An FPGA based Parallel Implementation for Point Cloud Neural Network

Point cloud is an important type of data structure for 3d visual information processing. Convolutional Neural Networks(CNNs) like Pointnet [4] are proposed for effective feature learning of point clouds recently. Pointnet is a novel type of neural network that directly consumes point clouds whcih can select informative points. In this paper, we design a fast, low-power FPGA accelerator for Pointnet. Most of the current FPGA accelerators currently do not handle point cloud data very well. The network designed for the FPGA implement, called O-Pointnet, optimizes Pointnet's nonlinear implementation, multi-layer sensing layer and maximum pooling layer by analyzing the implementation of the network and the space-time complexity. Experiments show that the computational performance is 1.208 GMAC/s with 2.149W power dissipation, and the the accuracy rate of the O-Pointnet is 88.48%, which outperforms previous approaches.