3D Interacting Hands Diffusion Model
Supplementary Material: zip
Primary Area: representation learning for computer vision, audio, language, and other modalities
Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.
Keywords: 3D interacting hands, generative model
Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2024/AuthorGuide.
TL;DR: We introduce a new generative model for 3D interacting hands.
Abstract: Humans make two-hands interactions in a variety of ways. Learning prior distributions of interactions between hands is critical for 1) generating new interacting hands and 2) recovering plausible and accurate interacting hands. Unfortunately, there have been no attempts to learn the prior distribution of interactions between two hands. Due to the lack of prior distribution, previous 3D interacting hands recovery methods often produce hands with physically implausible interactions, such as severe collisions, and semantically meaningless interactions. We present IHDiff, the first generative model for learning the prior distribution of interacting hands. Motivated by the strong performance of recent diffusion models, we learn the prior distributions using the diffusion process. For the reverse diffusion process, we design a novel Transformer-based network, which effectively captures correlations between joints of two hands using self- and cross-attention. We showcase three applications of IHDiff including random sampling, conditional random sampling, and fitting to observations.
The code and pre-trained model will be publicly available.
Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors' identity.
No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.
Submission Number: 5798
Loading