Abstract:While Signed Distance Fields (SDF) are well-established for modeling watertight surfaces, Unsigned Distance Fields (UDF) broaden the scope to include open surfaces and models with complex inner structures. Despite their flexibility, UDFs encounter significant challenges in high-fidelity 3D reconstruction, such as non-differentiability at the zero level set, difficulty in achieving the exact zero value, numerous local minima, vanishing gradients, and oscillating gradient directions near the zero level set. To address these challenges, we propose Details Enhanced UDF (DEUDF) learning that integrates normal alignment and the SIREN network for capturing fine geometric details, adaptively weighted Eikonal constraints to address vanishing gradients near the target surface, unconditioned MLP-based UDF representation to relax non-negativity constraints, and a UDF-tailored method for extracting iso-surface with non-constant iso-values. These strategies collectively stabilize the learning process from unoriented point clouds and enhance the accuracy of UDFs. Our computational results demonstrate that DEUDF outperforms existing UDF learning methods in both accuracy and the quality of reconstructed surfaces. We will make the source code publicly available.

What problem does this paper attempt to address?

The paper aims to address the challenging issues of Unsigned Distance Fields (UDF) in high-fidelity 3D surface reconstruction. Compared to Signed Distance Fields (SDF), UDF can better represent open surfaces and objects with complex internal structures, but it faces many difficulties in accurately reconstructing 3D surfaces, such as non-differentiability at the zero level set, difficulty in achieving precise zero values, gradient vanishing, and gradient direction oscillation. To solve the above problems, the authors propose a Details Enhanced UDF (DEUDF) learning method. This method mainly includes the following aspects: 1. **Improvement of UDF representation**: By using an unconditional multi-layer perceptron (MLP) to represent UDF and relaxing the non-negativity constraint. This helps alleviate the gradient vanishing problem and stabilize the changes in gradient direction. 2. **Normal alignment**: Enhancing detail capture capability using local normal vector information. By aligning the UDF gradient with the normal vector, the accuracy of detail expression can be effectively improved even without considering global consistency. 3. **Adaptive weighted Eikonal constraint**: To address the problem of gradient vanishing near the zero level set in UDF, an adaptive weighted Eikonal constraint method is proposed to more effectively stabilize the learning process. 4. **UDF-customized isosurface extraction technique**: Considering that UDF may not strictly reach zero values, an optimization algorithm (DCUDF) is used to identify local minima as the target surface, thereby avoiding surface missing or breaking issues that may occur in traditional methods. Experimental results show that the DEUDF method outperforms existing UDF learning methods on multiple benchmark datasets, with improvements in detail expression and surface reconstruction quality. Additionally, comparative analysis and ablation experiments further validate the effectiveness and necessity of each component.

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

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

NeUDF: Leaning Neural Unsigned Distance Fields with Volume Rendering

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

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

SuperUDF: Self-supervised UDF Estimation for Surface Reconstruction

DUDF: Differentiable Unsigned Distance Fields with Hyperbolic Scaling

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

HSDF: Hybrid Sign and Distance Field for Neural Representation of Surfaces with Arbitrary Topologies

DCUDF2: Improving Efficiency and Accuracy in Extracting Zero Level Sets from Unsigned Distance Fields

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

Neural Surface Detection for Unsigned Distance Fields

Robust Zero Level-Set Extraction from Unsigned Distance Fields Based on Double Covering

Monge-Ampere Regularization for Learning Arbitrary Shapes from Point Clouds

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

Depth-NeuS: Neural Implicit Surfaces Learning for Multi-view Reconstruction Based on Depth Information Optimization

ND-SDF: Learning Normal Deflection Fields for High-Fidelity Indoor Reconstruction

Statistical Edge Detection And UDF Learning For Shape Representation

Recovering Fine Details for Neural Implicit Surface Reconstruction

Surface-Centric Modeling for High-Fidelity Generalizable Neural Surface Reconstruction