Exploring rich intermediate representations for reconstructing 3D shapes from 2D images
作者：
Highlights：
• Embed the rich intermediate representations to the 3D reconstruction network to mine the information about shape priors and visible surface geometry of the targets in 2D images respectively.
• Combine the global priors and visible surface geometry information to estimate the invisible part of the 3D object as well as perform high-quality 3D reconstruction.
• Propose the geometric feature compensation generation unit to estimate the complete geometric structure information of the target in a single view image.
• Propose the shape transformation unit to model the shape transformation as the feature compensation learning process in the feature space to directly deform the three-dimensional voxel.
摘要
•Embed the rich intermediate representations to the 3D reconstruction network to mine the information about shape priors and visible surface geometry of the targets in 2D images respectively.•Combine the global priors and visible surface geometry information to estimate the invisible part of the 3D object as well as perform high-quality 3D reconstruction.•Propose the geometric feature compensation generation unit to estimate the complete geometric structure information of the target in a single view image.•Propose the shape transformation unit to model the shape transformation as the feature compensation learning process in the feature space to directly deform the three-dimensional voxel.
论文关键词：3D Reconstruction,Shape transformation,Intermediate representations
论文评审过程：Received 17 March 2021, Revised 2 August 2021, Accepted 31 August 2021, Available online 1 September 2021, Version of Record 10 September 2021.
论文官网地址：https://doi.org/10.1016/j.patcog.2021.108295
