Video Search Using a Visual Dictionary

This paper presents an original approach to video shot retrieval, which is an adaptation of the common text-based paradigm. The idea of our approach is to describe images using a small number of visual elements chosen in a visual dictionary. The user may select some elements from the visual dictionary to compose a query and search for specific video shots. In our approach, we automatically compute those visual elements to compose a global visual dictionary, and we select the most representative to build a query visual dictionary. We propose a method to evaluate automatically the efficiency of this query visual dictionary for retrieving shots, and we present experimental results on the tree video BBC rushes task.