Meshing Neural Unsigned Distance Fields


Event details

Date 12.07.2022 10:3012:30  
Speaker Federico Stella
Category Conferences - Seminars
EDIC candidacy exam
Exam president: Prof. Amir Zamir
Thesis advisor: Prof. Pascal Fua
Co-examiner: Prof. Nicolas Flammarion

Reconstructing shapes from noisy or partial input data is a fundamental problem with several applications in different fields such as robotics, graphics, medicine, architecture and virtual reality. Recently, implicit representations parametrized by neural networks and based on Signed Distance Fields or on Occupancy Fields have emerged as a very powerful tool to represent, manipulate and generate watertight shapes without being constrained by resolution or by a fixed topology. Explicit meshes can be obtained from them by using isosurface extraction algorithms such as the popular Marching Cubes. However, a major limitation of the existing approaches based on implicit representations is that they can only work with watertight surfaces with a clear concept of inside and outside, whereas real world data usually comes from 3D scans and needs to be pre-processed in order to be made watertight, losing the inner details in the process, for example the inside of a car. Moreover, 3D shapes like garments could be more naturally parametrized as open surfaces, allowing their use in applications that require single-sided surfaces, such as draping simulators. For these reasons implicit representations based on Unsigned Distance Fields have emerged, for which classical meshing algorithms do not directly work, posing additional challenges in obtaining open meshes as outputs.
In this work, we select three papers on the topics of geometry learning, neural unsigned distance fields approximation and isosurface extraction, highlighting the strengths and shortcomings of current methods and identifying interesting research directions to generate high-quality open meshes from learned (and thus imperfect) Unsigned Distance Fields.

Background papers
- "A survey on deep geometry learning: From a representation perspective” by Yun-Peng Xiao, Yu-Kun Lai, Fang-Lue Zhang, Chunpeng Li and Lin Gao.
- "Neural Unsigned Distance Fields for Implicit Function Learning” by Julian Chibane, Aymen Mir and Gerard Pons-Moll.
- “Neural Dual Contouring” by Zhiqin Chen, A. Tagliasacchi, T. Funkhouser and Hao Zhang.

Practical information

  • General public
  • Free


EDIC candidacy exam