Abstract
dc:descriptionThis thesis shows a relighting method that can relight a scene from a single image of that scene. It is well established that multiple distinctly lit images of a scene yield a light transport matrix that maps illuminants to shading fields and so to relit images. The work presented in this theses uses novel theory which says that one can use images of similar scenes to estimate the different lightings that apply to a given scene, with bounded expected error and explores various ways of selecting those images of similar scenes. This theory yields losses to train a relighting method using only unpaired images of scenes – multiple view data, Computer Generated Imagery (CGI) data or multiple relight data is not required for training. The method learns to map an image to a scene-specific relighting model consisting of a shading estimate, a light transport matrix, and a probability distribution over illuminants. The resulting model allows us to produce random relightings of a given scene that are plausible. The probability representation also allows us to produce relightings that are (a) close to a demand shading and (b) likely according to the predicted probability distribution, yielding controllable relighting. Qualitatively, the method’s relightings are easy to get for found data, but are not as good qualitatively as those obtained when multiple view or CGI data is available.
Degree
thesis:*- Name thesis:degree_name
- M.S.
- Level thesis:degree_level
- Thesis
- Discipline thesis:degree_discipline
- Computer Science
- Grantor
- University of Illinois at Urbana-Champaign
- Year dc:date
- 2022
Author and committee
dc:creator, dc:contributor.*- Author dc:creator
-
- Asthana, Pranav Kumar
- Contributors dc:contributor
-
- Forsyth, David A
Subjects
dc:subject × 5Rights
dc:rights- Statement dc:rights
-
- Copyright 2022 Pranav Asthana
- Language dc:language
- en, eng
Identifiers
dc:identifier.*- Handle dc:identifier
- https://hdl.handle.net/2142/115440