Anatomy-guided domain adaptation for 3D in-bed human pose estimation

Area of Science:

• Computer Vision
• Medical Imaging
• Machine Learning

Background:

• 3D human pose estimation is crucial for clinical monitoring.
• Deep learning models struggle with domain shifts and require extensive labeled data.
• Existing methods lack robust generalization in real-world clinical settings.

Purpose of the Study:

• To develop a novel domain adaptation method for 3D human pose estimation.
• To improve model generalization from labeled source domains to unlabeled target domains.
• To address limitations in clinical applicability due to domain shifts.

Main Methods:

• A novel domain adaptation method using anatomical priors.
• Anatomical loss function penalizing implausible poses (limb/bone lengths, joint angles).
• Anatomically plausible pseudo-label filtering integrated with the Mean Teacher paradigm for self-training.

Main Results:

• Consistent outperformance over state-of-the-art domain adaptation methods.
• Significant improvement over baseline models (31%/66%) in in-bed pose estimation.
• Substantial reduction in domain gap (65%/82%) using public and newly created datasets.

Conclusions:

• The proposed method effectively adapts 3D human pose estimation models to new domains using anatomical knowledge.
• This approach enhances clinical applicability by improving generalization and reducing reliance on labeled data.
• The framework supports unsupervised and source-free domain adaptation for 3D human pose estimation.