Statistical methods for correspondence problems with branched structures.
Thesis DisciplineComputer Science
Degree GrantorUniversity of Canterbury
Degree NameDoctor of Philosophy
Correspondence problems are among the most studied topics in computer vision. They consist of two main steps: understanding the information present in images and associating features that represent the same information, between different images. Solving correspondence problems is a fundamental task in many applications, such as recovering a three dimensional structure from images. Correspondence problems become even more challenging when scenes show self-similar structures or have a high level of ambiguity. The aim of my research is to solve such problems in the context of self-similar branched structures, such as vine structures. The solution will be then implemented on an automated pruning robot. Several methods to solve correspondence problems have already been proposed and they proved very effective with different scenes. However, when they are applied to vine images, because of the structure of the plants, their performances are reduced. I review and analyse some of those methods in Chapter 2. A model-based approach proved effective in the present context. It enables decreasing the level of ambiguity of matches and it avoids making too many assumptions about the scene that could be violated with vine structures. First, in Chapter 3 I define and analyse variables describing branched structures. Several algorithms are proposed and compared. Three state-of-the-art methods, RANSAC, graph matching and maximum likelihood, are suitably modified to be applied to my case. Those are then compared to the algorithm used in the initial state of the project and with three novel algorithms I proposed. These newly proposed methods compute the probabilities of a correspondence being correct given some variables of the branches. My main approach represents an improvement of 90% over prior research in terms of branches reconstructed, thus providing a more complete and accurate reconstructions. Moreover, it has better performance than state-of-the-art algorithms applied to this specific problem.