SPSR Parameter determined by Neural Network

Information

• Publication Type: Student Project
• Workgroup(s)/Project(s):
  • Rendering and Modeling
• Date: 2025
• Date (Start): October 2023
• Date (End): January 2025
• Matrikelnummer: 1227109
• First Supervisor: Philipp Erler
• Keywords: Deep Learning, Surface Reconstruction, Parameter Optimization, Screened Poisson Surface Reconstruction

Abstract

In ’Parameter Optimization for Surface Reconstruction’ we tested different parameter optimization methods to find the most accurate and fast way to find optimal parameters for longer-running tasks that cannot be exhaustively tested. One strategy that we did not test is using machine learning to solve this problem. If it is possible to train a neural network to determine close-to-optimal parameters for a given task, then that would certainly be faster than all the other tested solutions. For this report, we generated training data for the problem of reconstructing meshes from point clouds using Screened Poisson Surface Reconstruction. We used ParamILS for this, and then tested two different networks to see if this is achievable. We describe our strategy for this, present our results, and discuss the encountered problems.

Additional Files and Images

Weblinks

• Repo

BibTeX

@studentproject{steinschorn-2025-par,
  title =      "SPSR Parameter determined by Neural Network",
  author =     "Florian Steinschorn",
  year =       "2025",
  abstract =   "In ’Parameter Optimization for Surface Reconstruction’
               we tested different parameter optimization methods to find
               the most accurate and fast way to find optimal parameters
               for longer-running tasks that cannot be exhaustively tested.
               One strategy that we did not test is using machine learning
               to solve this problem. If it is possible to train a neural
               network to determine close-to-optimal parameters for a given
               task, then that would certainly be faster than all the other
               tested solutions. For this report, we generated training
               data for the problem of reconstructing meshes from point
               clouds using Screened Poisson Surface Reconstruction. We
               used ParamILS for this, and then tested two different
               networks to see if this is achievable. We describe our
               strategy for this, present our results, and discuss the
               encountered problems.",
  month =      jan,
  keywords =   "Deep Learning, Surface Reconstruction, Parameter
               Optimization, Screened Poisson Surface Reconstruction",
  URL =        "https://www.cg.tuwien.ac.at/research/publications/2025/steinschorn-2025-par/",
}