Object detection in 3D Point Clouds Using Efficient Convolutional Neural Networks

Hello ; I am studying Object detection in 3D Point Clouds Using Efficient Convolutional Neural Networks and I read this article: https://arxiv.org/pdf/1609.06666v1.pdf In this article images convert to 3d point cloud and than passed CNN is it true? But KITTI dataset used in this study and KITTI has png format On the other hand this png format convert to 3d point cloud How is this? I try png format convert to 2d point cloud in matlab But matlab only use ply and pcd format file Secondly KITTI has so many dataset For this article which dataset should I use? Could you help me please?