DEEP LEARNING ON POINT CLOUD FOR 3D CLASSIFICATION BASED ON SPIKING NEURAL NETWORK

Point cloud is a real representation of the physical world in a digital way. As an important geometric data structure, it is widely used in visualization, animation, rendering and modeling. Because of its irregular format, we have to convert the data to a normal way thus it can be used, which also leads to the data becoming extremely large. In this paper, we build a spiking neural network (SNN) based on spike-timing-dependent plastic rules (STDP) [1-3], which can process point cloud data more efficiently by using the sparse characteristics of the spiking neural network. At the same time, the SNN naturally takes into account the displacement invariance and rotation invariance of point clouds. The accuracy of the neural network model we designed in the classification task is more or less equal to that of the existing point cloud processing technology, but it is deployed on edge devices with better performance as lower power consumption, higher efficiency and so on.