Full-BAPose: Bottom Up Framework for Full Body Pose Estimation.

Area of Science:

• Computer Vision
• Deep Learning
• Human Pose Estimation

Background:

• Accurate human pose estimation is crucial for various applications.
• Existing methods often rely on separate people detectors, limiting efficiency.
• Estimating full body keypoints, including hands, feet, and face, remains challenging.

Purpose of the Study:

• To introduce Full-BAPose, a novel bottom-up approach for comprehensive full body pose estimation.
• To develop a deep learning architecture capable of end-to-end training for pose estimation.
• To enhance accuracy and robustness in crowded and occluded scenarios.

Main Methods:

• Utilized an encoder-decoder deep learning architecture with an HRNet backbone.
• Implemented a disentangled waterfall atrous spatial pooling module for multi-scale feature representation.
• Integrated adaptive convolutions and disentangled adaptive regression for precise keypoint inference in complex scenes.

Main Results:

• Achieved state-of-the-art performance on CrowdPose (72.2% AP) and COCO-WholeBody (68.4% AP) datasets with 133 keypoints.
• Demonstrated superior efficiency and robustness across diverse conditions, including multiple people, scale variations, and occlusions.
• Successfully performed full body pose estimation without external people detectors.

Conclusions:

• Full-BAPose offers a highly effective and efficient solution for full body pose estimation.
• The novel architecture and modules enable precise keypoint detection in challenging environments.
• This method advances the state-of-the-art in bottom-up human pose estimation.