Learning object class detectors from weakly annotated video