Not all views are equal: active neural radiance fields
Bojja, Bharat
This item is only available for download by members of the University of Illinois community. Students, faculty, and staff at the U of I may log in with your NetID and password to view the item. If you are trying to access an Illinois-restricted dissertation or thesis, you can request a copy through your library's Inter-Library Loan office or purchase a copy directly from ProQuest.
Permalink
https://hdl.handle.net/2142/114021
Description
Title
Not all views are equal: active neural radiance fields
Author(s)
Bojja, Bharat
Issue Date
2021-12-09
Director of Research (if dissertation) or Advisor (if thesis)
Wang, Yuxiong
Department of Study
Computer Science
Discipline
Computer Science
Degree Granting Institution
University of Illinois at Urbana-Champaign
Degree Name
M.S.
Degree Level
Thesis
Keyword(s)
neural radiance fields
active learning
deep learning
computer vision
Abstract
By modelling complex scenes via a continuous volumetric scene function, neural radiance fields (NeRF) achieves state-of-the-art results in novel view synthesis. However, since NeRF does not take into account prior information about the object content in the scene, it requires many image views and significant training time to achieve high-quality view synthesis. To reduce the number of views required, some follow-up works build on the NeRF model to take into account depth and appearance information. In this paper, we show that NeRF does not necessarily require many image views to achieve quality view synthesis if there is a way to identify the views that are most informative. There is currently no method that actively identifies how informative each view is to the eventual scene reconstruction. This is especially important when the dataset is extremely large, and it would be computationally expensive to train the model on all the images. We propose a reinforcement learning framework using the REINFORCE algorithm to actively select which viewpoints yield in the best view synthesis. We also investigate the Monte-Carlo Tree Search method as a potentially promising approach. Our methods demonstrates that it is possible achieve a dramatic improvement in performance by actively selecting a limited number of views.
Use this login method if you
don't
have an
@illinois.edu
email address.
(Oops, I do have one)
IDEALS migrated to a new platform on June 23, 2022. If you created
your account prior to this date, you will have to reset your password
using the forgot-password link below.