2010 Apr 01 at 10:00
DC 2314
Jingyuan Huang, graduate student, David R. Cheriton School of Comp. Sci., Univ. Waterloo
This thesis proposes a method for adding and changing lighting for a subset of images. The process requires only a single input, either a studio photograph or a synthetic image consisting of several simple objects placed on a uniformly coloured background. Based on 2D information (contours, shadows, specular areas) extracted from the input, the method reconstructs 3D and 2$\frac{1}{2}$D models for the original lighting and shape. It then modifies the appearance of shading and shadows to achieve relighting. It can produce visually satisfactory results without a full 3D description of the scene geometry, and requires minimal user assistance.
While developing this method, the author considered the importance of a variety of cues, such as contours and shadows, for understanding 3D geometry. The author also explored constraints like symmetry that help determine surface shapes. This method has potential applications in improving the appearance of existing photographs. It can also be used in image compositing to achieve consistent lighting.