UID:
kobvindex_GFZ20231120123334
Format:
v, 175 Seiten
,
Illustrationen, Diagramme
ISBN:
978-3-7696-5301-4
,
9783769653014
ISSN:
0174-1454
Series Statement:
Wissenschaftliche Arbeiten der Fachrichtung Geodäsie und Geoinformatik der Leibniz Universität Hannover Nr. 380
Content:
Semantic segmentation is an important task in computer vision to help machines gain a high-level understanding of the environment, similar to the human vision system. For example it is used in self-driving cars which are equipped with various sensors such as cameras and 3D laser scanners to gain a complete understanding of their environment. In recent years the field has been dominated by Deep Neural Networks (DNNs), which are notorious for requiring large amounts of training data. Creating these datasets is very time consuming and costly. Moreover, the datasets can only be applied to a specific type of sensor. The present work addresses this problem. It will be shown that knowledge from publicly available image datasets can be reused to minimize the labeling costs for 3D point clouds. For this purpose, the labels from classified images are transferred to 3D point clouds. To bridge the gap between sensor modalities, the geometric relationship of the sensors in a fully calibrated system is used. Due to various errors the naive label transfer can lead to a significant amount of incorrect class label assignments in 3D. Within the work the different reasons and possible solutions are shown in order to improve the label transfer.
Note:
Auch veröffentlicht in: Deutsche Geodätische Kommission bei der Bayerischen Akademie der Wissenschaften, Reihe C, Nr. 889, München 2022, ISBN 978-3-7696-5301-4
,
Dissertation, Gottfried Wilhelm Leibniz Universität Hannover, 2022
,
Contents
1 Introduction
2 Theoretical Background
2.1 Cameras and Laserscanning
2.1.1 Cameras
2.1.2 Laserscanning
2.2 Machine Learning Fundamentals
2.2.1 Types of Learning
2.2.2 Supervised Learning - Illustrated by Decision Trees
2.2.3 Boosting
2.3 Deep Learning
2.3.1 Basics
2.3.2 Self-Attention
2.3.3 Generative Adversarial Networks
3 Related Work
3.1 Classification and Semantic Segmentation (2D)
3.2 Semantic Segmentation (3D)
3.3 Semi-Supervised Learning
3.4 Conditional Generative Adversarial Networks
3.5 Multi-View Fusion, Prediction and Labeling
3.6 Shape Completion
4 Multi-View Label Transfer and Correction
4.1 2D to 3D Label Transfer
4.1.1 Regular and Self-Occlusions
4.1.2 Dynamic Occlusions
4.1.3 Naive Label Transfer and Label Policy-Based Noise
4.2 Label Noise Correction
4.2.1 Scanstrip-Based Noise Correction
4.2.2 Semi-Supervised Scanstrip-Based Noise Correction
4.2.3 Conclusion
4.3 Multi-View Outlier Correction and Label Transfer
4.3.1 Multi-View Network
4.3.2 Label Transfer Network
4.3.3 Conclusion
5 Self-Supervised Point Cloud Rendering and Completion
5.1 Photo-Realistic Point Cloud Rendering
5.1.1 Network Architecture
5.1.2 Loss Function
5.1.3 Image Stitching
5.2 Self-Supervised Shape Completion
5.2.1 Subregion-Based GAN model
5.2.2 Loss Function
5.2.3 Network Architecture
6 Preparation of MMS data
6.1 Preprocessing of the Mobile Mapping Dataset
6.1.1 Semantic Segmentation of the MMS-Dataset
6.1.2 Human annotated MMS-Dataset
6.2 Massively Parallel Point Cloud Rendering Using Hadoop
6.3 Datasets of Self-Occluded Objects
6.3.1 Real Dataset
6.3.2 Synthetic Datasets
7 Experiments and Results for Multi-View Label Transfer
7.1 Introduction
7.2 Baseline
7.3 Training, Validation and Test Set
7.4 Scanstrip-Based Correction
7.4.1 Point-Wise Correction
7.4.2 Supervised Scanstrip-Based Correction
7.4.3 Semi-Supervised Scanstrip-Based Correction
7.4.4 Qualitative Evaluation
7.4.5 Conclusion and Discussion
7.5 Multi-View Error Correction
7.5.1 Baseline
7.5.2 Training, Validation and Test Sets
7.5.3 Training Procedure
7.5.4 Ablation Studies and Results
7.5.5 Qualitative Evaluation
7.5.6 Retraining Semantic Segmentation Network
7.5.7 Results of the Retraining Process
7.5.8 Conclusion and Discussion
7.6 Multi-View Label Transfer Learning
7.6.1 Training Procedure
7.6.2 Ablation Studies and Results
7.6.3 Qualitative Evaluation
7.6.4 Conclusion and Discussion
7.7 Summary and Conclusion
8 Experiments and Results for Self-Supervised Completion
8.1 Photorealistic Point Cloud Rendering
8.1.1 Training Procedure
8.1.2 Quantitative Evaluation
8.1.3 Qualitative Evaluation
8.1.4 Multi-View Error Correction in GAN Images
8.1.5 Conclusion and Discussion
8.2 Self-Supervised Shape Completion
8.2.1 Training Procedure
8.2.2 Quantitative Evaluation
8.2.3 Qualitative Evaluation
8.2.4 Conclusion and Discussion
9 Conclusion and Discussion
9.1 Summary and Discussion
9.1.1 Scanstrip-Based Label Error Correction
9.1.2 End-To-End Multi-View Label Transfer
9.1.3 Self-Supervised Completion
9.1.4 Conclusion
9.2 Outlook
List of Figures
List of Tables
Bibliography
Resume
Acknowledgements
,
Sprache der Kurzfassungen: Englisch, Deutsch
In:
Wissenschaftliche Arbeiten der Fachrichtung Geodäsie und Geoinformatik der Leibniz Universität Hannover, Nr. 380
Language:
English
Keywords:
Hochschulschrift
URL:
https://publikationen.badw.de/de/048254461/pdf/CC%20BY
Bookmarklink
