UEC at TRECVID 2011 SIN and MED task

In this paper, we describe our approach and results for the semantics indexing (SIN) task and Multimedia event detection (MED) task at TRECVID2011. In our runs of SIN task, we used six features, spatiotemporal (ST) features, SURF, color, face, sound features and word histogram. This year, we use multiple frames selected by calculating the color difference between frames, not all frame. All runs used Multiple Kernel Learning as a fusion method to combine all these features in the same way as last year. Our submitted runs are as follows: • UEC1 1: SURF, color, ST, face, sound features • UEC2 2: Run1 & word histogram • UEC3 3: Run2 & sort using a category and video name • UEC4 4: Run2 with TRECVID 2010 training data As a result of the full-category SIN task, Run4 yielded the best performance (infAP=0.0452) among four runs. In MED task, we divide videos to shots which is 150 frames at most and extract SURF, color, ST features from shots. We get the average of the top three shot scores as the original video score.