Riccardo Marin

A unified framework for generative 3D Human-Object Interaction, capable of operating in seven different modes. Built on top of the Unidiffuser formulation, our method is trained to process different combinations of input and infer the unobserved ones as a conditional denoising approach. Our method is general, covering the use cases of all previous works. At the same time, it outperforms networks specialized in individual tasks, demonstrating the benefit of a more comprehensive formulation of Human-Object Interaction.

We recover a 4D realistic deformation between two sparse unordered point clouds. After applying shape matching and registration methods to obtain rough and noisy correspondence, we leverage the theory of level-set methods to infer a deformation that is continuous and physically plausible.

We propose a novel localized Neural Field (LoVD), the first self-supervised task for tuning neural fields (INT), and an efficient (takes less than a minute) scalable registration pipeline (NSR), that works with out-of-distribution data (partial point clouds, clutter, different poses, …).