3D object detection plays an important role in au-tonomous driving and other robotics applications. How-ever, these detectors usually require training on largeamounts of annotated data that is expensive and time-consuming to collect. Instead, we propose leveraging largeamounts of unlabeled point cloud videos by semi-supervisedlearning of 3D object detectors via temporal graph neu-ral networks. Our insight is that temporal smoothing cancreate more accurate detection results on unlabeled data,and these smoothed detections can then be used to re-train the detector. We learn to perform this temporal rea-soning with a graph neural network, where edges repre-sent the relationship between candidate detections in dif-ferent time frames. After semi-supervised learning, ourmethod achieves state-of-the-art detection performance onthe challenging nuScenes [2] and H3D [16] benchmarks,compared to baselines trained on the same amount of la-beled data. Project and code are released athttps://www.jianrenw.com/SOD-TGNN/