PoseNet++: A multi-scale and optimized feature extraction network for high-precision human pose estimation.

Area of Science:

• Computer Vision
• Deep Learning
• Artificial Intelligence

Background:

• Deep learning has advanced human pose estimation (HPE), but challenges remain with occlusions, complex poses, and multi-person interactions.
• Existing models struggle to accurately identify key body joints in cluttered or partially obscured scenes.

Purpose of the Study:

• To develop a novel human pose estimation approach, PoseNet++, that overcomes limitations in accuracy and efficiency.
• To enhance the model's ability to handle occlusions, complex poses, and multi-person scenarios through innovative architectural modules.

Main Methods:

• Introduced PoseNet++, a 3-stacked hourglass architecture featuring three key innovations: multi-scale spatial pyramid attention hourglass module (MSPAHM), coordinate-channel prior convolutional attention (C-CPCA), and PinSK Bottleneck Residual Module (PBRM).
• MSPAHM improves long-range channel dependencies for better joint relationship capture under occlusion.
• C-CPCA prioritizes keypoint regions and reduces confusion in multi-person settings, while PBRM optimizes feature extraction for complex poses.

Main Results:

• PoseNet++ achieved a 3.3% relative improvement in PCKh score on the MPII validation set compared to the baseline.
• The model significantly reduced parameters by 60.3% and floating-point operations by 53.1%.
• Achieved state-of-the-art performance on MPII, LSP, COCO, and CrowdPose datasets with lower model complexity.

Conclusions:

• PoseNet++ offers a highly accurate and efficient solution for human pose estimation.
• The proposed architectural innovations effectively address key challenges in HPE, including occlusion and complex scenarios.
• PoseNet++ represents a significant advancement in deep learning-based human pose estimation, balancing performance and computational cost.