Citation Context

...es will be scattered, but the true correspondences will cluster on the epipolar curve. The epipolar curve can thus be determined. The most similar previous work is Triggs' joint feature distributions =-=[12]-=-, which may be regarded as a parametric (in the statistical sense) analogue of our nonparametric model. The major advance of our work here is a simple robust algorithm to estimate the joint distributi...

Citation Context

...ch is to use Monte Carlo methods, which is highly general but computationally expensive. Alternatively, analytical frameworks have been developed by Kanatani [1], Förstner [2], and others (e.g., [3], =-=[4]-=-). However, when these mappings are well approximated locally by an affine transform, a first-order model has proved to be sufficient [5]. This linearized approximation of the error model is commonly ...

Citation Context

...pondences and to use them in the RANSAC step. We use weak motion models (WMMs) to estimate these probabilities which are more informative than the correlation scores used for this purpose in [16]. In =-=[21]-=- a probabilistic motion model was presented for finding additional matches after an estimate for the fundamental matrix was found, limiting the motion to a segment of the epipolar line based on the mo...

Citation Context

... number features in one image that have no correspondence in the other image. We do conditional characterization of the probability space of correspondences based on Joint Feature Distributions (JFD) =-=[21]-=-. We seek to maximize the probabilistic support of the putative correspondence set over a number of MCMC iterations, guided by proposal distributions based on similarity or JFD. Similarity based guida...

Citation Context

...ir proposed modifications cover the propagation of first and second order moments neglecting modifications of the underlying distributions. Weise & Wöger (1999) instead propagate distributions providing a framework that is more generally applicable. The recent GUM Supplement 1 document (GUMS1, JCGM (2008b)) makes explicit the Bayesian foundations of the GUM. Thus removing inconsistencies in the information fusion of the guide. Different approaches to the treatment of uncertainty in the domain of computer vision have been reported in the literature (e.g. Havelock (1989); Kanungo et al. (1995); Triggs (2001); Ochoa & Belongie (2006)). Using the central limit theorem (CLT, Papoulis & Pillai (2002)) as a key argument, the use of the Gaussian assumption is suggested – and often validated – by many researchers. Heuel (2003) and Criminisi (2001) show that Gaussian densities can be applied to represent homogeneous geometric entities. In an earlier work (Brandner (2006)) we apply first order propagation of Gaussian quantities to a vision-based tracking application. Based on the exclusive use of Gaussian densities to describe both physical quantities and prior information, the Bayesian extensions of the ...

Citation Context

...ir proposed modifications cover the propagation of first and second order moments neglecting modifications of the underlying distributions. Weise & Wöger (1999) instead propagate distributions providing a framework that is more generally applicable. The recent GUM Supplement 1 document (GUMS1, JCGM (2008b)) makes explicit the Bayesian foundations of the GUM. Thus removing inconsistencies in the information fusion of the guide. Different approaches to the treatment of uncertainty in the domain of computer vision have been reported in the literature (e.g. Havelock (1989); Kanungo et al. (1995); Triggs (2001); Ochoa & Belongie (2006)). Using the central limit theorem (CLT, Papoulis & Pillai (2002)) as a key argument, the use of the Gaussian assumption is suggested – and often validated – by many researchers. Heuel (2003) and Criminisi (2001) show that Gaussian densities can be applied to represent homogeneous geometric entities. In an earlier work (Brandner (2006)) we apply first order propagation of Gaussian quantities to a vision-based tracking application. Based on the exclusive use of Gaussian densities to describe both physical quantities and prior information, the Bayesian extensions of the ...

Citation Context

... correct, although in practice the method correctly converges even with less matches. In our experiments the transformation is either a homography or a fundamental matrix. A model selection algorithm =-=[56, 125]-=- can be used to automatically decide which transformation is the most appropriate one. 6.1.6 Retrieval algorithm In the following we outline the image retrieval algorithm. The recognition process cons...