Dichotomic Decision Cascading for Video Shot Boundary Detection

In this paper, we present a shot boundary decision fusion strategy which implements a multi-modal cascaded dichotomic search on the boundary space. The initial and core step of the proposed method is narrowing the shot boundary decision space as long as the accuracy is improved. Instead of the default sequential change detection, a dichotomic change strategy which is supervised with a cascaded fusion, is implemented to achieve higher accuracy and less algorithmic complexity. The main decision sources are image color histograms, object recognizer results, motion comparators, audio pattern analyzers, key point extractors and edge descriptors which are selectively employed in a cascaded manner. We propose a shot boundary detection algorithm which is noise tolerant, video genre adaptable, context aware and computationally efficient. In order to reduce computational complexity, we construct a shot boundary search heuristic for pruning the set of candidate shot boundary frames. We employ both statistical and rule based approaches in a cascaded fashion in order to decide the size of the search space to be pruned for the purposes of improving computational efficiency. TRECVid 2006 and 2007 data sets are used in the evaluation process and the performance results are given for both cuts and gradual transitions.