Progressive background–foreground difference enhancement for few-shot 3D point cloud semantic segmentation

Few-shot 3D point cloud semantic segmentation aims to segment query point clouds given only few annotated support point clouds. Most existing methods focus on exploring the complex relationships between support data and query data within the prototype-based framework. However, the ignored background ambiguity issue, i.e., the foregrounds of a support class are treated as backgrounds by other support classes, severely limits few-shot models’ ability to distinguish foregrounds and backgrounds, resulting in biased prototypes. In this paper, we propose a progressive background–foreground difference enhancement method to eliminate background ambiguity. Firstly, based on the fact that the background ambiguity only affects background prototypes, we develop a background–foreground difference enhancement strategy, which eliminates background ambiguity via enhancing the difference between foregrounds and backgrounds in query data. Then, we present a geometric-guided feature aggregation module, which integrates geometrical information to improve the reliability of pseudo labels. Finally, we aggregate high-confidence query features as pseudo prototypes to refine the prototypes. The iteration of these steps further improves prototype quality. Comprehensive experiments suggest that our method achieves competing performance on both S3DIS and ScanNet datasets.