3D Human Pose Machines with Self-Supervised Learning.

Area of Science:

• Computer Vision
• Robotics
• Human Pose Estimation

Background:

• Recovering 3D human poses from monocular images is challenging due to appearance variations, occlusions, and geometric ambiguities.
• Existing methods struggle with limited 3D data and the 2D-3D domain gap, hindering scalability in real-world scenarios.

Purpose of the Study:

• To propose a self-supervised correction mechanism for learning intrinsic human pose structures.
• To develop an accurate and scalable 3D human pose estimation framework adaptable to diverse conditions.

Main Methods:

• Introduced a self-supervised mechanism with dual learning tasks: 2D-to-3D pose transformation and 3D-to-2D pose projection.
• The 2D-to-3D task regresses intermediate 3D poses using temporal context.
• The 3D-to-2D task refines 3D poses by enforcing geometric consistency with 2D projections.

Main Results:

• The dual learning tasks enable adaptive learning from both 3D and large-scale 2D human pose data.
• Developed a 3D human pose machine integrating 2D spatial relationships, temporal smoothness, and 3D geometric knowledge.
• Achieved superior performance and efficiency on Human3.6M and HumanEva-I benchmarks.

Conclusions:

• The proposed self-supervised correction mechanism effectively addresses limitations of existing 3D human pose estimation methods.
• The framework demonstrates robust performance and scalability for practical applications.
• This approach offers a promising direction for accurate and efficient 3D human pose recovery.