Rank-test similarity measure between video segments for local descriptors

Abstract

This paper presents a novel and efficient similarity measure between video segments. We consider local spatio-temporal descriptors. They are considered to be realizations of an unknown, but class-specific distribution. The similarity of two video segments is calculated by evaluating an appropriate statistical criterion issued from a rank test. It does not require any matching of the local features between the two considered video segments, and can deal with a different number of computed local features in the two segments. Furthermore, our measure is self-normalized which allows for simple cue integration, and even on-line adapted class-dependent combination of the different descriptors. Satisfactory results have been obtained on real video sequences for two motion event recognition problems. © Springer-Verlag Berlin Heidelberg 2007.