GPr-Net: Geometric Prototypical Network for Point Cloud Few-Shot Learning

In the realm of 3D-computer vision applications, point cloud few-shotlearning plays a critical role. However, it poses an arduous challenge due tothe sparsity, irregularity, and unordered nature of the data. Current methodsrely on complex local geometric extraction techniques such as convolution,graph, and attention mechanisms, along with extensive data-driven pre-trainingtasks. These approaches contradict the fundamental goal of few-shot learning,which is to facilitate efficient learning. To address this issue, we proposeGPr-Net (Geometric Prototypical Network), a lightweight and computationallyefficient geometric prototypical network that captures the intrinsic topologyof point clouds and achieves superior performance. Our proposed method, IGI++(Intrinsic Geometry Interpreter++) employs vector-based hand-crafted intrinsicgeometry interpreters and Laplace vectors to extract and evaluate point cloudmorphology, resulting in improved representations for FSL (Few-Shot Learning).Additionally, Laplace vectors enable the extraction of valuable features frompoint clouds with fewer points. To tackle the distribution drift challenge infew-shot metric learning, we leverage hyperbolic space and demonstrate that ourapproach handles intra and inter-class variance better than existing pointcloud few-shot learning methods. Experimental results on the ModelNet40 datasetshow that GPr-Net outperforms state-of-the-art methods in few-shot learning onpoint clouds, achieving utmost computational efficiency that is $170\times$better than all existing works. The code is publicly available athttps://github.com/TejasAnvekar/GPr-Net.