Selective multi-cotraining for video concept detection