Regional Attention Based on SPD- Manifold for Point Cloud Classification

In point cloud-based tasks, the attention mechanism falls short in considering the local spatial information of point clouds, despite its good global learning ability. To address this issue, our paper proposes a regional attention mechanism that focuses on the local region features of point clouds. We model these features as manifold data, allowing us to obtain their non-Euclidean data features. Using a Riemannian manifold, we measure the similarity between these regional features and use it as an attention weight to obtain features with regional attention. We also employ a kernel method to ensure that the regional features always adhere to the manifold definition, enabling us to analyze their relationship on the manifold. Our network is tested on the ModelNet40 dataset, and we demonstrate that our method outperforms some state-of-the-art methods.