Internship: deep learning on aerial imagery
Readar was founded in 2016 and extracts all kind of information from aerial imagery. To do this we combine remote sensing with machine learning. Our customer base is very diverse: from grid operator to insurance companies. The knowledge we extract from our data has a strong environmental impact: as representative examples, we map solar panels to support the sustainable energy revolution and we help municipalities in banning asbestos. We combine the dynamics of a start-up with the professionalism of our established customers! The internship assignment is part of our international expansion plan.
About you and the project:
Readar has build methods to detect buildings in aerial imagery, to infer create height data from aerial imagery and to use this height data to create 3D polygon models of buildings. In the first two steps we successfully introduced CNN’s to replace classic remote sensing methods we developed earlier. We are now looking at the third step: how to create 3D polygons from RGB+Z data. We have a benchmark method which we have been using for the past 3 years but we are convinced that a CNN can perform better on this instance segmentation task. Test and training data is available and consists of 10cm RGB imagery + corresponding height data per pixel of the Netherlands and 3D building polygon models, both created from our previous method and hand drawn. We want you to focus on the segmentation task, we already have the data loaders for the imagery available and we have the tools to convert segmented imagery into polygon models. You will select promising architectures during literature review. From this selection you will perform a comparative study with respect to the benchmark and and propose an addition/alteration/improvement which is scientifically new. The architectures will be tested on our dataset. This research is new in our field of remote sensing/ GIS, therefore we think that publication of the result is feasible. Readar will use the result of the thesis to improve the existing 3D building models which are a.o. used to plan PV installations on buildings.
To make this a success you need:
- to be enthusiastic about this assignment.
- a “can be done mentality”, you really like to tackle this challenge.
- Programming experience in Python
- Experience with a deep learning library like Tensorflow.
If you do not already have experience in deep learning please state this clearly in your application, we have another assignment available.
For questions or applications please contact Erik Heeres: firstname.lastname@example.org