GGF: Global Geometric Feature for Rotation-Invariant Point Cloud Understanding

Most 3D vision tasks related to point cloud understanding require rotation-invariant solutions. However, existing deep learning techniques for point clouds do not always ensure rotation invariance, and achieving rotation invariance necessitates data augmentation or professional networks, which is not feasible for generic point cloud understanding tasks. To address this issue, we propose a plug-and-play feature called Global Geometric Feature (GGF), as an effective and efficient solution achieving rotation invariance for generic point cloud understanding networks. GGF extracts a distributed global description by capturing geometric relationships between points and projecting the point cloud into a rotation-invariant feature space. We find that GGF can be directly integrated into a variety of point cloud understanding tasks without network modification. Our experimental evaluation shows that GGF can improve the performance of generic point cloud networks for rotation-invariant understanding without data augmentation and is comparable to dedicated rotation-invariant methods.