Discriminant local features selection using efficient density estimation in a large database

In this paper, we propose a density-based method to select discriminant local features in images or videos. We first introduce a new fast density estimation technique using a simple grid index structure and specific queries based on the energy of the gaussian function. This method enables the nonparametric density estimation of target features with very large sets of source features. We then apply it to the selection of discriminant local features: the principle is to keep only the features having the lowest density in a feature database constructed from a large collection of representative objects (images or videos). Experiments are reported to evaluate the density estimation technique in terms of both quality and speed. The density-based selection of discriminant local features is evaluated in a complete video content-based copy detection framework using Harris interest points.