Change Detection in Graffiti Images

Abstract

The thesis presents an empirical evaluation of change detection in graffiti images, a largely unexplored domain for change detection. To conduct this evaluation three steps were conducted:1. Preparation of a dataset: As the available graffiti data is very scarce and highlycorrelated a new dataset had to be established. This was achieved by creating asynthetic dataset by adding digital graffiti to images of a graffiti wall.2. Training of the models: To increase the performance of the state-of-the-artchange detection models in the domain of graffiti images, the models were trained on graffiti data. The training was performed in different settings (finetuning and training from-scratch) as well as using different combinations of the available data. In this step, a simpler baseline model was implemented as well, to be able to evaluate the complexity of the task.3. Evaluation: Finally, the models were evaluated on the synthetic as well as on ahand-labeled real-world dataset.The evaluation showed that the original models cannot be used without finetuning for change detection in graffiti images, achieving an average F1-Score of 0.134. All models showed a significant improvement after finetuning, on verage the F1-Score increased to 0.612 for the models trained on all available data. For most models, synthetic data had an overall positive effect, especially since the precision could be improved for all models using synthetic data. The best-performing model was the baseline model, only trained on hand-labeled real-world data, achieving an F1-Score of 0.692.