基于超点图的点云实例分割方法

Point Cloud Instance Segmentation Method Based on Superpoint Graph

摘要:

提出一种基于超点图的点云实例分割（ISPG）方法。基于超点图结构提取点云对象相邻点之间的关联性特征，并且将传感器扫描的场景划分为均匀的几何元素，用来表示同属性的点云类，再由一个图卷积网络实现实例分割。结果表明：IoU阈值为0.5的情况下，该方法在斯坦福大型三维（3D）室内空间数据集S3DIS上精度达到了48.9%。

Abstract:

A point cloud instance segmentation method based on superpoint graph(ISPG) is proposed in this paper. The correlation between the adjacent points of point cloud objects is extracted based on superpoint graph structure. Then, the scene scanned by a sensor is divided into uniform geometric elements to represent the point cloud classes with same attributes. Finally, a graph convolution network is used to implement instance segmentation. It is shown that the method has an accuracy of 48.9% on Stanford’s large 3D indoor spatial dataset S3DIS when the IoU threshold is 0.5.