AIGTatt

AIGTatt was a research-style course project written in the format of a BIOSIG 2024 paper. It explored whether fully synthetic tattoo data could help address the scarcity and privacy constraints of public tattoo datasets used in biometric recognition research.

The dataset pipeline generated structured tattoo prompts with GPT-4 Turbo, produced tattoo images through SDXL and a tattoo-specific fine-tuned SDXL model, cropped the tattoo regions, and created realistic variations with rotation, scale, brightness, noise, and other augmentations. The final study used 250 isolated tattoo identities and 5,000 variations.

The retrieval experiments trained EfficientNetV2 and Swin embedding models with ArcFace-style objectives, then evaluated them on real-world tattoo datasets including WebTattoo and BIVTatt. The work showed that synthetic tattoo data can bootstrap recognition systems, while also making clear that larger and more diverse datasets are needed for stronger generalization.