A compact and retrieval-oriented video representation using mosaics

Abstract

Compact yet intuitive representations of digital videos are required to combine high quality storage with interactive video indexing and retrieval capabilities. The advent of video mosaicing has provided a natural way to obtain content-based video representations which are both retrieval-oriented and compression-efficient. In this paper, an algorithm for extracting a robust mosaic representation of video content from sparse interest image points is described. The representation, which is obtained via visual motion clustering and segmentation, features the geometric and kinematic description of all salient objects in the scene, being thus well suited for video browsing, indexing and retrieval by visual content. Results of experiments on several TV sequences provide an insight into the main characteristics of the approach.