FINet: Fast Point Cloud Interpolation Network Via Distance Transform

In recent years, sensors such as cameras and lidars have been widely used in in the field of perception. However, the differing frame rates between these sensors limit their effective fusion. While substantial progress has been made in the temporal interpolation of 2D video sequences, research on 3D point cloud sequence interpolation remains relatively underdeveloped. In this paper, we introduce FINet, a coordinate-based neural network scene flow model designed to address the challenge of rapid temporal frame interpolation for 3D dynamic point clouds. We develop a novel point cloud similarity loss function based on Distance Transform, which measures the similarity between two point cloud frames similarly to the Chamfer distance loss but with significantly reduced computational overhead by avoiding nearest neighbor search. We demonstrate the advantages of our proposed method in terms of performance and computational efficiency on point cloud sequence datasets. Our method achieves over a 50% improvement in quantitative comparison of computational overhead compared to other loss functions. Consequently, our point cloud frame interpolation network is better suited for real-time scene applications.