Learning isometry-invariant representations for point cloud analysis

Highlights：

• We propose a novel method to learn isometry-invariant representation for point cloud which is strictly rotation-invariant and capable of tolerating pose variations. Both mathematical analyses and experimental results demonstrate that the proposed method can extract isometry-invariant representations for 3D shape analysis tasks without rotation augmentation.

• In order to learn the local representation of point cloud, we build the local graph based the geodesic distance and propose the graph aggregation method to obtain the local feature, which is important for tolerating the shape variations in non-rigid shape analysis.

• We construct a novel dataset, PKUnon-rigid, which is composed of non-rigid 3D objects. Based on it we can evaluate the capacity of several mainstream methods in terms of processing non-rigid shapes.