MEMe: A Mutually Enhanced Modeling Method for Efficient and Effective Human Pose Estimation.

Area of Science:

• Computer Vision
• Machine Learning
• Deep Learning

Background:

• Traditional human pose estimation models achieve high accuracy but suffer from large scale and deployment difficulties.
• Lightweight models offer better deployment but exhibit a significant performance gap compared to heavier counterparts.
• There is a need for methods that enhance lightweight models without compromising efficiency.

Purpose of the Study:

• To propose a novel Mutually Enhanced Modeling method (MEMe) for improving lightweight human pose estimation.
• To reconstruct a lightweight baseline model (EffBase) into an efficient and effective pose (EEffPose) network.
• To achieve high accuracy in human pose estimation using a low-complexity, parameter-efficient approach.

Main Methods:

• The proposed MEMe reconstructs the EfficientDet-based EffBase model into the EEffPose net.
• The EEffPose net incorporates three mutually enhanced modules: Enhanced EffNet (EEffNet) backbone, Total Fusion Neck (TFNeck), and Final Attention Head (FAHead).
• Experiments were conducted on COCO and MPII benchmarks to evaluate model performance.

Main Results:

• MEMe-based models achieve state-of-the-art performance with limited parameters.
• The EEffPose-P0 model, using only 8.98 M parameters, achieved 75.4 AP on the COCO validation set at 256x192 resolution.
• This performance surpasses HRNet-W48 while utilizing only 14% of its parameters.

Conclusions:

• The MEMe approach effectively enhances lightweight human pose estimation models.
• EEffPose demonstrates superior performance-to-parameter ratio compared to existing methods.
• This work provides an efficient and accurate solution for real-world human pose estimation applications.