Files in this item



application/pdfSHIN-THESIS-2017.pdf (1MB)
(no description provided)PDF


Title:Learning viewer-centered projections for 3D shape completion
Author(s):Shin, Daeyun
Advisor(s):Hoiem, Derek
Department / Program:Computer Science
Discipline:Computer Science
Degree Granting Institution:University of Illinois at Urbana-Champaign
Subject(s):3D shape learning
Deep encoder-decoder networks
Vision for graphics
Multi-view reconstruction
Abstract:The goal of this study is to determine the effectiveness of different 3D shape representations in learning to generate volumetric shapes using deep neural networks. We propose to automatically reconstruct a 3D model from a single-view image of an object by synthesizing multiple depth images and inferring the volume through multi-view 3D reconstruction. The final output is a 3D mesh inferred without seeing voxels in the training process. This is similar to the intuition that humans remember (and inherently reproduce) 3D shapes without ever "seeing through" the underlying volume – we think of objects as seen from certain viewpoints and 3D structure is a derived concept. Most previous studies have focused on directly learning the voxel representations, deforming exemplars, or utilizing user interaction. In this paper, we want to learn category-independent object shape representations by simultaneously predicting multiple incomplete surfaces in relation to the viewer with the complete 3D structure in mind. Instead of predicting voxels which typically need to be in low resolution, we hypothesize that learning a representation that can consistently produce partial surfaces in a multi-task learning model enables inter-category 3D shape transfer. We perform shape completion in novel categories and evaluate quantitatively using voxel I/U and surface distance metrics. We also report that the learned representation improves 3D shape classification.
Issue Date:2017-07-20
Rights Information:Copyright 2017 Daeyun Shin
Date Available in IDEALS:2017-09-29
Date Deposited:2017-08

This item appears in the following Collection(s)

Item Statistics