Please use this identifier to cite or link to this item: http://hdl.handle.net/2440/120328
Type: Thesis
Title: Deeply Learned Priors for Geometric Reconstruction
Author: Weerasekera, Chamara Saroj
Issue Date: 2018
School/Discipline: School of Computer Science
Abstract: This thesis comprises of a body of work that investigates the use of deeply learned priors for dense geometric reconstruction of scenes. A typical image captured by a 2D camera sensor is a lossy two-dimensional (2D) projection of our three-dimensional (3D) world. Geometric reconstruction approaches usually recreate the lost structural information by taking in multiple images observing a scene from different views and solving a problem known as Structure from Motion (SfM) or Simultaneous Localization and Mapping (SLAM). Remarkably, by establishing correspondences across images and use of geometric models, these methods (under reasonable conditions) can reconstruct a scene's 3D structure as well as precisely localise the observed views relative to the scene. The success of dense every-pixel multi-view reconstruction is however limited by matching ambiguities that commonly arise due to uniform texture, occlusion, and appearance distortion, among several other factors. The standard approach to deal with matching ambiguities is to handcraft priors based on assumptions like piecewise smoothness or planarity in the 3D map, in order to "fill in" map regions supported by little or ambiguous matching evidence. In this thesis we propose learned priors that in comparison more closely model the true structure of the scene and are based on geometric information predicted from the images. The motivation stems from recent advancements in deep learning algorithms and availability of massive datasets, that have allowed Convolutional Neural Networks (CNNs) to predict geometric properties of a scene such as point-wise surface normals and depths, from just a single image, more reliably than what was possible using previous machine learning-based or hand-crafted methods. In particular, we first explore how single image-based surface normals from a CNN trained on massive amount of indoor data can benefit the accuracy of dense reconstruction given input images from a moving monocular camera. Here we propose a novel surface normal based inverse depth regularizer and compare its performance against the inverse depth smoothness prior that is typically used to regularize regions in the reconstruction that are textureless. We also propose the first real-time CNN-based framework for live dense monocular reconstruction using our learned normal prior. Next, we look at how we can use deep learning to learn features in order to improve the pixel matching process itself, which is at the heart of multi-view geometric reconstruction. We propose a self-supervised feature learning scheme using RGB-D data from a 3D sensor (that does not require any manual labelling) and a multi-scale CNN architecture for feature extraction that is fast and eficient to run inside our proposed real-time monocular reconstruction framework. We extensively analyze the combined benefits of using learned normals and deep features that are good-for-matching in the context of dense reconstruction, both quantitatively and qualitatively on large real world datasets. Lastly, we explore how learned depths, also predicted on a per-pixel basis from a single image using a CNN, can be used to inpaint sparse 3D maps obtained from monocular SLAM or a 3D sensor. We propose a novel model that uses predicted depths and confidences from CNNs as priors to inpaint maps with arbitrary scale and sparsity. We obtain more reliable reconstructions than those of traditional depth inpainting methods such as the cross-bilateral filter that in comparison offer few learnable parameters. Here we advocate the idea of "just-in-time reconstruction" where a higher level of scene understanding reliably inpaints the corresponding portion of a sparse map on-demand and in real-time.
Advisor: Reid, Ian
Dissertation Note: Thesis (Ph.D.) -- University of Adelaide, School of Computer Science, 2018
Keywords: SLAM
deep learning
learned priors
dense mapping
3D reconstruction
Provenance: This electronic version is made publicly available by the University of Adelaide in accordance with its open access policy for student theses. Copyright in this thesis remains with the author. This thesis may incorporate third party material which has been used by the author pursuant to Fair Dealing exceptions. If you are the owner of any included third party copyright material you wish to be removed from this electronic version, please complete the take down form located at: http://www.adelaide.edu.au/legals
Appears in Collections:Research Theses

Files in This Item:
File Description SizeFormat 
Weerasekera2018_PhD.pdf23.84 MBAdobe PDFView/Open


Items in DSpace are protected by copyright, with all rights reserved, unless otherwise indicated.