Reproduce “LEXIS: LatEnt ProXimal Interaction Signatures for 3D HOI from an Image” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Dimitrije Antić, Alvaro Budria, George Paschalidis, Sai Kumar Dwivedi, Dimitrios Tzionas
Researchers at the University of Amsterdam and Max Planck Institute for Intelligent Systems developed LEXIS-Flow, a one-stage generative framework that reconstructs 3D Human-Object Interaction from a single image using dense, continuous proximity representations called InterFields and learned interaction signatures. The method achieved state-of-the-art results on the Open3DHOI and BEHAVE datasets, improving human Chamfer Distance by 15.0% and object Chamfer Distance by 5.0% on BEHAVE.
Key insights
- Dense, continuous InterFields provide a significantly richer and more effective spatial signal for 3D HOI reconstruction compared to sparse, binary contact cues.
- Learning a compact, discrete lexicon of interaction signatures (LEXIS) via a VQ-VAE effectively regularizes the ill-posed problem of inferring high-dimensional 3D interaction fields from 2D images.
- Integrating InterField-guided refinement into a one-stage generative flow-matching framework allows for efficient and physically plausible HOI reconstruction directly from an image, bypassing the limitations of separate optimization steps.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.