NII-KAORI-PERSON-SEARCH: A General Framework for Indexing and Retrieving People's Appearance in Large Video Archives

Abstract

We introduce a general framework to index and retrieve people's appearance using their faces and names. All frontal faces are extracted from video programs and are grouped into clusters using temporal and visual information. These clusters are annotated by a semi-supervised method in which users help to annotate a small number of faces, and correct annotations suggested by the system. Once faces are annotated, users can search people by names or exemplar faces. In addition, trends of people's appearance and correlation between these trends in time space can be observed. We demonstrate such system for approximately 2,000 hours of videos broadcast by NHK Japan in 11 years from Mar 2001 to Mar 2011.