Unsupervised Learning of Category-Specific Symmetric 3D Keypoints from Point Sets

Automatic discovery of category-specific 3D keypoints from a collection of\nobjects of some category is a challenging problem. One reason is that not all\nobjects in a category necessarily have the same semantic parts. The level of\ndifficulty adds up further when objects are represented by 3D point clouds,\nwith variations in shape and unknown coordinate frames. We define keypoints to\nbe category-specific, if they meaningfully represent objects' shape and their\ncorrespondences can be simply established order-wise across all objects. This\npaper aims at learning category-specific 3D keypoints, in an unsupervised\nmanner, using a collection of misaligned 3D point clouds of objects from an\nunknown category. In order to do so, we model shapes defined by the keypoints,\nwithin a category, using the symmetric linear basis shapes without assuming the\nplane of symmetry to be known. The usage of symmetry prior leads us to learn\nstable keypoints suitable for higher misalignments. To the best of our\nknowledge, this is the first work on learning such keypoints directly from 3D\npoint clouds. Using categories from four benchmark datasets, we demonstrate the\nquality of our learned keypoints by quantitative and qualitative evaluations.\nOur experiments also show that the keypoints discovered by our method are\ngeometrically and semantically consistent.\n