Date Awarded

2017

Document Type

Dissertation

Degree Name

Doctor of Philosophy (Ph.D.)

Department

Computer Science

Advisor

Pieter Peers

Committee Member

Weizhen Mao

Committee Member

Peter Kemper

Committee Member

Robert Michael Lewis

Committee Member

Xing Tong

Abstract

In computer vision and computer graphics, a photograph is often considered a photometric representation of a scene. However, for most camera models, the relation between recorded pixel value and the amount of light received on the sensor is not linear. This non-linear relationship is modeled by the camera response function which maps the scene radiance to the image brightness. This non-linear transformation is unknown, and it can only be recovered via a rigorous radiometric calibration process. Classic radiometric calibration methods typically estimate a camera response function from an exposure stack (i.e., an image sequence captured with different exposures from the same viewpoint and time). However, for photographs in large image collections for which we do not have control over the capture process, traditional radiometric calibration methods cannot be applied. This thesis details two novel data-driven radiometric photo-linearization methods suit- able for photographs captured with unknown camera settings and under uncontrolled conditions. First, a novel example-based radiometric linearization method is pro- posed, that takes as input a radiometrically linear photograph of a scene (i.e., exemplar), and a standard (radiometrically uncalibrated) image of the same scene potentially from a different viewpoint and/or under different lighting, and which produces a radiometrically linear version of the latter. Key to this method is the observation that for many patches, their change in appearance (from different viewpoints and lighting) forms a 1D linear subspace. This observation allows the problem to be reformulated in a form similar to classic radiometric calibration from an exposure stack. In addition, practical solutions are proposed to automatically select and align the best matching patches/correspondences between the two photographs, and to robustly reject outliers/unreliable matches. Second, CRF-net (or Camera Response Function net), a robust single image radiometric calibration method based on convolutional neural net- works (CNNs) is presented. The proposed network takes as input a single photograph, and outputs an estimate of the camera response function in the form of the 11 PCA coefficients for the EMoR camera response model. CRF-net is able to accurately recover the camera response function from a single photograph under a wide range of conditions.

DOI

http://dx.doi.org/doi:10.21220/S25Q2M

Rights

© The Author

Share

COinS