Aristotle University Of Thessaloniki (AUTH)
Αντιστοίχιση σημείων με γεωμετρικούς περιορισμούς
Abstract
dc:descriptionIn this thesis we present and analyse the theoretical and experimental steps of a methodology for the evaluation of the geometric consistency of point-pair matches in weakly calibrated scenes. Given interest points that are extracted by a local feature algorithm and putative matches among them according to some local photometric measures, we attempt their verification using purely geometric characteristics in projective spaces. The initial approach is done in a weakly calibrated framework, which means that some information for the camera pair is available, but no assumptions are made for the observed point configurations in 3-dimensional space. During the course we comment on the theoretical and practical problems of the mathematical analysis, the experimental approach as well as the implementation of the proposed algorithms. The central idea of the proposed methodology utilises a virtual geometric construction through which a scalar quantity is calculated, which is subsequently evaluated statistically. The method is based on the cross-ratios which are projective invariants, in the sense that they are preserved by the projective transformations of a projective space. We describe the calculation of the proposed quantities through the virtual geometric construction and analyse their probability distribution under both assumptions of correct and false point correspondences. Main guideline in the whole approach has been to steer away as much as possible from the theoretical barrier that the absence of a metric in projective spaces imposes on calculations performed inside them. This is attempted by reducing the indexing of point matches as inliers and outliers in weekly calibrated scenarios into a problem of indexing probability functions characterising the proposed scalar quantities. The classification is binary and does not rely on the tuning of any thresholding variables. We study theoretically as well as by means of extensive experiments the distinction capability of the distributions as a criterion for the point matching problem and its dependence on the number of samples used for their estimation. This number is shown to be directly linked to the computational complexity of the proposed methodology and the accuracy of the results. At the same time, the indexing performance of the approach in terms of false positives is consistently high. That kind of behaviour is suited for different applications, depending on the available computational power and the required accuracy. Subsequently, we introduce techniques aimed at reducing the computational cost of the proposed methods, as well as for implementing them in parallel computing environments. The scalar values based on the cross-ratios are processed in a statistical framework through the use of weighted sampling. Our approach applies on the problem of geometric point-pair validation a novel modification on the voting based methods, in which the values of the cost function are evaluated according to their distributions instead of their sum. Finally, we also present an application of our approach in uncalibrated cameras using a robust RANSAC framework. Additionally time, this application is a partial extension of the cross-ratio based methodology in uncalibrated scenarios. The performance of the proposed schemes is compared to the commonly used algebraic distances and shown to be more strict by extensive experiments on synthetic and real data.
Degree
thesis:*- Grantor dc:publisher
- Aristotle University Of Thessaloniki (AUTH)
- Year dc:date
- 2011
Author and committee
dc:creator, dc:contributor.*- Authors dc:creator
-
- Papachristou, Christos
- Παπαχρήστου, Χρήστος
Subjects
dc:subject × 12Rights
- Language dc:language
- gre
Identifiers
dc:identifier.*- Identifier
- 10.12681/eadd/28493