DPC: Unsupervised Deep Point Correspondence via Cross and Self Construction

We present a new method for real-time non-rigid dense correspondence betweenpoint clouds based on structured shape construction. Our method, termed DeepPoint Correspondence (DPC), requires a fraction of the training data comparedto previous techniques and presents better generalization capabilities. Untilnow, two main approaches have been suggested for the dense correspondenceproblem. The first is a spectral-based approach that obtains great results onsynthetic datasets but requires mesh connectivity of the shapes and longinference processing time while being unstable in real-world scenarios. Thesecond is a spatial approach that uses an encoder-decoder framework to regressan ordered point cloud for the matching alignment from an irregular input.Unfortunately, the decoder brings considerable disadvantages, as it requires alarge amount of training data and struggles to generalize well in cross-datasetevaluations. DPC's novelty lies in its lack of a decoder component. Instead, weuse latent similarity and the input coordinates themselves to construct thepoint cloud and determine correspondence, replacing the coordinate regressiondone by the decoder. Extensive experiments show that our construction schemeleads to a performance boost in comparison to recent state-of-the-artcorrespondence methods. Our code is publicly available athttps://github.com/dvirginz/DPC.