We learn to predict correspondences by creating a self-supervised loop from a source 3D scan to a 3D mesh model. Correspondences are typically constrained to lie on the surface manifold using a 2D surface UV-map, but those have distortions and boundary discontinuities. The key idea here is to instead predict continuous correspondences in 3D to a canonical surface represented implicitly. This creates a differentiable loop without the need for UV-maps or 2D surface parameterizations.
Bharat Lal Bhatnagar, Cristian Sminchisescu, Christian Theobalt, Gerard Pons-Moll
LoopReg: Self-supervised Learning of Implicit Surface Correspondences, Pose and Shape for 3D Human Mesh Registration
in Advances in Neural Information Processing Systems (NeurIPS), 2020.
