Graph Structure-based Salient Object Detection for Point Clouds

Salient object detection aims to detect and extract the most attractive regions relative to their surroundings, which is more challenging than traditional target detection. A point cloud is a disorganised and discrete set of points in 3D space, which inherently lacks information about the topological relationships between points. The topological relationship between points is particularly important in the task of point cloud salient object detection. Therefore, designing a model that recovers the topological relationships between points can enrich the characterisation capability of point clouds. In this paper, we explore a variety of methods for recovering topological relationships between points. They are constructing directed graphs based on coordinate space, constructing directed graphs based on feature space, constructing directed graphs based on random sampling, and constructing directed graphs based on coordinate space and features. Through analysis and visualisation, we verified that the method of constructing a directed graphs based on feature space is most suitable for recovering the topological relationship information between points. Experimental results show that this method achieves state-of-the-art performance on PCSOD public data.