Multiresolution Deep Implicit Functions for 3D Shape Representation

  • 2021-09-12 19:14:51
  • Zhang Chen, Yinda Zhang, Kyle Genova, Sean Fanello, Sofien Bouaziz, Christian Haene, Ruofei Du, Cem Keskin, Thomas Funkhouser, Danhang Tang
  • 30

Abstract

We introduce Multiresolution Deep Implicit Functions (MDIF), a hierarchicalrepresentation that can recover fine geometry detail, while being able toperform global operations such as shape completion. Our model represents acomplex 3D shape with a hierarchy of latent grids, which can be decoded intodifferent levels of detail and also achieve better accuracy. For shapecompletion, we propose latent grid dropout to simulate partial data in thelatent space and therefore defer the completing functionality to the decoderside. This along with our multires design significantly improves the shapecompletion quality under decoder-only latent optimization. To the best of ourknowledge, MDIF is the first deep implicit function model that can at the sametime (1) represent different levels of detail and allow progressive decoding;(2) support both encoder-decoder inference and decoder-only latentoptimization, and fulfill multiple applications; (3) perform detaileddecoder-only shape completion. Experiments demonstrate its superior performanceagainst prior art in various 3D reconstruction tasks.

 

Quick Read (beta)

loading the full paper ...