NoiseSDF2NoiseSDF: Learning Clean Neural Fields from Noisy Supervision

2025-07-17

AI Summary

This work addresses the challenge of reconstructing high-fidelity implicit surfaces from low-quality, heavily noisy scanned point clouds. We propose NoiseSDF—the first method to extend the Noise2Noise paradigm to 3D neural implicit fields—enabling robust Signed Distance Function (SDF) learning without clean ground-truth supervision. Our approach constructs paired noisy SDF supervision signals and jointly optimizes point cloud denoising and implicit surface reconstruction via mean-squared-error minimization in a deep neural network. The key contribution is the first demonstration that high-quality neural SDFs can be trained exclusively from noisy point clouds, eliminating reliance on ideal geometric ground truth. Extensive evaluation on ShapeNet, ABC, Famous, and real-world scanned datasets shows that NoiseSDF significantly improves surface reconstruction accuracy and generalization over prior methods. This work establishes a new unsupervised paradigm for neural 3D field modeling.

Technology Category

• Computer Vision: 3D Computer Vision
• Machine Learning: Learning with Manifolds
• Search and Optimization: Learning to Search

Application Category

• Search and Retrieval-Augmented AI: Web learning to rank, online learning, and counterfactual learning for ranking
• Graph Algorithms and Modeling for the Web: Graph embeddings and representation learning for Web-related graphs
• Economics, Online Markets and Human Computation: LLM based quality controls for crowd work

Abstract

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 directly 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.

Problem

• Reconstructing accurate implicit surfaces from noisy point clouds
• Extending Noise2Noise paradigm to 3D neural fields
• Learning clean neural SDFs directly from noisy supervision

Innovation

• Extends Noise2Noise to 3D neural fields
• Learns clean SDFs from noisy point clouds
• Minimizes MSE loss for implicit denoising