Reproduce “4Deform: Neural Surface Deformation for Robust Shape Interpolation” 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

Lu Sang, Zehranaz Canfes, Dongliang Cao, Riccardo Marin, Florian Bernard, Daniel Cremers

Researchers at Technical University of Munich and collaborators developed 4Deform, a neural framework that generates realistic intermediate shapes between non-rigidly deformed 3D inputs. It robustly interpolates dynamic 3D data, including noisy point clouds and partial observations, by leveraging a physically grounded velocity field and novel deformation losses.

Key insights

  • A continuous velocity field, combined with an implicit neural field, enables direct incorporation of physical constraints and robust handling of unstructured input data.
  • Novel connectivity-agnostic distortion and stretching losses, derived from continuum mechanics, are crucial for producing physically plausible deformations, even with large or non-isometric shape changes.
  • Integrating an unsupervised correspondence estimation block allows the system to operate effectively with sparse, noisy, and incomplete real-world correspondences.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.