Abstract

Synthesizing accurate hand–object interactions (HOI) is critical for AR/VR and vision tasks.
Existing dense–correspondence methods improve contact fidelity but lack full differentiability or generality.
We propose CHOIR, a versatile, fully differentiable interaction field:

• Unsigned distance fields encode hand & object shapes continuously.
• Gaussian contact maps capture dense hand-centric contact distributions with few parameters.
  We integrate CHOIR into JointDiffusion, a diffusion model that learns CHOIR distributions for both:

1. Refinement: improves noisy reconstructions (contact F1 ↑ 5%).
2. Synthesis: generates grasps from object geometry alone (sim. displacement ↓ 46%).
   JointDiffusion+CHOIR outperforms SOTA on refinement and synthesis benchmarks.

Teaser

CHOIR jointly represents hand and object geometries via unsigned distance fields, and captures hand-centric contact distributions with 3D Gaussians—enabling a fully differentiable and versatile hand-object interaction model.

Results Carousel

Method Overview

Static Grasp Denoising

Static Grasp Synthesis

Reuses the above carousel of 11 object-specific grasp synthesis videos.

BibTeX

@article{morales2024choir,
  author  = {Morales, Théo and Taheri, Omid and Lacey, Gerard},
  title   = {A Versatile and Differentiable Hand-Object Interaction Representation},
  journal = {Winter Conference on Applications of Computer Vision (WACV)},
  year    = {2025},
}