2026-09-20T12:24:49Zhttps://ecommons.cornell.edu/server/oai/requestoai:ecommons.cornell.edu:1813/1148122026-07-07T15:02:19Zcom_1813_35col_1813_47
Yang, Guandao
Belongie, Serge
Hariharan, Bharath
De Sa, Christopher
Snavely, Keith
2024-04-05T18:48:38Z
2023-08
ProQuest Submission ID: 13719
ProQuest Publication ID: 30526098
https://hdl.handle.net/1813/114812
https://doi.org/10.7298/nxrc-ps68
16219181
196 pages
Geometry processing consists of algorithms for shape creation, manipulation, and analysis. These algorithms are cornerstones for the creation of almost every digital shapes. These shapes are consumed in our everyday lives, from digital contents to manufactured objects. These algorithms are indispensable parts of our every life. Traditionally, geometry processing algorithms are usually done with shape represented in explicit representations such as meshes, voxels, and their variants. These shape representations are easy to interpret by human, but they are difficult to be handled by automatic algorithms. For example, voxels can take up a lots of memory and meshes might need careful handle of the discretization quality during the processing. These challenges make development of geometry processing algorithms difficult and inefficient. In this thesis, we propose to use an alternative shape representation for the geometry processing pipeline. Specifically, each shape is represented by a neural network that takes a spatial coordinate as input and outputs a scalar or vector value. We coin such representation neural fields. Neural field can circumvent the common challenges of meshes and voxels because it doesn’t explicitly discretize the space and can be stored compactly. Moreover, it has the unique advantage of being easy to optimize in deep-learning frameworks, which makes them suitable for data-driven methods. This thesis demonstrates how to build a geometry processing pipeline using neural fields. Such a pipeline can improve shape creation efficiency, democratize the 3D assets creation process, and revolutionize the digital shape creation paradigm.
en
Artificial Intelligence
Computer Graphics
Computer Vision
Geometry Processing
Machine Learning
Shape Analysis
Geometry Processing with Neural Fields
dissertation or thesis
https://newcatalog.library.cornell.edu/catalog/16219181
application/pdf
Computer Science
Cornell University
Doctor of Philosophy
Ph. D., Computer Science
https://hdl.handle.net/1813/59810.2
Publication
Cornell Theses and Dissertations
Yang, Guandao
5893a6ea-7af3-41d7-abc6-04bcd26ab5df
COMPLETE
open.access
open.access
<Publication xmlns="https://www.openaire.eu/cerif-profile/1.1/" id="00249ba0-3764-4984-ac3e-b8c99eb5ea3e">
<Type xmlns="https://www.openaire.eu/cerif-profile/vocab/COAR_Publication_Types">http://purl.org/coar/resource_type/c_1843</Type>
<Language>en</Language>
<Title>Geometry Processing with Neural Fields</Title>
<PublishedIn>
<Publication>
</Publication>
</PublishedIn>
<PublicationDate>2023-08</PublicationDate>
<DOI>https://doi.org/10.7298/nxrc-ps68</DOI>
<Authors>
<Author>
<DisplayName>Yang, Guandao</DisplayName>
<Affiliation>
<OrgUnit>
</OrgUnit>
</Affiliation>
</Author>
</Authors>
<Editors>
</Editors>
<Publishers>
<Publisher>
<OrgUnit />
</Publisher>
</Publishers>
<Keyword>Artificial Intelligence</Keyword>
<Keyword>Computer Graphics</Keyword>
<Keyword>Computer Vision</Keyword>
<Keyword>Geometry Processing</Keyword>
<Keyword>Machine Learning</Keyword>
<Keyword>Shape Analysis</Keyword>
<Abstract>Geometry processing consists of algorithms for shape creation, manipulation, and analysis. These algorithms are cornerstones for the creation of almost every digital shapes. These shapes are consumed in our everyday lives, from digital contents to manufactured objects. These algorithms are indispensable parts of our every life. Traditionally, geometry processing algorithms are usually done with shape represented in explicit representations such as meshes, voxels, and their variants. These shape representations are easy to interpret by human, but they are difficult to be handled by automatic algorithms. For example, voxels can take up a lots of memory and meshes might need careful handle of the discretization quality during the processing. These challenges make development of geometry processing algorithms difficult and inefficient. In this thesis, we propose to use an alternative shape representation for the geometry processing pipeline. Specifically, each shape is represented by a neural network that takes a spatial coordinate as input and outputs a scalar or vector value. We coin such representation neural fields. Neural field can circumvent the common challenges of meshes and voxels because it doesn’t explicitly discretize the space and can be stored compactly. Moreover, it has the unique advantage of being easy to optimize in deep-learning frameworks, which makes them suitable for data-driven methods. This thesis demonstrates how to build a geometry processing pipeline using neural fields. Such a pipeline can improve shape creation efficiency, democratize the 3D assets creation process, and revolutionize the digital shape creation paradigm.</Abstract>
<Access xmlns="http://purl.org/coar/access_right"
>
</Access>
</Publication>