Talk on predicting Surface Elevation from a single RGB satellite image with Deep Learning.
More information about the event here (in Greek).
My colleagues and I were delighted to have been accepted to talk at the 2nd workshop on Remote Sensing and Space Applications of the Geological Society of Greece. We were given a chance to present our novel approach to creating Digital Elevation Models, a domain where insufficient accuracy is often achieved by classical methods. Among other researchers, we presented our results and layed the foundation for further work on the matter, both by us and many other inspired researchers. Our findings and approach engendered much interest by the organization committee, companies and fellow researchers, in turn inspiring us to pursue our research even further.
On a more technical note, we presented our initial idea and the problem we were trying to solve. Furthermore, after presenting some basic ideas regarding Machine Learning and discrete convolutions so that the audience not familiar with these ideas could follow the rest of our presentation, we examined related work that uses Deep Learning to solve similar problems and their results for our task, indicating the need for a new approach. After delineating our system, we proceeded by displaying some of our results, both as epochs elapsed and as a 3D visualization of landscapes. Finally, we discussed possible applications, problems and limitations of our work and the next steps we are planning to take on our research.
Presentation: [odp] [pdf] (Note: .odp contains GIFs, .pdf contains only the first/last frame of those)