UID:
kobvindex_GFZ176793534X
Format:
155 Seiten
,
Illustrationen, Diagramme
ISBN:
978-3-7696-5279-6
,
9783769652796
ISSN:
0174-1454
Series Statement:
Wissenschaftliche Arbeiten der Fachrichtung Geodäsie und Geoinformatik der Leibniz Universität Hannover Nr. 369
Note:
Auch veröffentlicht in: Deutsche Geodätische Kommission bei der Bayerischen Akademie der Wissenschaften, Reihe C, Nr. 867, München 2021, ISBN 978-3-7696-5279-6
,
Dissertation, Gottfried Wilhelm Leibniz Universität Hannover, 2021
,
Contents
1. Introduction
1.1. Motivation
1.2. Main Contributions
1.3. Thesis Outline
2. Basics
2.1. Feature based Image Matching
2.1.1. Overview: What is Feature based Image Matching?
2.1.2. Desired Properties for Detected Features and Descriptors
2.1.3. Scale-Invariant Feature Detection
2.1.4. Feature Affine Shape Estimation
2.1.5. Feature Orientation Assignment
2.1.6. Feature Description
2.1.7. Descriptor Matching
2.2. Convolutional Neural Network (CNN)
2.2.1. Architecture of CNN
2.2.2. Training of CNN
2.3. Siamese Convolutional Neural Network
3. Related Work
3.1. Local Feature Detection
3.1.1. Translation and Rotation Invariant Features
3.1.2. Scale Invariant Features
3.1.3. Detectors based on a Comparison of Grey Values or Saliency
3.1.4. Detectors based on Machine Learning
3.2. Feature Orientation and Affine Shape Estimation
3.2.1. Orientation Assignment
3.2.2. Affine Shape Estimation
3.3. Local Feature Description
3.3.1. Hand Crafted Descriptors
3.3.2. Machine Learning based Descriptors
3.4. An Application: Orientation of Oblique Aerial Images
3.5. Discussion
3.5.1. Orientation Assignment and Affine Shape Estimation
3.5.2. Descriptor Learning
3.5.3. An Aerial Photogrammetric Benchmark
3.5.4. Ability to Transfer Learned Modules
4. Deep Learning Feature Representation
4.1. Overview of the Methodology
4.2. Descriptor Learning using Active Weak Match Finder - WeMNet
4.2.1. Descriptor Learning Architecture
4.2.2. Generation of Training Pairs
4.2.3. Loss Function
4.2.4. Weak Match Branch
4.3. Self Supervised Feature Affine Shape Learning - MoNet
4.3.1. Affine Transformation Decomposition
4.3.2. Self Supervised Affine Shape Estimation Module
4.4. Self Supervised Orientation Assignment Module - MGNet
4.5. Full Affine Estimation Network - Full-AfFNet
4.5.1. Full Affine Network
4.5.2. Training Loss
4.5.3. Data Augmentation
4.6. Inference based on the Trained Networks
4.7. Discussion
4.7.1. Descriptor Learning
4.7.2. Affine Shape Estimation
4.7.3. Orientation Assignment Learning
4.7.4. The Inference Pipeline
5. Experiments and Results
5.1. Datasets
5.1.1. Datasets for Training
5.1.2. Datasets for Testing
5.2. Evaluation and Analysis Criteria
5.2.1. Task A: Patch based Image Matching
5.2.2. Task B: Descriptor Distance Analysis
5.2.3. Task C: Feature based Image Matching
5.2.4. Task D: Image Orientation
5.2.5. Summary of Tasks and Involved Datasets
5.3. Descriptor Learning and Patch Based Image Matching
5.3.1. Parameter Study for WeMNet
5.3.2. Comparison to Related Work
5.4. Descriptor Distance Analysis
5.4.1. Translation
5.4.2. Rotation
5.4.3. Affine Shape Transformation
5.5. Image Matching Analysis
5.5.1. Parameter Study for Affine Shape Learning
5.5.2. Image Matching for Rotation Dataset
5.5.3. Image Matching for Hpatches Affine Dataset
5.6. Image Orientation
5.6.1. Determination of Image Orientation
5.6.2. Experiment Setup Details
5.6.3. Orientation Result of Different Blocks
5.6.4. Matching Quality Analysis
6. Discussion
6.1. Descriptor Learning and Patch Based Image Matching
6.1.1. Parameter Study
6.1.2. Comparison to Related Works
6.2. Descriptor Distance Analysis
6.2.1. Translation
6.2.2. Rotation
6.2.3. Affine Shape Transformation
6.3. Feature based Image Matching
6.3.1. Parameter Study
6.3.2. Rotation Set
6.3.3. Affine Set
6.4. Image Orientation
7. Conclusion and Outlook
Bibliography
A. Affine Shape Adaptation Theory
A.l. transformation of affine Gaussian scale-space
A.2. Local affine distortion measurement
A.3. More affine transformation
In:
Wissenschaftliche Arbeiten der Fachrichtung Geodäsie und Geoinformatik der Leibniz Universität Hannover
Language:
English
Keywords:
Hochschulschrift
Author information:
Chen, Lin 1987-