A novel method for video shot similarity measures

In this paper, a novel method is proposed to determine the similarity between video shots. A video shot is treated as an ensemble that consists of multiple video key frames, so the shot similarity can be measured by the ensemble similarity. Based on nonlinear mapping, the original space is mapped to a high dimension space where the ensemble distribution can be supposed as normal distribution. Kernel method is adapted to compute the probability distance that is equivalent to the ensemble similarity. Thus, the shot similarity is also obtained. Experimental results show that this method may achieve superior performance than the traditional methods based, on Euclidean distance and histogram intersection methods.