Coarse-to-Fine Hand-Object Pose Estimation with Interaction-Aware Graph Convolutional Network.

Area of Science:

• Computer Vision
• Robotics
• Human-Computer Interaction

Background:

• Accurate hand-object pose estimation from RGB images is crucial for understanding human behavior and enabling advanced applications.
• Existing methods often struggle with the dynamic and complex relationships between hands and objects in 3D space.

Purpose of the Study:

• To propose a novel coarse-to-fine framework for precise 3D hand-object pose estimation.
• To explicitly model and leverage hand-object relations for improved pose refinement.

Main Methods:

• A two-stage approach: coarse stage estimates initial 3D poses from 2D keypoints, and a fine stage refines them.
• Introduction of InterGCN, an interaction-aware graph convolutional network, to capture dynamic hand-object relationships in 3D.
• Utilizing both general and interaction-specific relation graphs to enhance network adaptability across diverse scenarios.

Main Results:

• The proposed framework achieves state-of-the-art performance on benchmark hand-object datasets.
• Demonstrated significant improvements in 3D pose refinement accuracy.
• The InterGCN effectively handles variations in hand-object interactions.

Conclusions:

• The novel coarse-to-fine framework with InterGCN offers a robust solution for 3D hand-object pose estimation.
• Explicitly modeling 3D hand-object relations is key to overcoming challenges in pose refinement.
• The approach shows great potential for applications requiring detailed human-object interaction analysis.