School of Science Department of Mathematics 54 Research in AI and Machine Learning Supervisor: ZHANG Tong / MATH Student: XU Yanbo / COSC Course: UROP2100, Fall The advancement in object detection has benefited many fields including autonomous driving, robotics, and so on. Among the detection problems, 3D detection has been extensively studied by researchers. However, most existing 3D object detection methods rely heavily on the massive amount of labeled data, which is hard to acquire. In the context of autonomous driving, the collection of unlabeled data, i.e., stereo data and video, is relatively easy. To leverage these data, we proposed a semi-supervised training framework with an objectlevel photometric loss which provides a supervision signal for the detection task by Structure From Motion. Extensive experiments on KITTI and nuScenes datasets show the improvement on 3d detection tasks using our method, compared with state-of-the-art approaches.