How many high-level concepts will fill the semantic gap in video retrieval ?

A number of researchers have been building high-level seman tic concept detectors such as outdoors, face, building, etc., t o help with semantic video retrieval. Using the TRECVID video collecti on and LSCOM truth annotations from 300 concepts, we simulate perf ormance of video retrieval under different assumptions of con cept detection accuracy. Even low detection accuracy provides g ood retrieval results, when sufficiently many concepts are used. C onsidering this extrapolation under reasonable assumptions, th is paper arrives at the conclusion that “concept-based” video retri eval with fewer than 5000 concepts, detected with minimal accuracy of 10% mean average precision is likely to provide high accuracy re sults, comparable to text retrieval on the web, in a typical broadca st news collection. We also derive evidence that it should be feasib le to find sufficiently many new, useful concepts that would be help ful for retrieval.