[CVPR 2021 Oral] ContactOpt: Optimizing Contact to Improve Grasps скачать в хорошем качестве

[CVPR 2021 Oral] ContactOpt: Optimizing Contact to Improve Grasps

Paper: https://arxiv.org/abs/2104.07267 Paper website: https://www.pgrady.net/contactopt/ Abstract: Physical contact between hands and objects plays a critical role in human grasps. We show that optimizing the pose of a hand to achieve expected contact with an object can improve hand poses inferred via image-based methods. Given a hand mesh and an object mesh, a deep model trained on ground truth contact data infers desirable contact across the surfaces of the meshes. Then, ContactOpt efficiently optimizes the pose of the hand to achieve desirable contact using a differentiable contact model. Notably, our contact model encourages mesh interpenetration to approximate deformable soft tissue in the hand. In our evaluations, our methods resulted in grasps that better matched ground truth contact, had lower kinematic error, and were significantly preferred by human participants.