Abstract:Surface reconstruction for point clouds is an important task in 3D computer vision. Most of the latest methods resolve this problem by learning signed distance functions from point clouds, which are limited to reconstructing closed surfaces. Some other methods tried to represent open surfaces using unsigned distance functions (UDF) which are learned from ground truth distances. However, the learned UDF is hard to provide smooth distance fields due to the discontinuous character of point clouds. In this paper, we propose CAP-UDF, a novel method to learn consistency-aware UDF from raw point clouds. We achieve this by learning to move queries onto the surface with a field consistency constraint, where we also enable to progressively estimate a more accurate surface. Specifically, we train a neural network to gradually infer the relationship between queries and the approximated surface by searching for the moving target of queries in a dynamic way. Meanwhile, we introduce a polygonization algorithm to extract surfaces using the gradients of the learned UDF. We conduct comprehensive experiments in surface reconstruction for point clouds, real scans or depth maps, and further explore our performance in unsupervised point normal estimation, which demonstrate non-trivial improvements of CAP-UDF over the state-of-the-art methods.

What problem does this paper attempt to address?

The paper primarily focuses on addressing the problem of surface reconstruction from 3D point cloud data, particularly for open and multi-layer surfaces. Traditional methods mostly rely on learning Signed Distance Functions (SDFs), which are limited to reconstructing closed surfaces. Additionally, some attempts using Unsigned Distance Functions (UDFs) can represent open surfaces but face difficulties in learning smooth distance fields. To address the above issues, the paper proposes a novel method called CAP-UDF, with the following core contributions: 1. **Proposes a new network architecture** for directly learning Consistency-Aware Unsigned Distance Functions (Consistency-Aware UDFs) from raw point cloud data. This method does not require ground truth distance values or point normals as training inputs. 2. **Designs dynamic optimization objectives** by using specially designed loss functions to constrain the consistency of the field, thereby encouraging the network to learn more consistent and smooth unsigned distance fields. This method allows the network to gradually infer the relationship between query positions and the approximate surface. 3. **Introduces an algorithm for directly extracting high-fidelity isosurface** by utilizing the gradients of the learned unsigned distance function to determine whether two queries are on the same side or opposite sides of the approximate surface, and further uses an improved Marching Cubes algorithm to extract the surface. 4. **Conducts comprehensive experimental evaluations**, including surface reconstruction tasks on synthetic point clouds, real scan data, and depth maps, as well as unsupervised point normal estimation tasks. Experimental results show that CAP-UDF has significant advantages in efficiency and accuracy compared to existing techniques. In summary, CAP-UDF aims to overcome the limitations of existing UDF-based methods in learning smooth distance fields while providing a direct and efficient method for extracting high-quality surfaces from raw point cloud data.

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

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

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

Learnable Chamfer Distance for Point Cloud Reconstruction

Learning a More Continuous Zero Level Set in Unsigned Distance Fields through Level Set Projection

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

Details Enhancement in Unsigned Distance Field Learning for High-fidelity 3D Surface Reconstruction

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

NeuralUDF: Learning Unsigned Distance Fields for Multi-view Reconstruction of Surfaces with Arbitrary Topologies

Learning Continuous Implicit Field with Local Distance Indicator for Arbitrary-Scale Point Cloud Upsampling

Unsigned Orthogonal Distance Fields: An Accurate Neural Implicit Representation for Diverse 3D Shapes

Reconstructing Surfaces for Sparse Point Clouds with On-Surface Priors

NeUDF: Leaning Neural Unsigned Distance Fields with Volume Rendering

DUDF: Differentiable Unsigned Distance Fields with Hyperbolic Scaling

SuperUDF: Self-supervised UDF Estimation for Surface Reconstruction

PU-Ray: Domain-Independent Point Cloud Upsampling via Ray Marching on Neural Implicit Surface

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

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

Surface Reconstruction from Point Clouds by Learning Predictive Context Priors

Monge-Ampere Regularization for Learning Arbitrary Shapes from Point Clouds

Grad-PU: Arbitrary-Scale Point Cloud Upsampling via Gradient Descent with Learned Distance Functions