The structure of Semi-supervised Co-training: Firstly, Faster-RCNN based on MoCov1 and MoCov2, which are unsupervised learning methods during pre-training, are implemented on the KITTI dataset. Secondly, the two models pre-trained with MoCov1 and MoCov2 make predictions on the unlabeled data respectively. Only the prediction results of objects higher than a defined threshold would be retained. After that, the one with a lower classification score from two predictions is discarded and the retained prediction result is regarded as pseudo-label with the information of class and bounding box. Finally, the unlabeled data with pseudo-labels are added to the training dataset and used for semi-supervised learning, where BBAug is also used during the re-training process to boost the robustness