2020/03/24 by Rohan Chabra, Chabra, Rohan, Jan Eric Lenssen +11 · 5 citations
Computer Science · Engineering · #3D Shape Modeling and Analysis #Advanced Vision and Imaging #Computational Geometry (cs.CG) #Computer Graphics and Visualization Techniques #Computer Vision and Pattern Recognition (cs.CV) #FOS: Computer and information sciences #Machine Learning (cs.LG)
paper · pdf · doi:10.48550/arxiv.2003.10983
openalex publication_date 2020/03/24 · openalex created_date 2022/07/26 · openalex updated_date 2026/07/28
Efficiently reconstructing complex and intricate surfaces at scale is a\nlong-standing goal in machine perception. To address this problem we introduce\nDeep Local Shapes (DeepLS), a deep shape representation that enables encoding\nand reconstruction of high-quality 3D shapes without prohibitive memory\nrequirements. DeepLS replaces the dense volumetric signed distance function\n(SDF) representation used in traditional surface reconstruction systems with a\nset of locally learned continuous SDFs defined by a neural network, inspired by\nrecent work such as DeepSDF. Unlike DeepSDF, which represents an object-level\nSDF with a neural network and a single latent code, we store a grid of\nindependent latent codes, each responsible for storing information about\nsurfaces in a small local neighborhood. This decomposition of scenes into local\nshapes simplifies the prior distribution that the network must learn, and also\nenables efficient inference. We demonstrate the effectiveness and\ngeneralization power of DeepLS by showing object shape encoding and\nreconstructions of full scenes, where DeepLS delivers high compression,\naccuracy, and local shape completion.\n