yin,
  • Xinyu Wang,
  • Hongliang Tian
Xu Yin
Shandong Deep Forest Information Technology Co Ltd
Author Profile
Xinyu Wang
Northeast Electric Power University

Corresponding Author:[email protected]

Author Profile
Hongliang Tian
Northeast Electric Power University
Author Profile

Abstract

Instance-level human parsing, critical for human-centric analysis, involves labeling pixels of human body parts and associating them with specific instances. Despite progress in multi-person parsing, segmenting individuals in dense crowds remains challenging. The Visual Saliency-Based Human Parsing (ViS-HuP) algorithm addresses this by using visual saliency to enhance body pixel clarity and incorporating edge detection to refine instance boundaries within a multi-task learning framework. Tested on the Crowd Instance-level Human Parsing (CIHP) dataset[1], ViS-HuP outperforms conventional methods, showing significant accuracy improvements in crowded scenes.