DUA: A Domain-Unified Approach for Cross-Dataset 3D Human Pose Estimation

Area of Science:

• Computer Vision
• Machine Learning
• Robotics

Background:

• Deep learning excels in 2D human pose estimation.
• 3D pose estimation offers more robust results but lacks sufficient labeled data.
• Current two-step methods using 2D pose inputs struggle with generalization across datasets.

Purpose of the Study:

• To address pose misalignment in cross-dataset 3D human pose estimation.
• To improve the generalization capability of 3D pose estimation systems.
• To leverage synthetic data for enhanced domain adaptation.

Main Methods:

• Proposed a novel Domain Unified approach.
• Integrated three modules: pose converter, uncertainty estimator, and domain classifier.
• Employed domain adaptation techniques using synthetic (SURREAL) and real (Human3.6M) datasets.

Main Results:

• Achieved a 44.1mm (29.24%) error reduction compared to a no-adaptation scenario.
• Demonstrated state-of-the-art performance in cross-dataset evaluation.
• Successfully mitigated pose misalignment issues.

Conclusions:

• The Domain Unified approach effectively improves 3D human pose estimation accuracy and robustness.
• Domain adaptation using synthetic data is crucial for generalizing pose estimation models.
• The proposed method offers a significant advancement for real-world human pose understanding.