Shlukování založené na Voronoiově dláždění pro klasifikaci a vyhledávání ve videu

Clustering, Classification, Video Search, Local Features

Although there are many clustering techniques, it is not possible to use them for all purposes. The initiative problem was to create as many clusters as possible (eg. thousands) for the local image features description in huge amount of video for TRECVid 2008 evaluation. These large dimensional vectors cover the space almost continuously and commonly used clustering methods are unable to create enough classes or to finish in serious time.

Therefore, we have invented a new method based on Voronoi tessellation that needs no more than two passes through the data. It is based on discovery of clusters in higher density locations. Because of large dataset, it is possible to create higher amount of candidate clusters and select appropriate number of classes (large but not huge) and the rest data assign to these classes. The method has been implemented as a set of SQL functions and queries and tested on a huge problem and large amount of classes. Performed experiments have proven that it is significantly faster than common techniques.