Neural Intrinsic Embedding for Non-Rigid Point Cloud Matching

Figure 1 . Given a point cloud, we select 5 landmarks (see red points on the right-most one) and assign each of the rest points to the cluster represented by its nearest neighbor among the landmarks in the respective embedded space. We compare our method to Euclidean coordinates, LIE [30], GPS [42] . Our method takes in only point cloud and produces segmentation that is intrinsic geometry-aware.