PhD Candidate “Deep learning based object delineation from airborne sensor data”

Updated: 4 months ago
Deadline: 09 Feb 2022

The Department of Earth Observation Science (EOS) of the Faculty of Geo-Information Science and Earth Observation (ITC), University of Twente in Enschede, the Netherlands, has a full-time vacant position for a PhD Candidate.

Your tasks

Maps are quickly becoming outdated, about 10 % of the objects change annually. This project focuses on methods to automatically interpret newly acquired sensor data to detect and to update the changed objects in the digital map. The sensor data includes high resolution 2D aerial image data and 3D laser scanner data. Our approach is to use the existing (old) digitals maps to learn how various objects appear in these 2D and 3D datasets. It is your task to design a deep learning approach to generate correct object boundaries in vector format from the airborne sensor data. Closely connected to the task of semantic segmentation of data is the delineation of the data into object boundaries. How does the boundary of an object appear in the map, and can we transfer this knowledge to draw the boundary from sensor data? What is the generalization ability of the approach for each of the object classes? Alongside this PhD position, one Postdoc will work on the generation of training data, and one other PhD Candidate will work on the semantic segmentation of the same airborne sensor data.

The specific tasks in this project are:

  • Design a method for object boundary generation from raw or classified sensor data
  • Learn and evaluate the generalization ability for different image resolutions and point densities within a dataset
  • Smart incorporation of expert knowledge for actively learning new situations

View or Apply

Similar Positions