PolarPose: Single-Stage Multi-Person Pose Estimation in Polar Coordinates

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Multi-person pose estimation is crucial for real-time applications.
• Regression-based methods face challenges in long-range 2D offset accuracy compared to heatmap methods.

Purpose of the Study:

• To address the performance gap in regression-based multi-person pose estimation.
• To develop a more accurate and efficient method for keypoint localization.

Main Methods:

• Proposed PolarPose, a method simplifying 2D offset regression to classification in polar coordinates.
• Transformed Cartesian offset regression into quantized orientation classification and 1D length estimation.
• Introduced multi-center regression to mitigate quantization errors.

Main Results:

• Achieved 70.2% AP on the COCO test-dev dataset in a single-model, single-scale setting.
• Outperformed state-of-the-art regression-based methods in accuracy.
• Demonstrated high efficiency with 71.5% AP at 21.5 FPS on COCO val2017.

Conclusions:

• PolarPose offers a reliable and accurate approach to keypoint localization in multi-person pose estimation.
• The method achieves superior performance and efficiency compared to existing regression-based techniques.
• PolarPose presents a promising direction for real-time human pose analysis.