Human Pose Estimation Using Consistent Max Covering.

Area of Science:

• Computer Vision
• Machine Learning

Background:

• Human pose estimation is crucial for understanding human actions in images and videos.
• Existing methods often struggle with complex poses and cluttered backgrounds.

Purpose of the Study:

• To propose a novel consistent max-covering method for robust human pose estimation.
• To improve accuracy by incorporating global shape constraints and multiple cues.

Main Methods:

• Formulating pose estimation as a jigsaw puzzle problem with body part tiles.
• Maximizing coverage of foreground regions while matching local image features and satisfying body plan constraints.
• Introducing hyperedges in the part relation graph for global shape constraints and anchoring multiple body parts simultaneously.
• Solving the consistent max-covering problem efficiently using a two-stage relaxation in polynomial time.

Main Results:

• The proposed method demonstrates robustness against cluttered foregrounds by utilizing multiple cues.
• Experiments on diverse images and videos show superior performance compared to previous locally constrained methods.
• The approach effectively handles complex poses through simultaneous body part anchoring and global shape constraints.

Conclusions:

• The novel consistent max-covering method offers a more robust solution for human pose estimation.
• This approach provides significant improvements, particularly in challenging scenarios with complex poses and cluttered environments.
• The efficient linear solution enables practical application in various computer vision tasks.