Electronic Thesis and Dissertation Repository

Degree

Doctor of Philosophy

Program

Computer Science

Supervisor

Dr. Olga Veksler

Abstract

Computer graphics and computer vision were traditionally two distinct research fields focusing on opposite topics. Lately, they have been increasingly borrowing ideas and tools from each other. In this thesis, we investigate two problems in computer vision and graphics that rely on the same tool, namely energy optimization with graph cuts.

In the area of computer graphics, we address the problem of generating artificial classic mosaics, still and animated. The main purpose of artificial mosaics is to help a user to create digital art. First we reformulate our previous static mosaic work in a more principled global optimization framework. Then, relying on our still mosaic algorithm, we develop a method for producing animated mosaics directly from real video sequences, which is the first such method, we believe. Our mosaic animation style is uniquely expressive. Our method estimates the motion of the pixels in the video, renders the frames with mosaic effect based on both the colour and motion information from the input video. This algorithm relies extensively on our novel motion segmentation approach, which is a computer vision problem.

To improve the quality of our animated mosaics, we need to improve the motion segmentation algorithm. Since motion and stereo problems have a similar setup, we start with the problem of finding visual correspondence for stereo, which has the advantage of having datasets with ground truth, useful for evaluation. Most previous methods for stereo correspondence do not provide any measure of reliability in their estimates. We aim to find the regions for which correspondence can be determined reliably. Our main idea is to find corresponding regions that have a sufficiently strong texture cue on the boundary, since texture is a reliable cue for matching. Unlike the previous work, we allow the disparity range within each such region to vary smoothly, instead of being constant. This produces blob-like semi-dense visual features for which we have a high confidence in their estimated ranges of disparities.


Share

COinS