Clustering for Video Retrieval

The paper deals with an application of clustering we used as one of data reduction methods included in processing huge amount of video data provided for TRECVid evaluations. The problem we solved by means of clustering was to partition the local feature descriptors space so that thousands of partitions represent visual words, which may be effectively employed in video retrieval using classical information retrieval techniques. It has proved that well-known algorithms as K-means do not work well in this task or their computational complexity is too high. Therefore we developed a simple clustering method (referred to as MLD) that partitions the high-dimensional feature space incrementally in one to two database scans. The paper describes the problem of video retrieval and the role of clustering in the process, the MLD method and experiments focused on comparison with other clustering methods in the video retrieval application context.