[CS] Richard Liu MS PresentationApr 11, 2025

[via cs]

This is an announcement of Richard Liu's MS Presentation
===============================================
Candidate: Richard Liu

Date: Friday, April 11, 2025

Time:  1 pm CST

Location: JCL 298

Title: 3D Self-Supervised Learning through Geometry Processing

Abstract: Many tasks in geometry processing are often lacking in labelled training data, well-defined objectives, or both. Local mesh parameterization is one such task in which there are two contrasting objectives of patch size and distortion minimization and no training data available. Instead of balancing the objectives heuristically as in prior algorithms, I propose instead to learn the balance through unlabelled data and objective functions grounded in geometry processing. I adapt a traditional mesh parameterization technique as a differentiable layer in a neural pipeline, and train a neural network to predict local point-conditioned surface patches using energy functions derived from geometry processing theory. Once trained, the network predicts user-conditioned patches and parameterizations at real-time rates. 

Advisors: Rana Hanocka

Committee Members: Rana Hanocka, Michael Maire, and Greg Shakhnarovich