At FZI Forschungszentrum Informatik I worked on the LDFA anonymization pipeline for self-driving data: detecting German license plates and anonymizing them before storage or sharing.
Approach
- Segment Anything (SAM): Fine-tuned Meta’s SAM for license plate regions in our datasets.
- Metrics: Lifted detection accuracy from 75% to 95% and cut false positives by 15%.
- Synthetic data: Used Flux Inpainting to generate synthetic plates for training and augmentation.
- Experiment tracking: W&B for runs, hyperparameters, and model versions.
The pipeline is used in research and data preparation for autonomous driving at FZI.