Off-campus UMass Amherst users: To download campus access dissertations, please use the following link to log into our proxy server with your UMass Amherst user name and password.

Non-UMass Amherst users: Please talk to your librarian about requesting this dissertation through interlibrary loan.

Dissertations that have an embargo placed on them will not be available to anyone until the embargo expires.

Author ORCID Identifier

https://orcid.org/0000-0002-7492-7808

AccessType

Open Access Dissertation

Document Type

dissertation

Degree Name

Doctor of Philosophy (PhD)

Degree Program

Computer Science

Year Degree Awarded

2022

Month Degree Awarded

September

First Advisor

Evangelos Kalogerakis

Second Advisor

Subhransu Maji

Abstract

The ability to parse 3D objects into their constituent parts is essential for humans to understand and interact with the surrounding world. Imparting this skill in machines is important for various computer graphics, computer vision, and robotics tasks. Machines endowed with this skill can better interact with its surroundings, perform shape editing, texturing, recomposing, tracking, and animation. In this thesis, we ask two questions. First, how can machines decompose 3D shapes into their fundamental parts? Second, does the ability to decompose the 3D shape into these parts help learn useful 3D shape representations?

In this thesis, we focus on parsing the shape into compact representations, such as parametric surface patches and Constructive Solid Geometry (CSG) primitives, which are also widely used representations in 3D modeling in computer graphics. Inspired by the advances in neural networks for 3D shape processing, we develop neural network approaches to tackle shape decomposition. First, we present CSGNet, a network architecture to parse shapes into CSG programs, which is trained using combination of supervised and reinforcement learning. Second, we present ParSeNet, a network architecture to decompose a shape into parametric surface patches (B-Spline) and geometric primitives (plane, cone, cylinder and sphere), trained on a large set of CAD models using supervised learning.

The training of deep neural network architectures for 3D recognition and generation tasks requires a large amount of labeled datasets. We explore ways to alleviate this problem by relying on shape decomposition methods to guide the learning process. Towards that end, we first study the use of freely available metadata, albeit inconsistent, from shape repositories to learn 3D shape features. Later we show that learning to decompose a 3D shape into geometric primitives also helps in learning shape representations useful for semantic segmentation tasks. Finally, since most 3D shapes encountered in real life are textured, consisting of several fine-grained semantic parts, we propose a method to learn fine-grained representations for textured 3D shapes in a self-supervised manner by incorporating 3D geometric priors.

DOI

https://doi.org/10.7275/30570375

Creative Commons License

Creative Commons Attribution 4.0 License
This work is licensed under a Creative Commons Attribution 4.0 License.

Share

COinS