An improved lightweight high-resolution network based on multi-dimensional weighting for human pose estimation

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Human pose estimation is crucial for applications like action recognition and human-computer interaction.
• Existing lightweight networks like Lite-HRNet have limitations in feature extraction scale and information interaction.
• Improving the performance of human pose estimation remains an active research area.

Purpose of the Study:

• To propose an improved lightweight high-resolution network (MDW-HRNet) for enhanced human pose estimation.
• To address the limitations of single-scale feature extraction and insufficient information channels in prior methods.
• To achieve state-of-the-art accuracy in human pose estimation without compromising computational efficiency.

Main Methods:

• Developed MDW-HRNet incorporating global context modeling for multi-channel, multi-scale feature weighting.
• Introduced a cross-channel dynamic convolution module for inter-channel attention aggregation, replacing standard convolutions.
• Simplified network architecture to facilitate information exchange between high-resolution modules while maintaining speed and accuracy.

Main Results:

• MDW-HRNet demonstrated strong performance on the COCO and MPII human pose estimation datasets.
• The proposed method achieved higher accuracy compared to existing mainstream lightweight pose estimation networks.
• Computational complexity was not increased, indicating an efficient and effective model.

Conclusions:

• MDW-HRNet offers a significant advancement in lightweight human pose estimation.
• The integration of multi-dimensional weighting and dynamic convolutions enhances feature representation and information interaction.
• MDW-HRNet provides a competitive and efficient solution for accurate human pose estimation tasks.