Fast and effective features for recognizing recurring video clips in very large databases

Three different frame features (color patches, color coherence vectors, and gradient histograms) are investigated for their suitability to recognize recurring video clips in very large databases. They are evaluated in a real-time processing and real-time recognition system. Real-time recognition means that each clip must be recognized one second after its start. As the experimental results show, only gradient histograms work satisfactorily across different video material with the same video domain independent parameter set. For instance, they are - in contrast to color features - not negatively affected by dark frame sequences in video clips and the live video stream. By means of pre-computation and subsequent table look-ups, gradient histograms can be implemented such that their computational costs come very close to that of color features.