The article presents HANDY PRIORS, a unified framework for estimating pose in human-object interaction scenarios. This approach leverages recent advances in differentiable physics and rendering. HANDY PRIORS uses rendering priors to align with input images and physics priors to prevent penetration and relative-sliding across frames. The authors also propose two alternatives for hand and object pose estimation: optimization-based pose estimation for higher accuracy and filtering-based tracking for faster execution. The study demonstrates that HANDY PRIORS achieves comparable or better results in pose estimation tasks and that the differentiable physics module can predict contact information for pose refinement.
Publication date: 29 Nov 2023
Project Page: Not provided
Paper: https://arxiv.org/pdf/2311.16552
