FRPoint: learning spatial correlations in the frequency domain for point cloud classification

The use of neural networks for identifying irregular 3D point cloud data is gaining increasing attention from researchers. Similar to CNN, in order to extract local features of point clouds, point cloud models typically construct neighborhoods, extract features from each neighborhood point, and finally use symmetric aggregation functions to capture local information. However, how to accurately utilizing spatial correlation in 3D space remains a challenging problem. To address this issue, we propose FRPoint, which transforms neighborhoods into frequency domain space and capture long and short distance dependencies in the frequency domain using a learnable weight matrices. Benefit to the frequency domain's ability to well describe the spatial relationships, our method efficiently extracts accurate spatial correlations. More importantly, our method only requires the introduction of a simple learnable matrix, without the need for any complex Attention operations or stacking of intricate feature modules. The effectiveness of our method has been demonstrated through experiments on the classification datasets ScanObjectNN and ModelNet40.