FairFace: Face Attribute Dataset for Balanced Race, Gender, and Age
Paper • 1908.04913 • Published • 1
embedding listlengths 512 512 | group_label stringclasses 7
values |
|---|---|
[
0.04989944025874138,
0.035298582166433334,
-0.05810416862368584,
0.03154600039124489,
0.04662526771426201,
-0.016130542382597923,
0.02094835229218006,
0.012997347861528397,
0.0354774035513401,
0.020288273692131042,
0.12829172611236572,
-0.017512358725070953,
0.024240585044026375,
0.0095949... | White |
[
0.05193764343857765,
0.02395002171397209,
-0.028627507388591766,
0.04929104447364807,
0.01510655228048563,
-0.018445683643221855,
-0.03821086883544922,
0.060787852853536606,
0.010436213575303555,
0.01633639447391033,
0.11381224542856216,
-0.0036038262769579887,
0.015656588599085808,
0.0138... | White |
[
0.06418692320585251,
-0.02888801321387291,
-0.03721599653363228,
0.001688881660811603,
0.0177492443472147,
-0.0500495620071888,
0.06108185276389122,
-0.009739937260746956,
0.04352611303329468,
0.013265772722661495,
0.09090138226747513,
-0.02264045923948288,
0.08787395060062408,
0.036290559... | White |
[
0.06311547756195068,
-0.004907905124127865,
0.005354320630431175,
0.03552599996328354,
0.07166098058223724,
-0.01660478115081787,
0.021231606602668762,
-0.024871960282325745,
0.041192129254341125,
-0.018212152644991875,
0.05222555994987488,
-0.02169618010520935,
0.01631261594593525,
-0.071... | Middle Eastern |
[
0.02234967239201069,
0.002808178775012493,
-0.06610208749771118,
0.020719118416309357,
0.034987810999155045,
-0.0325588621199131,
-0.014851134270429611,
0.040960893034935,
0.0032679729629307985,
0.08640623837709427,
0.11619868874549866,
0.039506662636995316,
0.012062541209161282,
-0.012816... | Latino_Hispanic |
[
0.008663332089781761,
0.023025190457701683,
-0.08859963715076447,
0.039600908756256104,
-0.028436819091439247,
-0.09399104863405228,
0.02275828644633293,
0.05445462837815285,
0.054327331483364105,
0.04138454794883728,
0.07361951470375061,
0.031142765656113625,
0.08092526346445084,
0.038400... | Indian |
[
0.020604530349373817,
0.03382868319749832,
0.013613621704280376,
0.03963566571474075,
0.011740264482796192,
0.006284855306148529,
-0.048125505447387695,
0.04789062961935997,
0.009376502595841885,
0.027778536081314087,
0.10532231628894806,
0.050674811005592346,
-0.017742063850164413,
0.0075... | White |
[
-0.0787607878446579,
-0.02124846540391445,
-0.09060321003198624,
0.02362566627562046,
0.07428907603025436,
0.014540119096636772,
0.08532827347517014,
-0.004136725328862667,
-0.018168147653341293,
0.022283751517534256,
-0.004071491304785013,
0.055111005902290344,
0.00922378245741129,
0.0355... | White |
[
0.03761313855648041,
-0.014170018024742603,
-0.056407827883958817,
0.03417835757136345,
-0.013875655829906464,
-0.025376902893185616,
0.008147058077156544,
-0.01361092459410429,
-0.01971122995018959,
0.05200137570500374,
0.049794141203165054,
0.04774767532944679,
0.01053091324865818,
0.018... | White |
[
-0.030997877940535545,
0.028067663311958313,
0.0191909521818161,
-0.023825978860259056,
0.02451685257256031,
0.02085232548415661,
0.00006777108501410112,
-0.07073034346103668,
0.05054721236228943,
-0.003778098849579692,
0.10381501913070679,
-0.003302463097497821,
0.012752051465213299,
-0.0... | White |
[
0.009895679540932178,
-0.010506493039429188,
-0.029085827991366386,
0.04456785321235657,
0.03540624678134918,
-0.028092660009860992,
0.014437340199947357,
0.005002874415367842,
0.04173891246318817,
0.04943228140473366,
0.06870487332344055,
0.016076495870947838,
0.03339913859963417,
0.00415... | Black |
[
0.046365633606910706,
0.052503831684589386,
-0.007199116982519627,
-0.0006208071717992425,
-0.033764809370040894,
0.02114863507449627,
0.0022854090202599764,
0.0097395870834589,
0.022992892190814018,
-0.010180882178246975,
-0.009936741553246975,
-0.0019346557091921568,
-0.009962580166757107,... | Latino_Hispanic |
[
0.01746322773396969,
-0.012457409873604774,
-0.015644289553165436,
-0.0015219022752717137,
0.03799651935696602,
-0.009110571816563606,
0.02499350905418396,
-0.04832496494054794,
0.007212701253592968,
0.07120005041360855,
0.16585670411586761,
-0.004803055431693792,
0.045259106904268265,
0.0... | Southeast Asian |
[
-0.05323249474167824,
0.001791134593077004,
