AIParsing: Anchor-Free Instance-Level Human Parsing.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• State-of-the-art instance-level human parsing models often rely on two-stage anchor-based detectors.
• These methods face challenges with heuristic anchor box design and lack of pixel-level analysis.
• Existing approaches struggle with distinguishing adjacent or overlapping human parts within a single instance.

Purpose of the Study:

• To develop an instance-level human parsing network that is anchor-free and operates at a pixel level.
• To address the limitations of heuristic anchor box design and improve pixel-level analysis in human parsing.
• To enhance the ability to distinguish numerous human parts, including overlapping ones, within a single instance.

Main Methods:

• Designed a novel instance-level human parsing network featuring an anchor-free detection head for bounding box predictions.
• Incorporated an edge-guided parsing head that utilizes part-aware boundary clues for precise human segmentation.
• Implemented a refinement head that integrates box-level scores and part-level parsing quality for improved results.

Main Results:

• The proposed anchor-free detector head demonstrates pixel-like merits and reduces hyper-parameter sensitivity.
• The edge-guided parsing head effectively distinguishes up to 58 human parts per instance, even with overlaps.
• Experiments on CIHP, LV-MHP-v2.0, and VIP datasets show superior global and instance-level performance compared to alternatives.

Conclusions:

• The developed anchor-free, pixel-level human parsing network significantly advances the state-of-the-art.
• The method offers improved accuracy and robustness in segmenting complex human instances and parts.
• This approach provides a more effective solution for instance-level human parsing tasks.