NoiseSDF2NoiseSDF: Learning Clean Neural Fields from Noisy Supervision

Reconstructing accurate implicit surface representations from point clouds remains a challenging task, particularly when data is captured using low-quality scanning devices. These point clouds often contain substantial noise, leading to inaccurate surface reconstructions. Inspired by the Noise2Noise paradigm for 2D images, we introduce NoiseSDF2NoiseSDF, a novel method designed to extend this concept to 3D neural fields. Our approach enables learning clean neural SDFs from noisy point clouds through noisy supervision by minimizing the MSE loss between noisy SDF representations, allowing the network to implicitly denoise and refine surface estimations. We evaluate the effectiveness of NoiseSDF2NoiseSDF on benchmarks, including the ShapeNet, ABC, Famous, and Real datasets. Experimental results demonstrate that our framework significantly improves surface reconstruction quality from noisy inputs.

Neural fields are a way for computers to describe 3D shapes smoothly, but they usually need clean training data to learn well. In practice, 3D measurements are often noisy, and clean examples can be hard or expensive to obtain. This paper introduces NoiseSDF2NoiseSDF, a method that learns a cleaner 3D field from noisy versions of the same shape. The main idea is simple: although two noisy measurements may each contain mistakes, they still share the same underlying shape information. By learning from one noisy field to another, the model keeps what is consistent and reduces what is random. This adapts the noise-to-noise learning idea to 3D neural fields. It makes it easier to learn useful 3D shape representations when only imperfect supervision is available.