Skin tone and computer vision
summary
Describes Google replacing the 6-point Fitzpatrick standard after Cloud Vision missed a darker student, developing the Monk Skin Tone Scale with Dr. Ellis Monk as a 10-bin taxonomy, running a US study (under peer review) that finds it more inclusive and comparable to a major cosmetics brand, releasing it publicly at skintone.google, and documenting practices: include representation in training and evaluation, use human annotators, and do not use as a proxy for race.
AI-written summary · about
chapters (8)
- 0:00Skin tone in computer vision
- 1:05Problems with skin tone detection
- 2:36Fitzpatrick Scale limitations
- 4:30Monk Skin Tone Scale development
- 6:23Validating the Monk Scale
- 7:51Best practices for Monk Scale
- 9:39Decoupling skin tone from race
- 10:10Summary and resources
related talks
56mGlobal Bias in AI: When Western Data Shapes the World
47mDiversity and Inclusion for Algorithmic Bias Considerations
23mMultimodal Video Face Liveness: A Better Alternative?
18mCustomizable Computer Vision Expands Data Access Without Compromising Privacy
35mComputer vision: British Clay
25mComputer Vision in Melanoma Detection
34mI, Human
49mAffective Computing – Bringing Humans and Machines Closer Through Emotions