Co-Training for Deep Object Detection : Comparing Single-Modal and Multi-Modal Approaches