Recent works in object recognition often use visual words, i.e. vector quantized local descriptors extracted from the images. In this paper we present a novel method to build such a codebook with
vectors. This method, coined
Cluster Precision Maximization
(CPM), is based on a new measure of the cluster precision and on an optimization procedure that leads any clustering algorithm towards class representative visual words. We compare our procedure with other measures of cluster precision and present the integration of a Reciprocal Nearest Neighbor (RNN) clustering algorithm in the CPM method. In the experiments, on a subset of the the Caltech101 database, we analyze several vocabularies obtained with different local descriptors and different clustering algorithms, and we show that the vocabularies obtained with the CPM process perform best in a category-level object recognition system using a Support Vector Machine (SVM).
Bitte loggen Sie sich ein, um Zugang zu diesem Inhalt zu erhalten