Locally Connected Network for Monocular 3D Human Pose Estimation.

Area of Science:

• Computer Vision
• Machine Learning
• Human Pose Estimation

Background:

• Graph Convolutional Networks (GCNs) are standard for human pose tasks.
• GCNs have limitations in 3D pose estimation due to weight sharing.
• Existing methods often struggle with generalizing across datasets.

Purpose of the Study:

• To address limitations of GCNs in 3D human pose estimation.
• To introduce a novel network architecture, Locally Connected Network (LCN), for improved 3D pose estimation.
• To enhance cross-dataset generalization capabilities.

Main Methods:

• Developed a novel reformulation of GCNs, showing GCN and Fully Connected Networks (FCNs) as special cases.
• Introduced Locally Connected Network (LCN) with dedicated filters for each joint to overcome GCN limitations.
• Jointly trained LCN with a 2D pose estimator to handle inaccurate 2D inputs.

Main Results:

• LCN significantly outperforms GCN, FCN, and state-of-the-art methods on benchmark datasets.
• The proposed approach demonstrates strong cross-dataset generalization.
• Sparse connections in LCN contribute to its generalization ability.

Conclusions:

• Locally Connected Networks (LCNs) offer a superior alternative to GCNs for 3D human pose estimation.
• The LCN approach enhances accuracy and robustness, particularly in cross-dataset scenarios.
• This work provides a new direction for developing more effective human pose estimation models.