-0.04255831241607666,
-0.0023875900078564882,
-0.025508856400847435,
-0.05462013557553291,
0.0019690480548888445,
-0.015794072300195694,
0.03286433219909668,
0.1008627861738205,
0.06682897359132767,
0.020386699587106705,
0.006861773785203695,
-0.... | White |
[
-0.05112062767148018,
-0.04429017752408981,
-0.026968248188495636,
0.004793744068592787,
-0.025131039321422577,
-0.0647548958659172,
0.08002004027366638,
0.01260481309145689,
0.018698031082749367,
0.030304772779345512,
0.09806925058364868,
0.008731042966246605,
0.048634033650159836,
0.0178... | Indian |
[
0.029468290507793427,
-0.036708056926727295,
-0.02353317104279995,
-0.0396808460354805,
0.059413276612758636,
-0.0043835449032485485,
0.030015073716640472,
0.021267449483275414,
0.020461445674300194,
0.03447950258851051,
-0.01927020214498043,
0.02813297137618065,
-0.023157736286520958,
-0.... | White |
[
-0.011702745221555233,
0.018872112035751343,
-0.0033954421523958445,
-0.0388527475297451,
0.025315793231129646,
0.019952237606048584,
0.005599041003733873,
-0.054125670343637466,
0.04514506459236145,
0.0564083606004715,
0.15156550705432892,
-0.016689835116267204,
0.02179224230349064,
-0.03... | White |
[
0.02769925445318222,
0.06625032424926758,
0.03311329334974289,
0.02213616855442524,
0.010292883962392807,
0.017692742869257927,
-0.034836750477552414,
-0.02176179364323616,
-0.01370242703706026,
-0.005470165051519871,
0.06152840331196785,
-0.03730408847332001,
-0.04017491638660431,
0.02405... | White |
End of preview. Expand in Data Studio
RAF-Ready FairFace
This repository provides the processed FairFace dataset used in the paper:
Retrieval-Augmented Dataset Assembly for Fair Clustering
The data are prepared for evaluating RAF, a retrieval-augmented dataset assembly framework for downstream fair clustering. RAF studies a data-level solution to fair clustering: instead of only imposing fairness constraints on a fixed dataset, it retrieves and selects real external samples to reduce minority under-representation before clustering.
This release contains the RAF-ready processed version of FairFace, including the query dataset, external source files, source metadata, visual embeddings, and sensitive-group labels required to run RAF.
This repository does not redistribute the full original FairFace image dataset unless explicitly permitted by the original license and release terms. It only provides the processed files used for RAF experiments. Users should also comply with the license and terms of the original FairFace dataset.
Dataset Description
Background
Fair clustering aims to produce clustering results that are both useful and fair with respect to sensitive groups. However, when the original dataset severely under-represents a minority group in some semantic regions, fair clustering algorithms are limited by the available data. They can rebalance assignments, but they cannot create missing real samples.
RAF addresses this limitation from the data side. Given an initial query dataset and a collection of external data sources, RAF selectively retrieves candidate samples under a budget constraint. It then evaluates whether each candidate improves majority-minority distributional alignment in the embedding space. The accepted samples are used to augment the initial dataset before downstream clustering.
This processed FairFace release is designed to support this setting in the visual-data domain, where different demographic groups may be unevenly distributed across visual embedding regions.
Source Dataset
The original dataset is based on FairFace, a face attribute dataset designed for balanced race, gender, and age analysis.
Original dataset:
- FairFace: https://github.com/joojs/fairface
- FairFace paper: https://arxiv.org/abs/1908.04913
- WACV paper: https://openaccess.thecvf.com/content/WACV2021/html/Karkkainen_FairFace_Face_Attribute_Dataset_for_Balanced_Race_Gender_and_Age_WACV_2021_paper.html
In RAF, we use FairFace to construct visual embeddings and source partitions for retrieval-augmented fair clustering.
Processed Dataset
This repository contains the processed RAF-ready FairFace files:
- Query dataset used as the initial dataset
D_q - External source files used as candidate data sources
- Downloads last month
- 40