Reproduce “Stochastic Preconditioning for Neural Field Optimization” 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
Selena Ling, Merlin Nimier-David, Alec Jacobson, Nicholas Sharp
Stochastic preconditioning improves neural-field fitting by adding Gaussian noise to query coordinates and gradually reducing its scale, smoothing optimization before restoring detail. On DTU surface reconstruction, Neuralangelo’s mean PSNR rises from 35.87 to 36.26 and Chamfer distance falls from 0.87 to 0.76 across 15 scenes. It also recovers ReLU-field quality without hierarchy, but takes longer than hierarchical training.
Key insights
- Spatial perturbation during training implicitly regularizes neural field optimization by smoothing the landscape, guiding the network away from poor local minima.
- Optimizing with stochastically perturbed queries in expectation fits a Gaussian-blurred version of the neural field, effectively adopting a coarse-to-fine strategy.
- The blurring acts on the neural field being optimized, not the ground truth, enabling universal applicability to both directly and indirectly supervised problems.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.