[CS] William Gao MS Presentation/Feb 14, 2025

[via cs]

This is an announcement of William Gao's MS Presentation
===============================================
Candidate: William Gao

Date: Friday, February 14, 2025

Time: 10 am CST

Location: JCL 356

Title: Neural Methods for 3D Shape Editing

Abstract: 3D content generation from scratch has emerged as a focus in computer vision research due to applications in multiple visual mediums including games, animation, and virtual reality. However, editing existing 3D models remains a challenging task. In this talk, I will introduce two paradigms for neural manipulation of 3D shapes and discuss my research in tackling their shortcomings. In the first, 3D shapes are represented as explicit meshes with vertices and edges, and editing involves optimizing vertex positions against some objective function. Recent research uses text-image foundation models for this objective in order to enable text-driven editing. Our work demonstrates that a different parametrization of these vertex deformations combined with this text-based approach allows us to represent both high-frequency details and low-frequency semantic changes in a much more realistic manner.

The second paradigm revolves around using multiview image-to-3D models to generate 3D content. We employ a new masking strategy during training that enables us to use these reconstruction models to edit in a single forward pass as opposed to the optimization processes described in the first part.

Advisor: Michael Maire

Committee Members: Rana Hanocka and Michael Maire