GraphPointNet: Graph Convolutional NeuralNetwork for Point Cloud Denoising

The project proposed is finalized to develop a novel network for Point Cloud denoising based on Graphs. A Point Cloud is an object representation composed by a collection of 3-D space coordinates. This data is usually acquired by radar, laser, Electro-Optical systems or by reconstruction of 2-D images: all these methods lead to point cloud typically affected by noise. The aim of the project is to design a network able to efficiently produce cleaned 3-D point cloud from a noisy observation. Denoising task is a typical problem addressed in Image Processing and the current state-of-the-art is Convolutional Neural Network (CNN), that leads to promising results for noise removal in images; the idea developed in this project is to exploit a Deep Neural Network, composed by convolutional layers, introducing appropriate adjustment for the Point Cloud denoising task. The novelty of the project is the introduction of a graph-convolutional layer,that exploits the Edge-Conditioned-Convolution [1] to implements a graph-convolution operation over point cloud.A graph is computed for each Point Cloud, where each single point is a node and the weighted connections between them are the edges. The ECC is performed ina nested Deep Neural Network structure, where the feature vector associated to one node at layer l+ 1 is computed as a weighted local aggregation of the feature vector at layer l of the node itself and the nodes in the neighborhood. Traditional methods to denoise a point cloud are geometrical method and graph representation. Some popular methods are based on the computation of the surface of the point cloud from the noisy observation and then project the noisy points, other represent the point cloud on graph and exploit the graph-regularization method. All this approaches need a optimization process.Recently, due to the increasing interest in the point cloud denoising, neural network project have been published. All this methods achieve promising results, sometimes outperforming the traditional method. None of these networks exploits a graph representation of the data, proposing a quite simple architecture. The network proposed in this thesis, called GraphPointNet, would be the first neural network based on a convolution able to process point cloud. In this thesis after a introduction section where the basic concepts of neural network and graph theory are presented, the current state-of-arts are summarized. Then the development of the project is described, from the creation of the dataset, to the presentation of the architecture analyzed in details. Finally,the performance evaluations of the network proposed are reported. In order to evaluate the results obtained, quantitative and qualitative test are performed.In particular, the point-to-point distance is taken into account to evaluate the goodness of the results obtained and to make comparisons with other methods.It is shown that the method proposed is able to outperform or at least match the current state-of-arts.