Point Cloud Processing with Neural Networks

In this project, we explore new techniques and architectures for applying deep neural networks when the input is point cloud data. We first consider applying convolutions on regular pixel and voxel grids, using polynomials of point coordinates and Fourier transforms to get a rich feature representation for all points mapped to the same pixel or voxel. We also apply these ideas to generalize the recently proposed interpolated convolution , by learning continuous-space kernels as a combination of polynomial and Fourier basis kernels. Experiments on the ModelNet40 dataset demonstrate that our methods have superior performance over the baselines in 3D object recognition