(A) study on deep learning-based methods for point cloud processing

Abstract

Point cloud is an important data type for 3D representation of objects because it is the raw output of sensors e.g. Lidar and depth sensor etc. It is cheap to obtain and store but difficult to process due to its unstructured and permutation invariant nature. This research presents new point cloud processing techniques using deep learning based methods. A new architecture has been proposed which is named RL-GAN-Net. RL-GAN-Net is a reinforcement learning agent controlled GAN network designed specifically for the task of shape completion and classification for the incomplete point cloud. The primary benefit of this architecture is that it is robust, real-time and modular. Besides, this research also addresses the problem of point cloud reconstruction by improving the performance of a capsules network based auto-encoder. This is achieved by introducing a channel attention mechanism in the capsule layer. All findings have been verified through extensive experimentation.