Abstract
We propose a method for unsupervised reconstruction of atemporally-consistent sequence of surfaces from a sequence of time-evolvingpoint clouds. It yields dense and semantically meaningful correspondencesbetween frames. We represent the reconstructed surfaces as atlases computed bya neural network, which enables us to establish correspondences between frames.The key to making these correspondences semantically meaningful is to guaranteethat the metric tensors computed at corresponding points are as similar aspossible. We have devised an optimization strategy that makes our method robustto noise and global motions, without a priori correspondences or pre-alignmentsteps. As a result, our approach outperforms state-of-the-art ones on severalchallenging datasets. The code is available athttps://github.com/bednarikjan/temporally_coherent_surface_reconstruction.