Graph-based stereo correspondence algorithms and occlusion detection

Abstract

Stereo matching, or 3D reconstruction from two or multiple 2D images of a scene, is one of the most active research areas in computer vision. Due to its ill-posed nature, stereo matching problem still remains challenging although many stereo algorithms have been developed. Amongst recent algorithms, graph-based methods are the most popular because of their excellent performance. On the other hand, occlusion detection is an important part in stereo problem. Many heuristics have been presented to handle occlusions. In this dissertation, three stereo algorithms are presented. The first one is a local occlusion detection algorithm, so called the competitive approach. The objective of this algorithm is to label occlusions in the obtained dense disparity map without occlusions after matching. This algorithm has two main steps: first, all matching conflicts are labelled in a dense disparity map after matching without occlusions; second, occlusions are extracted from matching conflicts via a maximum-value rule. In this dissertation, an improved cooperative algorithm is presented to improve the original cooperative algorithm by using the competitive occlusion detection approach. This algorithm is implemented on some standard test image pairs to demonstrate its improvement. The second part of this dissertation proposes two global stereo algorithms, named left-right consistency graph cuts (LRC-GC) algorithms, to solve correspondence problem with occlusions. Our objective of this part is to design good stereo matchers for accurate correspondence computation including occlusion labelling. These two algorithms are based on graph cuts and handle occlusions within the global energy minimization process by enforcing the left-right matching consistency constraint, which states that only the non-occluded pixels have correspondences and each pair of correspondences should have the same disparity value. The two novel algorithms give unique and symmetric results. We compare them with another graph-based stereo algorithm, which handles occlusions by enforcing the uniqueness constraint. Experimental results are reported to demonstrate the excellent performance of the two novel algorithms and to show that they outperform the algorithm based on the uniqueness constraint. In the conclusion part, some limitations of these novel algorithms are discussed and some directions for further research are suggested.

Date of Award	15 Jul 2005
Original language	English
Awarding Institution	City University of Hong Kong
Supervisor	Shiu Yin Kelvin YUEN (Supervisor)