GTIGNet: Global Topology Interaction Graphormer Network for 3D hand pose estimation.

Area of Science:

• Computer Vision
• Machine Learning
• Deep Learning

Background:

• Estimating 3D hand poses from monocular RGB images is difficult due to complex hand structures, self-occlusions, and depth ambiguities.
• Current methods struggle to capture long-range dependencies in hand joint connections.

Purpose of the Study:

• To introduce a novel deep learning architecture, the Global Topology Interaction Graphormer Network (GTIGNet), for enhanced 3D hand pose estimation.
• To address limitations in capturing skeletal and non-skeletal connections for hand joints.

Main Methods:

• Developed GTIGNet, incorporating a Context-Aware Attention Block (CAAB) for improved multi-scale feature extraction in 2D pose estimation.
• Introduced a High-Order Graphormer to explicitly and implicitly model the topological structure of hand joints, enhancing feature interaction.

Main Results:

• GTIGNet achieved state-of-the-art performance across four challenging datasets: RHD, STB, FPHA, and FreiHAND.
• Achieved low Mean Per Joint Position Error (MPJPE): 9.98 mm (RHD), 6.12 mm (STB), 11.15 mm (FPHA), and 10.97 mm (FreiHAND).

Conclusions:

• GTIGNet effectively addresses limitations in existing 3D hand pose estimation methods.
• The proposed architecture demonstrates superior performance in accurately estimating 3D hand poses from monocular RGB images.