Intermediate Report

Jul 1, 2022

Introduction

Introduction

Fairness evaluation studies across Biometric Modalities

Studies

Fairness in Face Recognition

Fairness

Ocular : Motivation

Ocular Motivation

Research Objectives

Research

Covariates to be Studied

Covariates

Datasets Used

Datasets

Gender Classification

Gender Classification

Across Models

Across Models on UFPR Periocular - RGB

Across Models UFPR

Across Models on Notredame Ocular - NIR

Across Models Notredame

Across Varying Data Balances and Gender

UFPR - RGB

UFPR Data Balance 1

Notredame - NIR

Notredame Data Balance 1

Across Age

UFPR - RGB

TABLE I: Accuracy of CNN-based Gender Classification on Left Ocular Region among Young (18 to 39 years), Middle (40 to 59 years) and Older Adults (60 to 79 years).
CNN Young Middle-Aged Older
Male[%] Female[%] Male[%] Female[%] Male[%] Female[%]
ResNet-50 98.39 98.19 100 96.67 99.17 98.06
MobileNet-V2 99.97 99.9 100 99.7 100 100
ShuffleNet-V2-50 98.23 97.57 98.28 97.56 94.76 98.89
EfficientNet-B0 95.89 97.54 98.58 96.44 95.23 86.94

  • Middle Aged adults slightly outperformed the other two groups in gender classification by about 1% − 2%.
  • Possible explanation: Stable and distinct gender cues for middle aged adults when compared to young and older adults.
  • Younger adult population performed the best in age classification by about 25%.
  • This could be due to distinct variation in the features attributed to the growing stage of the youth population over middle-aged and older adults

Across Race

VISOB + Notredame- Ocular RGB

TABLE II: Gender Classification results across race groups - Ocular Datasets.
Race Male Female
White 82.42 89.92
South Asian 96.51 88.88
Black 78.57 72.10
Middle Eastern 71.12 93.08
Latino 84.83 91.61
  • South Asians and Caucasians overall works better.
  • This make sense since it contains the majority of the dataset

Face VS Ocular

FlickrFaceHQ-Aging - RGB

  • Face performs better than Ocular in Gender Classification on RGB
  • This may be because of the low quality of the crop
TABLE III: Gender Classification results across Face vs Periocular on FFHQ-Aging.
Male Female Modality
97.6 96.8 Face
92.5 91.6 Ocular

Notredame - NIR

Face

Notredame Face

Ocular

Notredame Ocular

  • Face performs better than Ocular in Gender Classification on NIR
  • This may be because of the low quality of the crop

Subject Verification

Subject Verification

Across Gender

UFPR - RGB

Training Set : All

Training Set : Male

Training Set : Female

Notredame - NIR

Training Set : All

Training Set : Male

Training Set : Female

VISOB - RGB

Across Varying Data Balances and Models

UFPR - RGB

UFPR Subject 1 UFPR Subject 2

Notredame - NIR

Notredame Subject

Across Age

UFPR - RGB

TABLE I: EER and FNMR at 0.01 and 0.1 FMR for user authentication using CNN models for Left (L), Right (R) ocular region,and their score-level fusion (L+R) for Young, Middle-Aged and Older adults evaluated on balanced version of UFPR ocular datasets.
CNN
 Age-Group
 EER(%)   FNMR(%) @ FMR
0.01 0.1
L R L+R L R L+R L R L+R
ResNet-50 Young 8.60 9.52 9.06 37.63 35.35 36.49 53.74 54.03 53.89
Middle-Aged 8.62 9.08 8.85 30.76 25.04 27.9 52.23 48.55 50.39
Older 11.01 11.00 11.01 15.47 19.34 17.405 30.67 30.67 30.67
MobileNet-V2 Young 7.75 7.32 7.54 33.21 26.11 29.66 51.61 48.28 49.95
Middle-Aged 9.08 8.68 8.88 29.18 28.14 28.66 51.17 51.31 51.24
Older 8.04 9.63 8.84 18.54 13.47 16.005 39.47 36.00 37.74
ShuffleNet-V2 Young 6.93 6.96 6.95 37.63 38.09 37.86 56.44 56.26 56.35
Middle-Aged 8.32 9.25 8.79 37.38 46.07 41.73 55.08 61.10 58.09
Older 9.72 8.18 8.95 30.53 30.80 30.67 44.93 53.47 49.20
EfficientNet-B0 Young 7.64 9.61 8.63 32.54 27.00 29.77 55.40 48.38 51.89
Middle-Aged 6.95 9.03 7.99 38.07 26.69 32.38 49.12 49.31 49.22
Older 9.34 12.08 10.71 20.80 23.33 22.065 39.73 41.74 40.74
  • Younger adults obtained performance identical to middle-aged individuals in user verification.
  • Older adults’ performance differs slightly in terms of EER with only a 1% decrease, but the performance dropped at lower FMR points.
  • The possible reason could be due to likely inferior quality of image capture, and relatively higher inter-class similarity due to wrinkles and folds on the skin.
Sreeraj Ramachandran
Sreeraj Ramachandran
Graduate Research Assistant

My research interests include Computer Vision, Biometrics, Bias AI, GANs and adversarial Attacks

comments powered by Disqus