My bachelor's degree concluded with this project — detecting collapsed buildings using UNET++ and shadow-based heuristics from satellite imagery. It is quite expansive, so I highly recommend reading through the thesis itself if you'd like to understand it in its entirety:
📄 Thesis_Hugo_SK.pdf — the full document.
Notes & Reflections
But I can give some notes of my own, since some time has passed since I handed this in (as of the time of writing this).
Firstly I would say that this thesis was a giant jump in terms of data volume and project scope from anything I had done before. I managed to cobble it together, but if I were to do it all again, much would be different:
Lessons Learned
- Trusted my predecessor too much — since someone had given a quite successful crack at it the previous year, I thought that I would simply run what they made and expand upon it. But in the end I lost lots of time trying to fix or interpret what they did, rather than pursue my own original direction.
- Did not analyse my data enough — sure, some graphs and distributions are shown, but I should have spent at least triple the amount of time delving deep into the images, keeping the data imbalance in mind.
- Got lost in my own code — having such a large scope I ended up forgetting, not understanding and/or patching together code snippets from sources which I did not fully understand. I spent significant amounts of time walking in circles just because I had not built a solid foundation of understanding.
- Didn't carve out a test set early — with the data volume being staggeringly high, lots was wasted on errored or faulty transformations of this dataset. More dry runs and end-to-end runs on small fragments were definitely in order.
- Didn't use LaTeX — it seemed too daunting and impractical, but over the last couple of months I've grown to like it and appreciate its efficiency and looks. Will definitely use Overleaf LaTeX next time.
Takeaway
You might read all these and think — oh my god, this dude had no idea what he was doing. And you would be right. But in the end, I learned so much from these failures and shortcomings. Every time I see a dataset, or every time I inherit some architecture or review an AI code snippet, this project comes to mind. It has taught me to be more thoughtful, careful, and most importantly — patient. Taking a step back and not rushing saves you time in the long run.