PointINet: Point Cloud Frame Interpolation Network

LiDAR point cloud streams are usually sparse in time dimension, which is limited by hardware performance. Generally, the frame rates of mechanical sensors 10 to 20 Hz, much lower than other commonly used like cameras. To overcome temporal limitations sensors, a novel task named Point Cloud Frame Interpolation studied this paper. Given two consecutive frames, aims generate intermediate frame(s) between them. achieve that, we propose framework, namely Network (PointINet). Based on proposed method, low rate can be upsampled higher rates. We start estimating bi-directional 3D scene flow clouds and then warp them given step based flow. fuse warped frames cloud(s), learning-based points fusion module, simultaneously takes into consideration. design both quantitative qualitative experiments evaluate performance interpolation method extensive large scale outdoor datasets demonstrate effectiveness PointINet. Our code available at https://github.com/ispc-lab/PointINet.git.