Investigating Bias and Fairness in Appearance-based Gaze Estimation

Burak Akgül; Erol Şahin; Sinan Kalkan

cs.CVApr 12, 2026

Investigating Bias and Fairness in Appearance-based Gaze Estimation

Authors: Burak Akgül, Erol Şahin, Sinan Kalkan

Organizations: Dept. of Computer Eng. and ROMER Robotics Center, Middle East Technical University, Ankara, Turkey

Abstract

While appearance-based gaze estimation has achieved significant improvements in accuracy and domain adaptation, the fairness of these systems across different demographic groups remains largely unexplored. To date, there is no comprehensive benchmark quantifying algorithmic bias in gaze estimation. This paper presents the first extensive evaluation of fairness in appearance-based gaze estimation, focusing on ethnicity and gender attributes. We establish a fairness baseline by analyzing state-of-the-art models using standard fairness metrics, revealing significant performance disparities. Furthermore, we evaluate the effectiveness of existing bias mitigation strategies when applied to the gaze domain and show that their fairness contributions are limited. We summarize key insights and open issues. Overall, our work calls for research into developing robust, equitable gaze estimators. To support future research and reproducibility, we publicly release our annotations, code, and trained models at: github.com/akgulburak/gaze-estimation-fairness

Figures & tables

Fig. 1 : Contributions of the paper for fairness in gaze estimation. Contribution 1: Annotating the selected datasets with sensitive attributes. Contribution 2: Analysis of fairness between demographic groups, using the sensitive attribute annotations. Contribution 3: Applying bias mitigation methods to reduce bias and achieve fairness.

Fig. 2 : Overall pipeline for bias mitigation methods and fairness calculations.

Fig. 3 : The ethnicity and gender annotation pipeline. A deep learning model (Fairface [ 30 ] ) is used to annotate the subjects. Conflicting annotations are validated by the human annotator.

	Ethnicity				Gender
Dataset	Cau.	Afro-A.	Asian	Unsure	Male	Female	Total
Gaze360	118,650	12,950	64,041	1,947	103,805	93,783	197,588
GazeCapture	1,827,875	178,250	241,127	198,252	1,244,510	1,200,994	2,445,504

TABLE I : Exp 1: The distribution of samples across demographic groups for the two datasets. Cau: Caucasian, Afro-A: Afro-American.

TABLE II : Exp 2: Analysis of bias in existing gaze estimation models. WS and KS do not have accepted thresholds to identify fairness whereas

1-p\leq 0.95

indicates fairness. Red and green numbers indicate bias and fairness, respectively. ∗ denotes models that are trained on Gaze360 dataset and tested on GazeCapture dataset. To make it easier to follow the tables, we make all fairness metrics lower-better by converting the

p

-value metric and its threshold (

p>0.05

) into

1-p\leq 0.95

TABLE III : Exp 3: Bias mitigation results for oversampling experiment. WS and KS do not have accepted thresholds to identify fairness whereas

1-p\leq 0.95

indicates fairness. Red and green numbers indicate bias and fairness, respectively.

{\color[rgb]{1,0,0}\medtriangleup}

{\color[rgb]{0,0.3906,0}\medtriangledown}

{\color[rgb]{0.5,0.5,0.5}\mathbf{\sim}}

denote respectively improvement, reduction or no-change in fairness compared to Table II . ∗ denotes models that are trained on Gaze360 dataset and tested on GazeCapture dataset. To make it easier to follow the tables, we make all fairness metrics lower-better by converting the

p

-value metric and its threshold (

p>0.05

) into

1-p\leq 0.95

TABLE IV : Exp 3: Bias mitigation results for resampling experiment. WS and KS do not have accepted thresholds to identify fairness whereas

p\leq 0.95

indicates fairness. Red and green numbers indicate bias and fairness, respectively.

{\color[rgb]{1,0,0}\medtriangleup}

{\color[rgb]{0,0.3906,0}\medtriangledown}

{\color[rgb]{0.5,0.5,0.5}\mathbf{\sim}}

p

-value metric and its threshold (

p>0.05

) into

1-p\leq 0.95

TABLE V : Exp 3: Bias mitigation results for loss reweighting experiment. WS and KS do not have accepted thresholds to identify fairness whereas

p\leq 0.95

indicate fairness. Red and green numbers indicate bias and fairness, respectively.

