3D object detection and scene understanding are the key technologies for autonomous driving scenarios. Due to the differences in configuration and datasets used by each 3D object detection algorithm, it is difficult to evaluate the performance of each method. In this work, we provide a comparison of the advanced 3D object detection networks based on LiDAR point cloud in recent two years and analyze each network structure in detail. For the open-sourced networks, we reproduce them on KITTI dataset benchmark with following their original algorithms. Meanwhile, in order to provide more powerful results, we also utilize nuScenes dataset to retrain the networks as mentioned above. The experimental results show that the performance of the networks with point cloud and images as input is better than that of a single input network.