Federated Learning for Human Pose Estimation on Non-IID Data via Gradient Coordination.

Area of Science:

• Computer Vision
• Machine Learning
• Robotics

Background:

• Decentralized data in human pose estimation often presents non-independent and identically distributed (non-IID) characteristics.
• Traditional federated learning aggregation strategies struggle with gradient conflicts, hindering model convergence and accuracy in non-IID environments.

Purpose of the Study:

• To propose a novel federated learning aggregation strategy, Federated Gradient Harmonization (FedGH), to address gradient conflicts in non-IID settings for human pose estimation.
• To enhance the accuracy and robustness of distributed human pose estimation models.

Main Methods:

• FedGH coordinates client update directions by measuring gradient discrepancies.
• It integrates gradient-projection correction with a parameter-reconstruction mechanism to harmonize updates.
• The strategy was tested on a self-constructed robotic dataset and the MPII Human Pose Dataset.

Main Results:

• FedGH achieved average Percentage of Correct Keypoints (PCK) of 47.14% and 66.31% on the MPII dataset, outperforming Federated Adaptive Weighting (FedAW).
• On the self-constructed dataset, FedGH reached 86.4% PCK for shoulder detection, surpassing other methods by 20-30%.
• The keypoint heatmap regression model using FedGH achieved over 98% accuracy within 10 rounds on the self-constructed dataset.

Conclusions:

• FedGH effectively mitigates gradient conflicts in non-IID environments, offering a more robust optimization solution for distributed human pose estimation.
• The proposed method demonstrates significant improvements in accuracy and convergence for federated human pose estimation tasks.