Reproduce “RDA-INR: Riemannian Diffeomorphic Autoencoding via Implicit Neural Representations” 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
Sven Dummer, Nicola Strisciuglio, Christoph Brune
The RDA-INR framework integrates LDDMM's Riemannian geometry with implicit neural representations to create a resolution-independent and geometrically consistent deep learning model for statistical shape analysis. This approach accurately learns Fréchet means and geodesic deformations, outperforming existing methods in reconstruction quality, generalization, and noise robustness for point cloud and mesh data.
Key insights
- Riemannian regularization is crucial for learning data-consistent Fréchet means and generating physically plausible geodesic deformations, especially when specific geometric priors (e.g., rigidity) are relevant to the data.
- Implicit Neural Representations inherently provide resolution independence, allowing the model to learn high-quality, smooth templates and reconstructions regardless of input data resolution or initial quality.
- The combination of LDDMM Riemannian geometry and INRs significantly improves model generalization and robustness to noise in the latent space by constraining deformations to a geometrically sound manifold.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.