Reproduce “ODeform: Learning Continuous 4D Motion for Shape Deformation with Neural ODEs” 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
Yordanka Velikova, Mahdi Saleh, Liming Kuang, Benjamin Busam
ODeform models continuous 4D object deformation using parallel Neural Ordinary Differential Equations (nODEs) to evolve rigid and local motion components in a learned latent space, integrating physical parameters. The framework demonstrates robust generalization to unseen geometries and physical conditions, achieving a 1.279 mm RMSE on unseen mass-bending parameters and superior performance in temporal interpolation and extrapolation compared to discrete learning methods.
Key insights
- Separating rigid and local deformation components within an nODE framework improves numerical stability and preserves physical structures, leading to more consistent and realistic deformation modeling.
- Representing 4D dynamics as continuous flows in a learned latent space enables seamless interpolation and extrapolation of deformation states, generalizing effectively to unseen physical parameters and object geometries.
- The model's ability to inversely identify physical parameters from observed deformations indicates that its learned latent space captures a meaningful representation of underlying physical properties.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.