Motion retrieval by temporal slices analysis

We investigate video shots retrieval based on the analysis of temporal slice images. Temporal slices are a set of 2D images extracted along the time dimension of image sequences. They encode a rich set of motion clues for shot similarity measure. Because motion is depicted as texture orientation in temporal slices, we utilize various texture features such as tensor histogram, Gabor feature, and the statistical feature of a co-occurrence matrix extracted directly from slices for motion description and retrieval. In this way, motion retrieval can be treated in a similar way as a texture retrieval problem. Experimental results indicate that the features extracted from slices perform satisfactorily in the sport video domain and are, in general, superior to the histogram of MPEG motion vector.