- The paper introduces a neural network that converts UDFs into pseudo-SDFs, enabling traditional meshing methods to extract surfaces accurately.
- The authors demonstrate superior performance over state-of-the-art techniques on datasets like ABC, with a median Chamfer Distance of 2.33×10^-5 at high resolutions.
- The approach effectively integrates UDF values and gradients, resulting in robust surface reconstruction for noisy and non-watertight data.
Neural Surface Detection for Unsigned Distance Fields
The paper "Neural Surface Detection for Unsigned Distance Fields" by Stella et al. tackles the problem of surface extraction from Unsigned Distance Fields (UDFs) using deep learning techniques. Unlike Signed Distance Fields (SDFs) that naturally facilitate surface detection through sign changes, UDFs lack this property, making conventional algorithms like Marching Cubes (MC) and Dual Contouring (DC) less effective. The authors present a neural network-based approach to convert UDFs into pseudo-SDFs, which allows the use of standard meshing algorithms.
Overview and Problem Statement
Extracting surfaces from three-dimensional fields is crucial for various applications in graphics and vision. While SDFs are well-studied and commonly used for this purpose, UDFs are better suited for representing non-watertight surfaces, such as garments or open shapes. Traditional methods to triangulate UDFs either rely on hand-crafted rules or neural networks that ignore important gradient information. These methods frequently introduce artifacts such as holes or biased structures. The paper aims to address these shortcomings by leveraging a deep learning method that integrates both UDF values and their gradients for more accurate surface detection.
Methodology
The core of the proposed method is a multi-layer perceptron (MLP) that classifies each cell in the UDF into one of 128 possible pseudo-sign configurations. The network is trained using ground-truth SDFs but can generalize to non-watertight surfaces due to its local cell-wise decision-making process. The training is augmented with Gaussian noise to improve robustness.
The process begins with a voxel grid of UDF values and gradients. The MLP then estimates pseudo-signs at the cell corners, transforming the UDF into a pseudo-SDF that is meshed using MC. Additionally, the authors integrate this approach with DualMeshUDF, a variant of DC, to enhance its performance, particularly in noisy or complex settings.
Experimental Evaluation
The authors conduct extensive evaluations on the ABC, MGN, and ShapeNet-Cars datasets. Their method consistently outperforms state-of-the-art algorithms including CAP-UDF, MeshUDF, and DCUDF across different resolutions and scenarios. Specifically, the proposed approach yields more accurate meshes and fewer artifacts such as spurious faces or holes. For instance, their method achieves a median Chamfer Distance (CD) of 2.33×10-5 on ABC at a resolution of 256, compared to 2.42×10-5 by MeshUDF.
When applied to neural UDFs obtained from auto-decoders, the method demonstrates robustness to noise, outperforming existing methods particularly at lower resolutions. This robustness is attributed to the noise augmentation strategy employed during training.
Implications and Future Work
Practically, this work facilitates the use of UDFs in applications requiring high-quality surface reconstruction without manual parameter tuning. Theoretically, it underscores the importance of incorporating gradient information in surface detection from implicit fields. The proposed method is scalable, adaptation-friendly, and capable of handling various geometric complexities.
Future research could focus on tighter integration with DualMeshUDF by reformulating its quadratic optimization to better leverage the pseudo-sign information. Addressing the inherent noisiness of gradients around the surface in UDF-based methods is another promising direction.
Conclusion
Stella et al. introduce a robust and efficient approach for neural surface detection in UDFs, bridging the gap between implicit and explicit surface representations. Their method, devoid of hand-crafted rules, shows significant improvements over existing techniques, both in terms of accuracy and generalizability. By turning UDFs into pseudo-SDFs, they open new avenues for utilizing UDFs in broader applications, setting a solid foundation for future developments in this area.