Learning Dynamical Human-Joint Affinity for 3D Pose Estimation in Videos.

Area of Science:

• Computer Vision
• Machine Learning
• Human Pose Estimation

Background:

• Graph Convolutional Networks (GCNs) are used for 3D human pose estimation.
• Fixed human-joint affinities in GCNs limit adaptation to complex video pose variations.

Purpose of the Study:

• To propose a Dynamical Graph Network (DG-Net) for adaptive 3D human pose estimation.
• To dynamically identify human-joint affinity and learn spatial/temporal joint relations from videos.

Main Methods:

• Introduced Dynamical Spatial Graph convolution (DSG) and Dynamical Temporal Graph convolution (DTG).
• DSG/DTG discover joint affinities based on spatial distance and temporal movement similarity.
• These methods reduce depth ambiguity and motion uncertainty in 2D to 3D pose lifting.

Main Results:

• DG-Net outperforms state-of-the-art (SOTA) approaches on Human3.6M, HumanEva-I, and MPI-INF-3DHP benchmarks.
• Achieved superior performance with fewer input frames and a smaller model size.
• Demonstrated effective understanding of spatial and temporal joint relationships.

Conclusions:

• DG-Net offers a more adaptive and accurate approach to 3D human pose estimation in videos.
• Dynamic identification of joint affinity enhances robustness to pose variations.
• The proposed method advances the field by improving efficiency and performance.