Learning a Contact Potential Field for Modeling the Hand-Object Interaction.

Area of Science:

• Robotics and Human-Computer Interaction
• Computer Vision
• Machine Learning

Background:

• Understanding human hand-object manipulation is crucial for human behavior analysis.
• Accurate modeling of hand-object interactions requires considering both pose and contact.
• Existing methods often overlook the importance of contact for generating plausible grasps.

Purpose of the Study:

• To propose an explicit contact representation, Contact Potential Field (CPF), for hand-object interactions.
• To enhance the accuracy and physical plausibility of hand-object pose estimation and grasp generation.
• To demonstrate the effectiveness of CPF on challenging datasets and tasks.

Main Methods:

• Developed Contact Potential Field (CPF) to represent hand-object contact as a spring-mass system.
• Modeled contact stability and distance using the spring-mass system, creating an elastic potential field.
• Applied CPF to hand-object pose estimation and grasping pose generation tasks.

Main Results:

• Achieved state-of-the-art results in several reconstruction metrics on challenging tasks and datasets.
• Demonstrated the ability to generate more physically plausible hand-object poses.
• Showcased improved performance even with ground-truth data exhibiting interpenetration or disjointedness.

Conclusions:

• Contact Potential Field (CPF) offers a robust method for modeling hand-object contact.
• CPF significantly improves the realism and accuracy of hand-object pose and grasp generation.
• The proposed method advances the understanding and simulation of human manipulation.