Differentiable Multi-Granularity Human Parsing.

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Instance-aware human body part parsing is a complex task.
• Existing methods often require post-processing or heuristic algorithms.

Purpose of the Study:

• To develop a novel bottom-up framework for instance-aware human body part parsing.
• To achieve joint and end-to-end learning of human semantic segmentation and multi-person pose estimation.

Main Methods:

• A dense-to-sparse projection field is learned to associate dense semantics with sparse keypoints.
• Pixel grouping is framed as a multi-person joint assembling task solved via maximum-weight bipartite matching.
• Two novel differentiable algorithms (projected gradient descent, unbalanced optimal transport) are developed for joint association.

Main Results:

• The proposed framework achieves compact, efficient, and powerful human parsing.
• It outperforms existing human parsers on MHP-v2, DensePose-COCO, and PASCAL-Person-Part datasets.
• The method demonstrates significantly more efficient inference compared to prior approaches.

Conclusions:

• The end-to-end trainable method effectively handles multi-granularity human representation learning.
• This approach distinguishes itself from current parsers by avoiding complex post-processing.
• The framework offers a more efficient and accurate solution for instance-aware human parsing.