SPINet: self-supervised point cloud frame interpolation network

For autonomous vehicles, the acquisition frequency difference between LiDAR (10-20 Hz) and camera (over 100 Hz) makes simultaneous update of two perceptive systems (2D/3D) less efficient. Nowadays, frame interpolation is in urgent need for increasing frame rate of point cloud sequences obtained by LiDAR. However, a major limitation of current full supervised methods is that high frame rate ground truth sequences are hard to access. We propose a novel Self-supervised Point Cloud Frame Interpolation Network (SPINet) accommodating with variable motion situation, retaining geometric consistency, but without the necessity of utilizing G.T. data. Extensive experiments show that our proposed SPINet outperforms the current full supervised methods.