Selective multi-cotraining for video concept detection

Research interest in cotraining is increasing which combines information from usually two classiers to iteratively increase training resources and strengthen the classiers. We try to select classiers for cotraining when more than two representations of the data are available. The classier based on the selected representation or data descriptor is expected to provide the most complementary information as new labels for the target classier. These labels are critical for the next learning iteration. We present two criteria to select the complementary classier where classication results on a validation set are used to calculate statistics for all the available classiers. These statistics are used not only to pick the best classier but also ascertain the number of new labels to be added for the target classier. We demonstrate the eectiveness of classier selection for semantic indexing task on the TRECIVD 2013 dataset and compare it to the self-training.