Neural Splines: Fitting 3D Surfaces with Infinitely-Wide Neural Networks
About
We present Neural Splines, a technique for 3D surface reconstruction that is based on random feature kernels arising from infinitely-wide shallow ReLU networks. Our method achieves state-of-the-art results, outperforming recent neural network-based techniques and widely used Poisson Surface Reconstruction (which, as we demonstrate, can also be viewed as a type of kernel method). Because our approach is based on a simple kernel formulation, it is easy to analyze and can be accelerated by general techniques designed for kernel-based learning. We provide explicit analytical expressions for our kernel and argue that our formulation can be seen as a generalization of cubic spline interpolation to higher dimensions. In particular, the RKHS norm associated with Neural Splines biases toward smooth interpolants.
Related benchmarks
| Task | Dataset | Result | Rank | |
|---|---|---|---|---|
| Scene Reconstruction | SceneNet (test) | Chamfer Distance (CD)0.6 | 16 | |
| Surface Reconstruction | DC benchmark 1.0 (test) | Chamfer Distance (GT)0.14 | 9 | |
| Surface Reconstruction | ShapeNet 260 shapes 15 | sCD (mean)5.36e-5 | 9 | |
| Surface Reconstruction | Gargoyle 1.0 (test) | Chamfer Distance (GT)0.16 | 9 | |
| Surface Reconstruction | Lord Quas benchmark 1.0 (test) | Chamfer Distance (GT)0.12 | 9 | |
| 3D surface reconstruction | ShapeNet 3k points | Reconstruction Time12.7 | 9 | |
| Surface Reconstruction | Anchor 1.0 (test) | Chamfer Distance (GT)0.22 | 9 | |
| Surface Reconstruction | Daratech benchmark 1.0 (test) | Chamfer Distance (GT)0.21 | 9 |