VOICI: Video Overview for Image Cluster Indexing -- a swift browsing tool for a large digital image database using similarities

The Leiden Imaging and Multi-media (LIM) group develops visual search tools for content-based image indexing and retrieval. Because visual searches take place in an interactive setting, the methods used have to be fast. Research addressed in this paper deals with ”How to get quickly acquainted with the contents of a large digital image collection”. More formally spoken we have tried to speed-up browsing using video presentation of database content and a bi-level hierarchy of video loops. At the lowest level images from a similarity cluster are shown and at the global highest level the video loop is composed of representative images per cluster. Using image feature vectors, image similarity measures are computed and stored in a distance matrix. The clustering of images is achieved by applying a threshold in this distance matrix. The threshold separates like images from less-like images. The video overview consists of a global path through the distance matrix by using one index image for each image cluster. Image cluster contents can be viewed separately by selecting an index frame from the global overview and by showing a video loop of cluster frames in a minimum variance order. For different threshold settings, MPEG files for the global and cluster overviews have been generated for the image files in the Leiden 19th-Century Portrait Database (LCPD).