3D Object Recognition Method Based on Point Cloud Sequential Coding

Abstract

Point cloud is the most usual manner to describe 3-dimensional (3D) objects. It contains the position of each point of the object in spatial rectangular coordinate system and their corresponding RGB pixel values, which can describe the object appearance completely. However, point cloud data is dense and unordered, and hence not suitable for convolutional neural networks (CNNs) that is the basic of deep learning. In this work, a new 3D object recognition method based on sequential coding of point cloud is proposed, with which a point cloud can be transformed into an ordered multi-channel 2D array that is suitable for efficient 2D convolutional operation. Point cloud can be sequential coded in spherical coordinate or cylindrical coordinate. Based on the new ordered data, a classifying network and a retrieval framework is design for recognition of 3D objects. The proposed method has achieved better results in both classification and retrieval tasks than other methods.