Scalable Spatio-Temporal Video Indexing Using Sparse Multiscale Patches

In this paper we address the problem of scalable video indexing. We propose a new framework combining sparse spatial multiscale patches and Group of Pictures (GoP) motion patches. The distributions of these sets of patches are compared via the Kullback-Leibler divergence estimated in a non-parametric framework using a k-th Nearest Neighbor (kNN) estimator. We evaluated this similarity measure on selected videos from the ICOS-HD ANR project, probing in particular its robustness to resampling and compression and thus showing its scalability on heterogeneous networks.