Differentiable volume rendering using signed distance functions
Name
1145123039-MIT.pdf
Size
3.98 MB
Format
Adobe PDF
Checksum (MD5)
5d1a92ba532bda1d1a1912bf49b7f034
Author(s)
Kaza, Srinivas,M. Eng.Massachusetts Institute of Technology.
Advisor(s)
Fredo Durand.
Date Issued
2019
Publisher
Massachusetts Institute of Technology
Abstract
Gradient-based methods are often used in a computer graphics and computer vision context to solve inverse rendering problems. These methods can be used to infer camera parameters, material properties, and even object pose and geometry from 2D images. One of the challenges that faces differentiable rendering systems is handling visibility terms in the rendering equation, which are not continuous on object boundaries. We present a renderer that solves this problem by introducing a form of visibility that is not discontinuous, and thus can be differentiated. This "soft visibility" is inspired by volumetric rendering, and is facilitated by our decision to represent geometry within the scene as a signed distance function. We also present methods for performing gradient descent upon distance fields while preserving Lipschitz continuity. Unlike most differentiable mesh-based renderers, our renderer can optimize between geometry of different homeomorphism classes in a variety of image-based shape fitting tasks.
Description
This electronic version was submitted by the student author. The certified thesis is available in the Institute Archives and Special Collections.
Thesis: M. Eng., Massachusetts Institute of Technology, Department of Electrical Engineering and Computer Science, 2019
Cataloged from student-submitted PDF version of thesis.
Includes bibliographical references (pages 55-58).
Subjects
Electrical Engineering and Computer Science.
MIT Department
Massachusetts Institute of Technology. Department of Electrical Engineering and Computer Science
Terms of Use
MIT theses are protected by copyright. They may be viewed, downloaded, or printed from this source but further reproduction or distribution in any format is prohibited without written permission.
Persistent DSpace Link
