Research On Deep Convolutional Neural Networks For Point Cloud Perception With Adaptive Permutation Module

Point clouds are gaining increased popularity recently in areas such as autonomous driving and robotic systems and have become one of the most important data formats for 3D representation.Since it is a necessary part of point cloud processing,point feature extraction has also attracted increasing attention lately.Due to the significant success achieved by CNN in 2D computer vision tasks,many researchers have tried to extend this success from 2D to 3D point clouds.However,there is a problem with deep convolutional networks which directly consume on point clouds:point clouds are irregular and unordered.In general,there are two kinds of deep learning methods to handle point order ambiguity in point cloud data:one is that uses symmetric functions like max-pooling to get the similar features with input point clouds in different order.The another one is to perform a permutation to the point cloud.However,existing methods have been inadequately explored in terms of permutation,which may limit the effectiveness of point cloud feature extraction.To that end,we present an adaptive permutation module(APM)that calculates a particular permutation for each input local point subsets to achieve permutation invariance.We conduct thorough experiments to demonstrate the superiority of this module.In addition,the APM can be plugged into other state-of-the-art approaches flexibly to further improve performance in classification task.We build an end-to-end deep convolutional neural network applying PointCNN as our backbone combined with the adaptive permutation module and achieve state-of-art performance on point cloud classification task.Our work demonstrates that the latent spatially-local correlations play a critical role in feature learning on point clouds.The main contributions of this thesis are as follows:1.Proposing a novel adaptive permutation module named APM that can permute the input point clouds to handle point order ambiguity and can be flexibly plugged into other approaches to further improve performance.2.Building an end-to-end effective deep convolutional neural network with a PointCNN backbone combined with the adaptive permutation module and achieving the state-of-art performance on classification and segmentation tasks.3.Conducting a thorough empirical analysis on the superiority of our adaptive permutation module and demonstrating that the latent spatially-local correlations play a critical role in feature learning on point clouds.