基于邻域参与的形状感知卷积网络的点云分析

Neighborhood participated relation-shape convolutional neural network for point cloud analysis

点云作为三维传感器的主要输出,在许多领域有着广泛的应用.但是,由于点云中点的不规则性,使得它不像二维图像那样,可以直接使用深度网络进行分析处理.提出了一种新的直接在点云上进行深度卷积的方式——邻域参与的形状感知卷积.对于给定邻域,首先通过多层感知器(Multi-layer Perceptron,MLP)对邻域内每一点提取点特征,然后利用平均池(Average Pooling)操作获得邻域的全局特征,再将全局特征与点特征串联,并通过MLP学习出与输入特征维度相同的权重.利用权重对输入特征进行加权,最终通过最大池(Max Pooling)和MLP得到邻域的特征描述.实验结果表明,算法能够提高三维点云分类和语义分割的准确率.在ModelNet40数据集上的分类准确率为93.2%,在ShapeNet Parts数据集上语义分割的平均交并比为86.4%.

Point clouds,as the main output of three-dimensional(3 D)sensors,have extensive applications in some fields.However,since the points in the point cloud are irregular,they can not be trivially consumed by convolutional networks in the same way the 2 D image.In this paper,we propose a new convolution operation in point clouds,named neighborhood participated relation-shape convolution.For a given neighborhood,first we extract features for each point in the neighborhood through a multi-layer perceptron(MLP).Then using the average pooling operations to obtain the global features of the neighborhood,and concatenating the global features and the point features to learn a weight for each point.The weight is regressed by MLP and has the same dimension with the input features.The input features are weighted using computed weights and the final features for the input neighborhood is obtained by the maximum pooling(Max Pooling)and MLP.Experimental results show that,this algorithm improves the accuracy of 3 D point cloud classification and semantic segmentation.Classification accuracy on ModelNet40 dataset is 93.2%,average intersection ratio of semantic segmentation on the ShapeNet Parts dataset is 86.4%.