The model was the easy part. The hard part was creating a dataset that represented a place we could not repeatedly fly over.
The problem
ISeeUNet was my six-month bachelor thesis on real-time change detection for drone footage in scrubland environments. The core research problem was a data problem as much as a model problem: public training datasets did not cover the conditions needed for reliable detection from a moving drone.
I created a hybrid dataset from HD source material and synthetic scenes generated with Unity and Real Virtuality 4. Synthetic data let me control scene composition, object movement, and lighting conditions that would have been expensive or unsafe to capture repeatedly in the field. I used Python, Keras, TensorFlow, and OpenCV for the training and evaluation workflow, and trained the model on a pool of Nvidia 1080 GPUs for roughly 12 GPU-months.
The build
The model was a modified U-Net architecture, chosen because pixel-level change detection needs spatial detail beyond an image-level label. I evaluated architecture and dataset choices against the conflicting constraints of detection quality and runtime cost. The final system was designed to identify small movements, tolerate changing light, and operate during omnidirectional flight. It also had to cope with camera ego-motion, which makes naive frame differencing unreliable.
Deployment mattered from the beginning. I targeted an Nvidia Jetson NX rather than treating inference hardware as an afterthought. That forced tradeoffs between model size, input processing, and real-time performance. Working with RWU and HATtec GmbH gave the thesis access to domain feedback and a more realistic benchmark for the work.
The most interesting engineering was the loop between synthetic scene generation, training, and evaluation. No single layer or loss function captured it. If the synthetic data was too clean, the model learned the renderer. If it was too close to the test footage, the evaluation stopped being an honest test of transfer. I had to keep asking which part of the result came from the model and which came from the dataset.
The result
The result was a real-time change-detection model that ran effectively on Jetson NX hardware and performed well on standard benchmarks. The thesis received a grade of 1.0. The code was later removed at a company's request, but the thesis and result videos remain in the repository as evidence of the system and its evaluation.
What I took from it
The thesis taught me to treat datasets, model architecture, and deployment hardware as one system rather than separate phases. If I continued the project, I would expand field validation, measure failure modes by weather and terrain type, and formalise the synthetic-to-real transfer evaluation. The work gave me a durable interest in computer vision that has to survive contact with a real device.