Example-based image relighting
Date of Issue2014
School of Computer Engineering
Centre for Multimedia and Network Technology
Synthesizing photo-realistic images under different illumination conditions remains an important but challenging problem in both computer vision and graphics communities. Over the past few years, many researchers in both communities have tried inventing new methods to capture the complex light interactions existing in nature. However, most of them either rely on expensive custom-built equipment, or only focus on a class of objects, such as faces, to make the problem tractable with their approaches. Our research focuses on an example-based relighting framework for general objects. In this thesis, a review of the representative existing works on image relighting is provided. To avoid explicit 3D model acquisition as required in geometry-based methods and tedious laboratory setups as required for existing image-based methods, an example-based framework is proposed and analyzed. In this framework, we use a database of reference objects captured under different illumination conditions. Given one or several input image(s) of a new object captured under illumination conditions present in the reference database, new images can be synthesized for the input object under novel illumination conditions. We start from a single image approach with spherical surface and Lambertian model assumptions. The analysis is extended to non-spherical surface and more general BRDF models. Finally, a multiple image approach is introduced. Given limited sample images of the input object, the example-based relighting method can synthesize realistic images under different lighting conditions. It is a general and purely data-driven method which does not require an analytical model of the surface geometry or reflectance. The framework is demonstrated on standard real image databases and our own collected data, with comparisons showing that it outperforms image-based model fitting methods. This thesis concludes with a discussion of future work for example-based relighting. These include applying the framework to more complex images using sophisticated similarity estimation methods. Besides exploring physically correct solutions, we may also consider research methods for achieving perceptually appropriate relighting.
DRNTU::Engineering::Computer science and engineering::Computing methodologies::Image processing and computer vision