The paper proposes a method for learning Unsigned Distance Functions (UDFs) that improves the fidelity of the obtained Neural UDF to the original 3D surface. The key idea is to concentrate the learning effort of the Neural UDF on surface edges.
To detect surface edges, the authors propose a new statistical method based on the calculation of a p-value at each point on the surface. This method is shown to detect surface edges more accurately than a commonly used local geometric descriptor.
The authors first describe the problem of encoding 3D shapes using implicit distance functions, and the DeepSDF model for learning a latent representation of 3D surfaces. They then introduce their edge detection method, which projects the neighboring points of a surface point onto the average plane and performs a central symmetry test on the projected points. The p-value of this test is used as a local descriptor to identify surface edges.
The authors then explain how this edge detection method is used to improve the training of a Neural UDF. By sampling more training points around surface edges, the local accuracy of the trained Neural UDF is improved, leading to better global expressiveness in terms of Hausdorff distance.
Finally, the authors present results on edge detection and the application to UDF learning, showing the benefits of their approach.
To Another Language
from source content
arxiv.org
Key Insights Distilled From
by Virgile Foy ... at arxiv.org 05-07-2024
https://arxiv.org/pdf/2405.03381.pdfDeeper Inquiries