Few-shot learning (FSL) offers a solution when large-scale annotation is difficult or costly. However, extending FSL to three-dimensional (3D) point clouds remains challenging because point clouds are unordered and irregular, while large, well-annotated 3D datasets are relatively limited. This makes it difficult for feature learning methods to fully exploit the geometric structure of point cloud data.

In a study published in Knowledge-Based Systems, a team led by Prof. CHAO Jianshu from the Fujian Institute of Research on the Structure of Matter of the Chinese Academy of Sciences developed a dual-hierarchy learning framework, DHNet, for few-shot 3D point cloud classification, introducing the concept of dual-hierarchy and jointly modeling hierarchical structures at patch and object levels.

Researchers observed that point clouds contain hierarchical structures at two complementary levels. At the patch level, local geometric components are composed of structurally related patches, while at the object level, different instances can share a common semantic category and exhibit relationships among related classes. DHNet exploits these geometric priors, and adopted hyperbolic geometry.

DHNet comprises two key modules. PHFormer is a Transformer-based embedding network that captures patch-level structural dependencies via hierarchy attention. HyperHead is a classification head that leverages hyperbolic geometry to model object-level hierarchies, enabling more discriminative representations in a non-Euclidean space.

Experiments on four benchmark datasets, i.e. ModelNet40, ModelNet40-C, ScanObjectNN, and ShapeNet70, demonstrated that DHNet consistently outperformed state-of-the-art methods under standard 5-way 1-shot and 5-way 5-shot FSL settings. On ShapeNet70, the 5-way 1-shot accuracy showed an improvement of 4.41%. These results highlight the effectiveness of dual-hierarchy modeling and hyperbolic geometry in addressing the issues of few-shot 3D point cloud classification.

This study shows that explicitly modeling the dual-hierarchy inherent in 3D point clouds improves data-efficient representation learning, offering a new direction for few-shot 3D perception, particularly in settings where labeled point-cloud data are scarce. DHNet provides a unified approach for exploiting local geometric structure and global semantic relationships.