The Fisher-Rao metric and its applications in computer vision

old_uid3898
titleThe Fisher-Rao metric and its applications in computer vision
start_date2008/01/24
schedule14h
onlineno
location_infocouloir 55-65, salle 211
summaryMany computer vision algorithms involve measurements of distance. For example, in a clustering algorithm it is of interest to measure the distances between the different clusters. The measurement of distance is made more complicated because the form of the data often depends on an arbitrary choice of a coordinate system by the computer vision practitioner. Under these circumstances it is necessary for the measurements to be independent of the choice of coordinate system. This requirement places a severe restriction on the choice of measurement. If, in the above example, the clusters are modeled by probability density functions, then a natural measurement of the distance between two probability density functions is the Kullback-Leibler divergence, which has the essential property that it is invariant under changes of coordinate system. If the probability density functions belong to a parameterised family of densities then the Kullback Leibler distance defines a Riemannian metric on the parameter space. This metric is known as the Fisher-Rao metric. The Kullback-Leibler distance is widely used in computer vision, but there have until now been very few applications of the Fisher-Rao metric. It is shown how the Fisher-Rao metric is the basis of a new approach to structure detection in which the parameter space for the structures in question is sampled at a finite number of points and each point is tested in turn to see if the presence of the corresponding structure is supported by the measurements. The sample points are chosen such that every point in the parameter space is near, under the Fisher-Rao metric, to at least one sample point. The Fisher-Rao metric is used to obtain an estimate of the number of sample points that are required. The applications of this method to the detection of lines and ellipses are described.
oncancelchangement de lieu
responsiblesClady