Fast coarse-to-fine video retrieval via shot-level statistics

We propose a fast coarse-to-fine video retrieval scheme using shot-level spatio-temporal statistics. The proposed scheme consists of a two-step coarse search and a fine search. At the coarse-search stage, the shot-level motion and color distributions are computed as the spatio-temporal features for shot matching. The first-pass coarse search uses the shot-level global statistics to cut down the size of the search space drastically. By adding an adjacent shot of the first query shot, the second-pass coarse-search introduces the "causality" relation between two consecutive shots to improve the search accuracy. As a result, the final fine-search step based on local color features of key-frames of the query shot is performed to further refine the search result. Experimental results show that the proposed methods can achieve good retrieval performance with a much reduced complexity compared to single-pass methods.