FairFace: Face Attribute Dataset for Balanced Race, Gender, and Age

By Kimmo Kärkkäinen, Jungseock Joo. 2019

206 citations recorded by the source corpus.

Existing public face datasets are strongly biased toward Caucasian faces, and other races (e.g., Latino) are significantly underrepresented. This can lead to inconsistent model accuracy, limit the applicability of face analytic systems to non-White race groups, and adversely affect research findings based on such skewed data. To mitigate the race bias in these datasets, we construct a novel face image dataset, containing 108,501 images, with an emphasis of balanced race composition in the dataset. We define 7 race groups: White, Black, Indian, East Asian, Southeast Asian, Middle East, and Latino. Images were collected from the YFCC-100M Flickr dataset and labeled with race, gender, and age groups. Evaluations were performed on existing face attribute datasets as well as novel image datasets to measure generalization performance. We find that the model trained from our dataset is substantially more accurate on novel datasets and the accuracy is consistent between race and gender groups.

Subfields: Computer Vision and Pattern Recognition, Neurology, Evaluation & Benchmarks, Signal Processing

How researchers see this paper

The researchers whose trusted networks rank it highest, strongest advocate first. Every researcher gets a different order.

In Computer Vision and Pattern Recognition

In Neurology

In Evaluation & Benchmarks

In Signal Processing

Citations in the corpus

Learning Transferable Visual Models From Natural Language Supervision (2021, methodology): “We hypothesize that one of the reasons for this is that all the images in the Members of Congress dataset were high quality and clear, with the people clearly centered, unlike those in the FairFace dataset.”
DISCO: Dynamic and Invariant Sensitive Channel Obfuscation for deep neural networks (2020, methodology): “Datasets We conduct experiments with the following datasets: • Fairface [59] dataset consists of 108,501 images, with race, gender, and age groups.”
Algorithmic Fairness (2020, background): “However, the task becomes much more complex when a fair description of images is required, such as in multi-label classification tasks [82, 129, 134, 148], in face attribute detection [81, 121], or in the task of image captioning [67].”
Designing an Online Infrastructure for Collecting AI Data From People With Disabilities (2021, unknown): “representation of various races and genders [30].”
GPT-4V(ision) System Card (2023, methodology): “These evaluations were constructed using publicly available datasets such as FairFace [11] and Labeled Faces in the Wild [12].”
Towards causal benchmarking of bias in face analysis algorithms (2020, methodology): “The ﬁrst was trained on the CelebA dataset [49], and the second on the FairFace dataset [35].”
Taxonomizing and Measuring Representational Harms: A Look at Image Tagging (2023, background): “One distribution-based approach compares the distribution of images that depict different social groups in a training dataset to some reference distribution, typically the uniform distribution (Buolamwini and Gebru 2018; K¨arkk¨ainen and Joo 2021; Wang et al. 2021; Yang et al. 2020).”
A Review on Fairness in Machine Learning (2022, background): “However, the task becomes much more complex when a fair description of images is required, such as in multi-label classification tasks [117, 193, 202, 221], in face attribute detection [116, 177], or in the task of image captioning [96].”
One Label, One Billion Faces: Usage and Consistency of Racial Categories in Computer Vision (2021, methodology): “These categories are partially based on US Census categories and visual distinctness of the categories [28] and believed to "encompass enough of the global population to be useful" [15].”