Georgia Institute of Technology
Differentiable 3D Scene Representations with Point-based Neural Methods
Abstract
dc:description.abstractThis thesis explores particle-based approaches for scene representation. We introduce moving patches by fitting geometry pixels ("gexels") of moving geometry images onto the scene geometry. We propose to represent a signed distance field as a weighted sum of local functions attached to moving particles. We consider optimization frameworks to fit both 2D and 3D target shapes and propose a splitting strategy to control the density of particles. Experiments on canonical 3D models demonstrate our method’s viability to capture coarse and fine geometric structures with explainable local patches, offering a storage‑efficient alternative to current representations. Finally, we outline avenues for future work, such as continuous neural patch parametrization, adaptive domain decomposition, and interactive editing. We hope to inspire further investigation into explainable, flexible, and differentiable explicit geometry paradigms that balance fidelity, efficiency and editability.
Degree
thesis:*- Level thesis:degree_level
- Masters
- Department dc:contributor.department
- Interactive Computing
- Grantor dc:publisher
- Georgia Institute of Technology
- Year dc:date.issued
- 2025
Author and committee
dc:creator, dc:contributor.*- Author dc:creator
-
- Borcsok, Barnabas Barney
- Advisors dc:contributor.advisor
-
- Zhu, Bo
- Turk, Greg
- Committee member dc:contributor.committeemember
-
- Ha, Sehoon
Subjects
dc:subject × 4Rights
- Language dc:language.iso
- en_US
Identifiers
dc:identifier.*- Handle dc:identifier.uri
- https://hdl.handle.net/1853/77851