Geometry-Aware 3D Salient Object Detection Network

TL;DR

The paper tackles 3D salient object detection in point clouds, where prior methods struggle with blurry boundaries due to limited geometric context. It introduces a geometry-aware network (3DGAS) that explicitly leverages local geometry by partitioning points into superpoints and then enhancing point features through a geometry fusion mechanism via superpoint–point attention, aided by a point-cloud class-agnostic loss to improve superpoint quality without semantic labels. The two modular components—superpoint partition and geometry enhancement—along with the CA loss achieve state-of-the-art performance on PCSOD while offering faster inference. This approach advances boundary-aware 3D segmentation, enabling more reliable 3D scene understanding in cluttered environments.

Abstract

Point cloud salient object detection has attracted the attention of researchers in recent years. Since existing works do not fully utilize the geometry context of 3D objects, blurry boundaries are generated when segmenting objects with complex backgrounds. In this paper, we propose a geometry-aware 3D salient object detection network that explicitly clusters points into superpoints to enhance the geometric boundaries of objects, thereby segmenting complete objects with clear boundaries. Specifically, we first propose a simple yet effective superpoint partition module to cluster points into superpoints. In order to improve the quality of superpoints, we present a point cloud class-agnostic loss to learn discriminative point features for clustering superpoints from the object. After obtaining superpoints, we then propose a geometry enhancement module that utilizes superpoint-point attention to aggregate geometric information into point features for predicting the salient map of the object with clear boundaries. Extensive experiments show that our method achieves new state-of-the-art performance on the PCSOD dataset.

Figures (6)

• Figure 1: Comparison between our method and existing methods (PointSal fan2022salient, EPFNet zhang2023enhanced and PSOD-Net wei2024point) in 3D salient object detection task. We explicitly utilize the structural information of the point cloud by introducing superpoint representation. It can be observed that using superpoints can effectively characterize the structure of the green cup.
• Figure 2: The pipeline of our geometry-aware 3D salient object detection network. Given a point cloud, we first use the 3D CNN backbone to extract point features. Then, we adopt the superpoint partition module to obtain superpoints. After that, we propose the geometry enhancement module to encode structural information into point clouds. In addition, we propose a point cloud class-agnostic loss $\mathcal{L}_{agn}$ to learn discriminative point features for improving superpoint quality.
• Figure 3: Superpoint Generation Algorithm
• Figure 4: Visualization results of five methods under different views in the test set of the PCSOD dataset. Note that "GT","PSOD", "PSal", "PNext", and "PTrans" represent the ground truth, PSOD-Net wei2024point, PointSal fan2022salient, PointNeXt qian2022pointnext, Point Transformer zhao2021point, respectively.
• Figure 5: Visualization results of the generated superpoints by our method in the test set of the PCSOD dataset. Please note that superpoints are randomly colored.