Skin Tone in the Wild: A Dataset and Benchmark for 10-tone MST Classification

Better skin tone classification tools can help identify and fix bias in AI systems: The paper demonstrates that existing face recognition datasets like CelebA and VGGFace2 have "substantial representation imbalances" across different skin tones, meaning AI trained on these datasets may perform worse for people with underrepresented skin tones—a key source of algorithmic bias that can now be measured and addressed.

The 6-tone Fitzpatrick scale used in medical and AI contexts is insufficient: By introducing a 10-tone Monk Skin Tone (MST) scale with over 42,000 openly available images, the researchers show that finer-grained classification is possible and necessary, particularly for capturing "extreme skin tones" that were previously underrepresented and harder to study.

Lighting and photography conditions significantly complicate skin tone detection: The paper notes that current methods "often fail to disentangle true skin tone from in-the-wild imaging conditions," meaning poor lighting or camera settings can make AI systems misclassify skin tone—a problem the new SkinToneNet model addresses by achieving 90% accuracy even in real-world, uncontrolled settings.

Open-access datasets are critical for responsible AI development: Previous skin tone datasets either had restrictive licenses preventing research or lacked diversity, highlighting why the authors made their 42K-image dataset publicly available—enabling independent audits and broader efforts to identify and mitigate bias in commercial AI systems.