Reproduce “Grounding and Enhancing Grid-based Models for Neural Fields” 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
Zelin Zhao, Fenglei Fan, Wenlong Liao, Junchi Yan
This research introduces Grid Tangent Kernels (GTKs) to provide the first systematic theoretical framework for grid-based neural fields, revealing their stationary behavior during training. Guided by this theory, the proposed MulFAGrid model achieves state-of-the-art performance in 2D image fitting (56.19 dB PSNR), 3D SDF reconstruction (4.51 NAE), and novel view synthesis (e.g., 28.98 dB PSNR on Mip-NeRF-360) across multiple benchmarks.
Key insights
- The Grid Tangent Kernel (GTK) of a grid-based model remains stationary (unchanged) throughout the training process, allowing it to be modeled as a linear kernelized system.
- GTK directly governs the training dynamics of grid-based neural fields, connecting model output evolution to training error.
- A derived generalization bound, dependent on the term Y^T G_g^-1 Y, provides a theoretical metric to assess and compare the generalization capabilities of different grid architectures.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.