Call number:
S 99.0139(378)
In:
Wissenschaftliche Arbeiten der Fachrichtung Geodäsie und Geoinformatik der Leibniz Universität Hannover, Nr. 378
Type of Medium:
Series available for loan
Pages:
viii, 117 Seiten
,
Illustrationen, Diagramme
ISSN:
0174-1454
Series Statement:
Wissenschaftliche Arbeiten der Fachrichtung Geodäsie und Geoinformatik der Leibniz Universität Hannover Nr. 378
URL:
https://d-nb.info/1264856385/04
Language:
English
Note:
Dissertation, Fakultät für Bauingenieurwesen und Geodäsie der Gottfried Wilhelm Leibniz Universität Hannover, 2021
,
1 Introduction
1.1 Problem Statement
1.2 Contributions
1.3 Thesis Outline
2 Basics
2.1 Dense Stereo Matching
2.1.1 Terminology and Practical Simplifications
2.1.2 Taxonomy of the Matching Process
2.1.3 Challenges and Common Assumptions
2.2 Uncertainty Quantification
2.3 Deep Learning
2.3.1 Convolutional Neural Networks
2.3.2 Bayesian Neural Networks
3 Related Work
3.1 Dense Stereo Matching
3.2 Aleatoric Uncertainty Estimation
3.3 Epistemic Uncertainty Estimation
3.4 Discussion
4 Uncertainty Estimation for Dense Stereo Matching - A New Method
4.1 Overview
4.2 Aleatoric Uncertainty Estimation
4.2.1 CNN-based Cost Volume Analysis
4.2.2 Uncertainty Models
4.3 Epistemic Uncertainty Estimation
4.3.1 Functional Model
4.3.2 Stochastic Model
4.4 Joint Uncertainty Estimation
4.5 Discussion
5 Experimental Setup
5.1 Objectives
5.2 Datasets
5.3 Training and Hyper-parameter Settings
5.3.1 General Remarks
5.3.2 CVA-Net
5.3.3 Probabilistic GC-Net
5.3.4 Combined Approach
5.4 Evaluation Strategy andCriteria
5.4.1 Disparity Error Metrics
5.4.2 Confidence Error Metric
5.4.3 Uncertainty Error Metric
5.4.4 Region Masks
5.4.5 Monte Carlo Sampling
6 Results and Discussion
6.1 CVA-Net Architecture
6.2 Aleatoric Uncertainty Models
6.3 Dense Stereo Matching using a Bayesian Neural Network
6.3.1 Comparison to the Deterministic Baseline
6.3.2 On the Relevance of Aleatoric and Epistemic Uncertainty
6.3.3 The Kullback-Leibler Divergence and the Mode Collapse Problem
6.4 Discussion
7 Conclusions and Outlook
Bibliography
Acknowledgment
Location:
Lower compact magazine
Branch Library:
GFZ Library
Permalink
