Semi-synthetic Data Generation for Tattoo Segmentation

Tattoos have been successfully employed to assist law enforcement in the identification of criminals and victims. Due to various privacy issues in acquiring images containing tattoos, only a limited number of databases exist. This lack of databases has slowed down the development of new tattoo segmentation and retrieval methods. In our work, we propose a new unsupervised generator that allows generating a large number of semi-synthetic images with tattooed subjects. To successfully generate realistic images, a database including the respective skin segmentation map is also proposed. Using this new generator and the skin database, 5,500 semi-synthetic images were created and evaluated for the tattoo segmentation use case. Experimental results on real data show the usefulness of using semi-synthetic images to train semantic segmentation algorithms: several manually mislabelled real samples were successfully corrected. The tattoo generator code, the skin database and generated images have been made available at https://dasec.h-da.de/hda-sstd/.