Spatiogram-Based Shot Distances for Video Retrieval

We propose a video retrieval framework based on a novel combination of spatiograms and the Jensen-Shannon divergence, and validate its performance in two quantitative experiments on TRECVID BBC Rushes data. In the first experiment, color-based methods are tested by grouping redundant shots in an unsupervised clustering. Results of the second experiment show that motion-based spatiograms make a promising fast, compressed-domain descriptor for the detection of interview scenes. Experiment 1: Clustering Run-ID concept NN error rate (%) clustering error (%) CH-L1 baseline: color histograms 33.5 53.0 CS1D-JSD our framework 13.4 42.3 CS-JSD our framework 15.4 46.5 CS1D-PROB different similarity measure [1] 36.2 90.6 CS-PROB different similarity measure [1] 44.2 87.2 WH-L1 color histograms over local windows 16.1 50.3 BVW-Harris ”bag-of-visual-words” 24.8 91.9 Experiment 2: Interview Detection Run-ID concept Precision (%) Recall (%) MH-L1 baseline: motion histograms 49.0 71.4 MS1D-JSD our framework 69.0 84.3 MS-JSD our framework 49.4 78.1 MS1D-PROB different similarity measure [1] 44.1 67.6 MS-PROB different similarity measure [1] 46.3 65.7 WH-L1 motion histograms over local windows 59.7 81.9 CS1D-JSD color-based method 44.0 10.5