SPSN: Seed Point Selection Network in Point Cloud Instance Segmentation

The current mainstream point cloud instance segmentation methods are mainly divided into two steps. Firstly, the points of each instance are aggregated in the feature space by means of metric learning to make the features of the same instance are as similar as possible, and then the aggregated vector clusters are segmented to construct the proposal of each instance. Much of the previous work has focused on the aggregation of vectors and ignored how the instance is divided after the vector aggregation. In this paper, we propose a seed point selection network. The seed point selection network selects a better seed point generation proposal by judging the "seedness" of each point, and completes the instance-level segmentation of all points. In addition, the speed of instance segmentation effectively improved by the fast processing of the generated instance points and the low "seedness" points. In the experiment, we graft the seed point selection network onto different instance segmentation networks, and the accuracy and efficiency of segmentation are improved in different degrees.