3D Point Clouds Classification And Segmentation Based On Graph Convolutional Networks

3D Point Clouds is a very important kind of spatial geometry data,which is generally used to construct the surface shape of 3D geometric objects.At the same time,point clouds are also the original data form generated by Li DAR,3D sensors,and TOF Camera.With the development of autonomous driving,VR and AR technique in recent years,the acquirement of 3D point clouds has become more and more convenient,but 3D point clouds is an irregular and disorder data format,efficient modeling method for 3D point clouds has always been an arduous challenge.Deep Learning has achieved great success in many application fields recently,especially in image recognition and natural language processing.However,it shows strenuous in the irregular non-Euclidean domain data modeling aspect,the tensor represent mode of deep learning algorithms cannot be applied on irregular non-Euclidean domain directly.So as a typical kind of non-Euclidean domain data,3D point clouds processing methods with traditional deep learning algorithms are difficult and unsuitable.Graph Convolutional Networks(GCN)model is another emerging research hotspot in the past few years,the unique feature extraction mode of GCN is a suitable method for irregular non-Euclidean domain data modeling,so GCN models applies on 3D point clouds processing is also a very popular attempt in recent researches.In this paper,we propose a Multi-scale Dynamic Graph Convolutional Network model for 3D point clouds classification tasks,and a Graph Convolution U-Net model for 3D point clouds segmentation tasks.1.Multi-scale Dynamic Graph Convolutional Network:3D Point Clouds classification is a relatively coarse-grained image recognition task,similar to traditional deep learning image classification processing,3D point clouds classification models generally extract local features by a series of convolution operations and utilizing a global pooling layer to extract global semantic features for classification.In this paper,we propose a Multi-scale Dynamic Graph Convolutional Network model for 3D point clouds classification,a Farthest Point Sampling(FPS)method was applied firstly in this model to efficiently cover the entire points set,and then employed different scale k-NN group method to location k nearest neighborhood for each central node.Moreover,different depth Edge Convolution layers were used to extract and aggregate local features between neighbor connected nodes and the central node,a global max pooling layer was employed finally to explore 3D point clouds global semantic features for classification.Experiments show that our Multi-scale Dynamic Graph Convolutional Network model achieves a better performance on classification accuracy and model computational complexity than other state-of-the-art models,maintain the classification accuracy reaches a high level,and meanwhile reduces the model computational complexity dramatically.2.Graph Convolution U-Net:Compared with the classification task,3D point clouds segmentation belongs to a fine-grained,pixel-level image recognition task,each data point in point clouds needs to be divided into its own category.The image semantic segmentation model in deep learning generally uses an end-to-end symmetrical structure,on the first half of segmentation models,a series of convolutional layers and pooling layers were used to extract local information from its receptive fields first,the second half restores the encoded image features to the original image features dimension size by a series of deconvolution operation and output pixel-level classification results finally.In this paper we propose a Graph Convolutional U-Net model for 3D point clouds segmentation,following the classical U-Net model architecture,the left part of Graph Conv U-Net model extracts and encode the local features information of original point clouds by a series of Farthest Point Sampling method,k-NN Group and Edge Convolution operation,the right part utilized the k-NN interpolation algorithm to up-sampling the encoded data points,to restore the number of data points and feature dimensions to the original size and achieving end-to-end output of the model.Experiment shows that our Graph Conv U-Net model achieves a very mature performance on the 3D point clouds segmentation,all the test criteria are reached at a similar or a higher level compared with the current state-of-the-art 3D point clouds segmentation models.