Spatial Semantics and Classifier Cascades

In this paper, we describe the system developed by the Australian National University (ANU), National ICT Australia (NICTA) for multimedia event detection applied to the TRECVID-2011 video retrieval benchmark. Our system uses five audio and visual features, leverages training events with cascaded classifier training, and sees performance improvements with spatial semantic features. A summary of our submitted runs can be found below: Run number Short hand Description pRun01 01_combined_all Combination from all features, with classifier bagging, and cascade training from SIFT. cRun02 02_combined_nocas Same as Run01 except without cascade training. cRun03 03_combined_visual Run01 without the two audio features. cRun04 04_combined_single Combination from all features, without classifier bagging, with SIFT. The best run from our system ranks fourth in mean-ActualNDC, and third in meanF1 metric, averaged over all ten events among sixty runs from nineteen teams.