Jake Austin

EECS Department, University of California, Berkeley

Technical Report No. UCB/EECS-2024-104

May 15, 2024

http://www2.eecs.berkeley.edu/Pubs/TechRpts/2024/EECS-2024-104.pdf

3D scene reconstruction for static scenes is a challenging and difficult prob- lem. There has been a rise in highly unconstrained neural 3D representations like NeRF which use neural networks as black box functions to model the den- sity and color of the scene, with no constraints on geometry or feasibility. While the expressive nature of NeRFs can lead to high visual fidelity, it can also lead to overfitting and poor geometry. In this work, we detail a number of avenues for attempting to constrain the expressive nature of modern 3D reconstruction methods, through decomposing the scene into simpler basic components. We experiment with novel representations and seek to exploit the inductive bias that the world around us is simply a composition of basic shapes and parts, and examine the pros and cons of such a strong inductive bias. We first attempt to regularize geometry by constraining not the individual shapes, but rather constrain that the shapes be reusable, experimenting with neural field repre- sentations. We then explicitly regularize geometry as being a composition of simple parts and simple geometric shapes.

Advisors: Angjoo Kanazawa

---

BibTeX citation:

@mastersthesis{Austin:EECS-2024-104,
    Author= {Austin, Jake},
    Title= {3D Scene Part Decomposition},
    School= {EECS Department, University of California, Berkeley},
    Year= {2024},
    Month= {May},
    Url= {http://www2.eecs.berkeley.edu/Pubs/TechRpts/2024/EECS-2024-104.html},
    Number= {UCB/EECS-2024-104},
    Abstract= {3D scene reconstruction for static scenes is a challenging and difficult prob- lem. There has been a rise in highly unconstrained neural 3D representations like NeRF which use neural networks as black box functions to model the den- sity and color of the scene, with no constraints on geometry or feasibility. While the expressive nature of NeRFs can lead to high visual fidelity, it can also lead to overfitting and poor geometry. In this work, we detail a number of avenues for attempting to constrain the expressive nature of modern 3D reconstruction methods, through decomposing the scene into simpler basic components. We experiment with novel representations and seek to exploit the inductive bias that the world around us is simply a composition of basic shapes and parts, and examine the pros and cons of such a strong inductive bias. We first attempt to regularize geometry by constraining not the individual shapes, but rather constrain that the shapes be reusable, experimenting with neural field repre- sentations. We then explicitly regularize geometry as being a composition of simple parts and simple geometric shapes.},
}

EndNote citation:

%0 Thesis
%A Austin, Jake 
%T 3D Scene Part Decomposition
%I EECS Department, University of California, Berkeley
%D 2024
%8 May 15
%@ UCB/EECS-2024-104
%U http://www2.eecs.berkeley.edu/Pubs/TechRpts/2024/EECS-2024-104.html
%F Austin:EECS-2024-104