Abstract:Learning signed distance functions (SDFs) from point clouds is an important task in 3D computer vision. However, without ground truth signed distances, point normals or clean point clouds, current methods still struggle from learning SDFs from noisy point clouds. To overcome this challenge, we propose to learn SDFs via a noise to noise mapping, which does not require any clean point cloud or ground truth supervision. Our novelty lies in the noise to noise mapping which can infer a highly accurate SDF of a single object or scene from its multiple or even single noisy observations. We achieve this by a novel loss which enables statistical reasoning on point clouds and maintains geometric consistency although point clouds are irregular, unordered and have no point correspondence among noisy observations. To accelerate training, we use multi-resolution hash encodings implemented in CUDA in our framework, which reduces our training time by a factor of ten, achieving convergence within one minute. We further introduce a novel schema to improve multi-view reconstruction by estimating SDFs as a prior. Our evaluations under widely-used benchmarks demonstrate our superiority over the state-of-the-art methods in surface reconstruction from point clouds or multi-view images, point cloud denoising and upsampling.

What problem does this paper attempt to address?

This paper attempts to solve the problem of learning signed distance functions (SDFs) from noisy point clouds. Specifically, current methods still have difficulty in learning SDFs from noisy point clouds in the absence of ground - truth signed distances, point normals, or clean point clouds. This paper proposes a method for learning SDFs through noise - to - noise mapping, which does not require any clean point clouds or ground - truth supervision. Its innovation lies in the noise - to - noise mapping, which can infer highly accurate SDFs from multiple or even a single noisy observation. ### Main Contributions 1. **Unsupervised**: Proposed a method for learning SDFs from noisy point clouds without requiring ground - truth signed distances, point normals, or clean point clouds. 2. **Statistical Inference**: Proved that statistical inference can be carried out through the Earth Mover's Distance (EMD) even when 3D point clouds are irregular, unordered, and there is no point - to - point correspondence between different observations. 3. **Accelerated Training**: Significantly accelerated the training process by integrating multi - resolution hash encoding, reducing the convergence time of SDFs to within one minute. 4. **Multi - view Reconstruction**: Proposed a new scheme to use the SDFs learned from noisy SfM points as priors to improve the accuracy of multi - view reconstruction. 5. **State - of - the - Art Results**: Achieved state - of - the - art results in multiple applications such as surface reconstruction, point - cloud denoising, and up - sampling. ### Method Overview - **Statistical Inference**: Through noise - to - noise mapping, utilize the statistical relationships between multiple noisy observations to infer the structures unaffected by noise. - **Loss Function**: Designed a new loss function that includes geometric consistency regularization to ensure the geometric consistency of the learned SDFs. - **Fast Learning Framework**: Utilize multi - resolution hash encoding to improve the inference efficiency of SDFs and significantly reduce the training time. ### Specific Implementation - **Denoising Function**: Defined a denoising function \(F\), which pulls any point \(q\) on the noisy point cloud back to the surface through the signed distance \(d = f_{\theta}(q)\) and the gradient \(\nabla f_{\theta}(q)\). - **Distance Metric**: Use the Earth Mover's Distance (EMD) as the distance metric \(L\) to ensure that the denoised point cloud \(S'\) can converge to the surface unaffected by noise. - **Geometric Consistency**: Introduced a geometric consistency regularization term \(R\) to ensure the geometric consistency of the learned SDFs. ### Experimental Results - **Surface Reconstruction**: Demonstrated superior performance in reconstructing surfaces from noisy point clouds in multiple benchmarks. - **Point - Cloud Denoising and Up - sampling**: Achieved better results than existing methods in point - cloud denoising and up - sampling tasks. - **Multi - view Reconstruction**: Significantly improved the accuracy of multi - view reconstruction by using the learned SDF as a prior. In conclusion, this paper proposes an innovative method that can efficiently learn signed distance functions from noisy point clouds without ground - truth supervision and has achieved significant performance improvements in multiple application scenarios.

Fast Learning of Signed Distance Functions from Noisy Point Clouds via Noise to Noise Mapping

Learning Signed Distance Functions from Noisy 3D Point Clouds via Noise to Noise Mapping

SDFMAP: Neural Signed Distance Fields for Mapping and Positioning in Real-Time

Inferring Neural Signed Distance Functions by Overfitting on Single Noisy Point Clouds through Finetuning Data-Driven based Priors

Unsupervised Inference of Signed Distance Functions from Single Sparse Point Clouds Without Learning Priors

NeuralTPS: Learning Signed Distance Functions without Priors from Single Sparse Point Clouds

CAP-UDF: Learning Unsigned Distance Functions Progressively from Raw Point Clouds with Consistency-Aware Field Optimization

Neural-Pull: Learning Signed Distance Functions from Point Clouds by Learning to Pull Space Onto Surfaces

Learning Density Regulated and Multi-View Consistent Unsigned Distance Fields.

Learning Unsigned Distance Fields from Local Shape Functions for 3D Surface Reconstruction

Implicit Filtering for Learning Neural Signed Distance Functions from 3D Point Clouds

Reconstructing Surfaces for Sparse Point Clouds with On-Surface Priors

N$^{3}$-Mapping: Normal Guided Neural Non-Projective Signed Distance Fields for Large-scale 3D Mapping

PU-SDF: Arbitrary-Scale Uniformly Upsampling Point Clouds Via Signed Distance Functions

Fast Monocular Scene Reconstruction with Global-Sparse Local-Dense Grids

Monge-Ampere Regularization for Learning Arbitrary Shapes from Point Clouds

Neural Signed Distance Function Inference through Splatting 3D Gaussians Pulled on Zero-Level Set

Toward Robust Neural Reconstruction from Sparse Point Sets

SDFReg: Learning Signed Distance Functions for Point Cloud Registration

Research on a 3D Point Cloud Map Learning Algorithm Based on Point Normal Constraints

3D Point Cloud Denoising and Normal Estimation for 3D Surface Reconstruction.