Fine-grained Object Matching Based On Rotation-invariant Embeddings

Semantic understanding of 3D object and the rotation-invariant property are the crucial problems in real applications.Algorithms that designed for semantic analysis face challenges in real scenes.Besides,it is hard to give a canonical definition of semantics,because of its vagueness.Some datasets and methods of 3D object understanding are limited to different specific tasks.However,human beings always have common consensus on correspondences among objects by observing.Therefore,exploring fine-grained correspondences could be a novel way to avoid those problems.Therefore,the main contributions of this paper are:1.In order to overcome the challenge that the pose of object is unknown in real applications,we propose the unit sphere space and further extract the rotation-invariant features in this space.Besides,a point re-sampling module is designed for point cloud reconstruction from spherical signal,which ensures that the network can predict pointwise rotation-invariant features.2.Due to the vague definition of semantics,we explore the correspondences among different objects and design the geodesic consistency loss to train the network,which make the network can learn semantic representations of objects and ensure the differences between different points.3.To measure the performance of the matching task,we design mean geodesic error algorithm.Furthermore,two applications are studied to verify the significance of this paper.