PRNet: Self-Supervised Learning for Partial-to-Partial Registration Yue Wang Justin M. Solomon Massachusetts Institute of Technology {yuewangx, jsolomon}@mit.edu Motivation Empowering Iterative Closest Point with self-supervised Learning. Task: Registration Given find the optimal to align them. Partial Registration Network A general framework for partial-to-partial registration tasks. Analogously to ICP, PRNet is designed to be iterative. The basic steps are as follows, Iterative/Deep Closest Point If the correspondences are known, we can define the energy function (1), centroids of and (2), and cross-covariance matrix (3). (1) (3) (2) (4) What if correspondences are unknown? ICP! It alternates between (5) and (6), (5) (6) If we use the singular value decomposition (SVD) to decompose , then, the alignment is given in closed-form by (4). Results In contrast, Deep Closest Point using neural networks to parameterize mapping function as (7) (a) Network architecture for PRNet. (b) ACP. Keypoint Detection. We observe that use L2 norm of features tend to indicate importance of a point. Gumbel-Softmax. To obtain sharp correspondence, we use Gumbel-Softmax as follows, (8) (9) Examples of scans from The Stanford Scanning Repository Keypoint detection results Correspondences for pairs of objects Code Paper GDP group