Point Evolution Hierarchy Network for Weak Single-Point Human Parsing.

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Human parsing segments the body into detailed categories using labels.
• Sparse labels like single points reduce annotation effort but challenge performance.
• Existing methods struggle with performance under limited sparse point annotations.

Purpose of the Study:

• To propose an end-to-end network for fine-grained human parsing using only single-point supervision.
• To address the challenge of achieving high parsing accuracy with minimal annotation data.

Main Methods:

• Introduced the Point Evolution Hierarchy human parsing Network (PEHNet).
• Employed a divide-and-conquer strategy, partitioning pixels into labeled, pseudo-region, and unlabeled groups.
• Developed a point dissemination module to generate pseudo-region labels from sparse points.
• Implemented a hierarchy-wise constraint leveraging point-level spatial information for structural alignment.

Main Results:

• PEHNet demonstrates superior performance compared to state-of-the-art methods.
• Achieved high accuracy on human parsing benchmarks (LIP, ATR) and semantic segmentation (Pascal VOC 2012).
• Effectively utilizes single-point supervision for fine-grained human body segmentation.

Conclusions:

• PEHNet offers an effective solution for fine-grained human parsing with minimal supervision.
• The proposed methods successfully overcome limitations of sparse point annotations.
• This approach significantly alleviates the human labeling burden while maintaining high performance.