2021/11/12 by Jan Bednarik, Jan Bednařík, Bednarik, Jan +12 · 1 voice
Engineering · Computer Science · #3D Shape Modeling and Analysis #Computer Graphics and Visualization Techniques #Image Processing and 3D Reconstruction
paper · pdf · doi:10.48550/arxiv.2111.06838
We propose a method for unsupervised reconstruction of a\ntemporally-consistent sequence of surfaces from a sequence of time-evolving\npoint clouds. It yields dense and semantically meaningful correspondences\nbetween frames. We represent the reconstructed surfaces as atlases computed by\na neural network, which enables us to establish correspondences between frames.\nThe key to making these correspondences semantically meaningful is to guarantee\nthat the metric tensors computed at corresponding points are as similar as\npossible. We have devised an optimization strategy that makes our method robust\nto noise and global motions, without a priori correspondences or pre-alignment\nsteps. As a result, our approach outperforms state-of-the-art ones on several\nchallenging datasets. The code is available at\nhttps://github.com/bednarikjan/temporallycoherentsurfacereconstruction.\n