Workshop - 403

Automatic Vector Map Generation via Deep Neural Networks (DNNs): Theory and Practice

• Target audience:  Students and researchers
• Level: Beginner to intermediate
  
• Duration: 3 hours
  
• Language: English

Overview

Content

In the first section, for 45 minutes, we will explain the general concepts of DNNs including applications in Geomatics, the theoretical aspects of hidden layers, different architectures, and the training workflow such as data generation, parameter tuning and optimization.

In the second section, for 75 minutes, we will practically demonstrate how to customize a DNN for building extraction, prepare the training data and train a DNN from real data. Besides, we will investigate different challenges in practice and the effects of training parameters on the results.

In the third section, for 60 minutes, we present a sample of commercial software, “Supermap iDesktop 10” and its capabilities for vector map generation using DNNs and how to extract building vectors from real data only by a click on the software! We also discuss the practical examples and the experimental results.

Presenters Information

• Mohammad Saadatseresht¹: School of Surveying and Geospatial Engineering, College of Engineering, University of Tehran
• Fatemeh Alidoost²: Stuttgart University of Applied Sciences (HfT); vigram GmbH, Germany
• Mohsen Hosseinpour³: Supermap Comapany Agent in Iran
• Zahra Frajzadeh¹: School of Surveying and Geospatial Engineering, College of Engineering, University of Tehran
• Parsa Dajkhosh¹: School of Surveying and Geospatial Engineering, College of Engineering, University of Tehran