{\color[rgb]{1,0,0}\medtriangleup}

{\color[rgb]{0,0.3906,0}\medtriangledown}

{\color[rgb]{0.5,0.5,0.5}\mathbf{\sim}}

p

-value metric and its threshold (

p>0.05

) into

1-p\leq 0.95

Model	Oversampling	Resampling	Loss Reweighting
CrossGaze	12	14	19
MCGaze	9	11	14
L2CS-Net	16	16	18
PureGaze	18	23	16
GazeTR	13	13	10
Total	68	77	77

TABLE VI : Exp 3: The number of times each method improves fairness across any of the measures. Each improvement on a dataset with respect to a measure is counted independently as one (1).

Gaze Estimation

Gaze estimation predicts where a person is looking from eye, face, head-pose, and scene cues.

See all 66 papers

Is Geometry Enough? An Evaluation of Landmark-Based Gaze Estimation
Mar 25, 2026Daniele Agostinelli, Thomas Agostinelli, Andrea Generosi +1Gaze Estimation Domain Generalization
Department of Industrial Engineering and Mathematical Sciences, Universit`a Politecnica delle Marche, via Brecce Bianche, 12, Ancona, 60131, Italy · Department of Science and Information Technology, Universit`a Pegaso, piazza Trieste e Trento, 48, Napoli, 80132, Italy
Real-time Appearance-based Gaze Estimation for Open Domains
Mar 27, 2026Zhenhao Li, Zheng Liu, Seunghyun Lee +2Gaze Estimation Domain Generalization
Huawei Technologies Canada · University of Toronto
PAGE: Towards Practical Human-level Gaze Target Estimation
Jul 6, 2026Zhoutong Ye, Chengwen Zhang, Zhaibin Cui +10Gaze Estimation
1Tsinghua University · 2Jinan University

Mar 25, 2026J/K move · Enter open · S save

Is Geometry Enough? An Evaluation of Landmark-Based Gaze Estimation

Daniele Agostinelli, Thomas Agostinelli, Andrea Generosi, Maura Mengoni

Department of Industrial Engineering and Mathematical Sciences, Universit`a Politecnica delle Marche, via Brecce Bianche, 12, Ancona, 60131, Italy · Department of Science and Information Technology, Universit`a Pegaso, piazza Trieste e Trento, 48, Napoli, 80132, Italy

Gaze Estimation Domain Generalization

Open paper AskPDF arXiv

Appearance-based gaze estimation frequently relies on deep Convolutional Neural Networks (CNNs). These models are accurate, but computationally expensive and act as "black boxes", offering little interpretability. Geometric methods based on facial landmarks are a lightweight alternative, but their performance limits and generalization capabilities remain underexplored in modern benchmarks. In this study, we conduct a comprehensive evaluation of landmark-based gaze estimation. We introduce a standardized pipeline to extract and normalize landmarks from three large-scale datasets (Gaze360, ETH-XGaze, and GazeGene) and train lightweight regression models, specifically Extreme Gradient Boosted trees and two neural architectures: a holistic Multi-Layer Perceptron (MLP) and a siamese MLP designed to capture binocular geometry. We find that landmark-based models exhibit lower performance in within-domain evaluation, likely due to noise introduced into the datasets by the landmark detector. Nevertheless, in cross-domain evaluation, the proposed MLP architectures show generalization capabilities comparable to those of ResNet18 baselines. These findings suggest that sparse geometric features encode sufficient information for robust gaze estimation, paving the way for efficient, interpretable, and privacy-friendly edge applications. The source code and generated landmark-based datasets are available at https://github.com/daniele-agostinelli/LandmarkGaze.git.

Abstract

Gaze Estimation

Explore similar work

Is Geometry Enough? An Evaluation of Landmark-Based Gaze Estimation

Real-time Appearance-based Gaze Estimation for Open Domains

PAGE: Towards Practical Human-level Gaze Target Estimation