Student information

Msc thesis subject: Remote Sensing Hollywood: Computer Generated Forests

When Pieter Claesz painted "Still Life with a Skull and a Writing Quill" in 1628, (putting an enormous amount of effort to realistically portray light interactions with all the objects in the scene), he could not imagine that you would do the same almost 400 years later (with plants and CGI – however).

You will develop a parametric plant model using Blender (open-source software). Parametric plant model equals a CGI model which grows just like a real plant: same anatomy, canopy architecture and development cycle. If successful, you will have performed a great contribution to many areas of remote sensing: radiative transfer models, synthetic data for deep-learning models, optimization of camera pose for UAV+UAG photogrammetry.

Computer-generated imagery has improved dramatically in the last few years, to the point that it is becoming increasingly difficult to distinguish in between reality and CGI (the new Star Wars, anyone?). 

In remote sensing, a pressing issue is data collection (which is cumbersome, time-consuming and costly). If only we could generate a virtual scenario with all known biological attributes (LAI, LAD, DM, gap fraction), one could simulate/optimize so many data collection protocols for the real world: what is the necessary overlap for orthomosaic? What is the minimum pixel resolution for dry matter estimation? How can one pre-train a neural network with synthetic data so it could classify between different tree-species at different physiological stages?  You can imagine that the necessary data collection for all of these tasks would most likely be impractical. However, there is one thing we can do: generate a virtual world and collect data from it.

Check the video-links and the github page to have an idea how that is possible.

After that, you will come up with a real-world example (which we could ground-truth) and transfer the “knowledge” generated from our parametric virtual models.

Objectives  

  • Come up with a real-world example (which we could ground-truth) and transfer the “knowledge” generated from our parametric virtual models.

Literature

Requirements

  • Interest on CGI, computer graphics, machine learning

Theme(s): Sensing & measuring; Modelling & visualisation