Limb Pose Aware Networks for Monocular 3D Pose Estimation.

Area of Science:

• Computer Vision
• Human Pose Estimation
• Machine Learning

Background:

• Monocular 3D pose estimation faces challenges with higher degree of freedom (DOF) limb joints, leading to accumulated errors.
• Existing methods struggle with the complexity of joint trajectories and error propagation along human body structures.

Purpose of the Study:

• To develop a novel framework that enhances the accuracy of 3D limb joint position prediction.
• To address the issue of error accumulation in human pose estimation, particularly for limb joints.

Main Methods:

• Proposes a limb pose aware framework integrating a kinematic constraint aware network and a trajectory aware temporal module.
• Introduces two kinematic constraints: relative bone angles and absolute bone angles, to model angular relationships.
• Employs a Hierarchical Transformer network to process and fuse temporal joint trajectories.

Main Results:

• The framework effectively suppresses accumulated errors along human limbs.
• Achieves promising results in improving 3D pose prediction accuracy.
• Demonstrates compatibility with existing 2D pose estimators.

Conclusions:

• The proposed framework significantly alleviates error accumulation in 3D human pose estimation.
• Kinematic constraints and the trajectory-aware network are key to improving prediction accuracy.
• The method shows strong performance validated through extensive experiments and ablation studies.