LSIS TREC VIDEO 2008 High Level Feature Shot Segmentation using Compact Profile Entropy and Affinity Propagation Clustering

In this task, we build fast video indexing systems using a kind of efficient features based on the entropy of pixel projections. These features of 45 dimensions, called Profil Entropy Features (PEF), are derived using the projection in the horizontal orientation. These features are then fed to SVMs to produce the keyframe ranks, from which we can get the shot ranks. In the runs, we divided the training set into several subsets using randomly method or affinity propagation clustering in order to simplify learning, and then we combined the outputs of the SVMs on the subsets into the final output. Finally we also made some fusions of different runs using arithmetic and harmonic means. We got the inferred MAP of 0.05245 and the 19th rank among all the best by team of 37 automatic submitted runs, the average of which is 0.063 for a STD of 0.0458. Further, our system needs only 11 hours time consumption for training and testing on the whole TREC video sets (on a Linux Xeon 2.66GHZ).