"A new rotation equivariant neural network architecture for 3D point clouds"Kunsch, Robert J.We present a new rotation equivariant neural network architecture for 3D point cloud processing. Assuming that point clouds of interest are sampled from surfaces of objects in space, the network makes use of surface normals. When extracting features from the neighbourhood of a point, the point cloud is locally rotated around that point to let the surface tangent plane fit the xy-plane in relative coordinates. Feature extraction is repeated for a bunch of equidistant rotation angles within the xy-plane, we thus call the net “Surface Tangent Plane Rotation Net“ (STPRnet). This process is organized in a hierarchical manner of several layers. Processing oriented features from previous layers requires the use of spherical harmonics in new layers of the network. In the end, features that come along with geometrical information may help to reconstruct position and orientation parameters of objects. Rotation equivariance means that rotated objects can be detected by the trained network, while training data may have been given always in the same orientation. |
| http://univie.ac.at/projektservice-mathematik/e/talks/Kunsch_2021-07_bare_conf_ICCHA_kunsch.pdf |
« back