Learning to detect occluded objects in videos