DGPoint:用于三维点云语义分割的动态图卷积网络

DGPoint:A Dynamic Graph Convolution Network for Point Cloud Semantic Segmentation

三维点云语义分割在三维场景理解和重建、自动驾驶和机器人导航等领域发挥着重要作用。针对现有方法对点云的局部特征提取不足导致分割精度不高这一问题,基于PointNet++架构提出了一个动态图卷积网络DGPoint。首先,边缘卷积的特征聚合函数使用双通道池化操作来补偿信息丢失,可以更好地保留点云的细粒度局部特征;其次,在进行边缘卷积之前使用K近邻算法确定新的局部区域以达到动态图更新的效果;此外,为了保证边缘提取准确,将设计的编码器重复多次,并将提取的特征以跳跃连接的方式连接起来输入到解码器中。在S3DIS数据集上的实验结果表明,与其他方法相比,DGPoint有效解决了局部特征提取不足的缺点,并提高了语义分割的准确性:在S3DIS数据集的平均交并比为68.3%,总体准确度为86.2%。

Semantic segmentation of point cloud data plays an important role for 3D scene understanding and reconstruction,autonomous driving,and robot navigation.In this study,DGPoint,a dynamic graph convolution network based on the PointNet++architecture,is proposed to address the insufficient segmentation accuracy due to insufficient local feature extraction of point clouds by existing methods.First,the feature aggregation function in edge convolution compensates for loss using a dual-channel pooling operation,which can better retain the fine-grained local characteristics of the point cloud.Then,to accomplish the impact of dynamic graph updates,K-nearest neighbors algorithm is used to determine new local regions prior to edge convolution.Additionally,to ensure the accuracy of edge feature extraction,the designed encoder is repeated multiple times,and the extracted features are concatenated in a jump-connection style before being input to the decoder.Experimental results of the S3DIS data set show that DGPoint effectively solves the shortcomings of the insufficient local feature extraction and improves semantic segmentation accuracy with the mean intersection over union of 68.3%and overall accuracy of 86.2%compared with other methods.