Multi-Person Pose Estimation via Multi-Layer Fractal Network and Joints Kinship Pattern

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Accurate multi-person pose estimation is crucial for various applications.
• Existing methods face challenges in accurately localizing and associating human joints in complex scenes.

Purpose of the Study:

• To develop an effective method for boosting the accuracy of multi-person pose estimation.
• To improve the regression of multi-person joints location heatmap and optimize keypoint detection.

Main Methods:

• A three-layer fractal network was constructed for multi-person joints location heatmap regression.
• A hierarchical bi-directional inference algorithm was proposed to calculate joint relatedness (Kinship).
• A joints kinship pattern matching mechanism combined Kinship with spatial constraints for joint pair matching.

Main Results:

• The proposed method achieved approximately 1% and 0.6% increase in mean Average Precision (mAP) on the MPII multi-person subset and MSCOCO 2016 keypoints challenge, respectively.
• The approach demonstrated superior performance compared to state-of-the-art schemes.

Conclusions:

• The developed method effectively enhances multi-person pose estimation accuracy.
• The fractal network and kinship pattern matching mechanism offer a promising direction for human pose analysis.