Investigating Bias and Fairness in Appearance-based Gaze Estimation
Listen
Radio episode about this paper
Transcript
Introduction to the show: ident: AI Radio. Generated commentary on the latest Artificial Intelligence papers.
Tom: We're looking at the paper: "Investigating Bias and Fairness in Appearance-based Gaze Estimation".
Jane: Here's the gist of the paper. Appearance-based gaze estimation methods suffer from significant fairness issues across different demographic groups and this paper presents the first extensive evaluation of fairness in appearance-based gaze estimation focusing on ethnicity and gender attributes.
Tom: Research Gaps Addressed The paper addresses several research gaps concerning bias and fairness in this domain.
Jane: The identified gaps include: 1.
Tom: Many high-stake applications, e.g., driver monitoring systems, can lead to life threatening outcomes owing to biases across demographic groups.
Title and authors: Tom: So, we're diving into this paper called "Investigating Bias and Fairness in Appearance-based Gaze Estimation". It's by Burak Akgul, Erol S¸ahin, and Sinan Kalkan. Basically, they’re looking at how systems that guess where someone is looking based just on their appearance are working for different groups of people.
Jane: That sounds really important because we all use these kinds of tools now, like in driver monitoring or other applications. What's the main takeaway from this paper about what they found?
Tom: Well, the core issue they point out is that appearance-based gaze estimation models have significant fairness problems across different groups, specifically focusing on ethnicity and gender. They show that accuracy isn't enough if you don't check for bias in these systems.
Lu: What did they do to find this bias? Did they just look at a few models, or did they build something big?
Tom: They set up a fairness baseline by taking several state-of-the-art models and running them against standard fairness metrics after annotating the existing datasets with gender and ethnicity labels. It showed significant performance disparities.
Jane: That brings up the question of data, right? If you can't measure what you're trying to fix, how can you fix it? What are the specific gaps this paper is trying to fill for us as researchers?
Tom: They identified three main research gaps. First, high-stakes applications like driver monitoring could lead to dangerous outcomes because of these biases across different demographic groups. Second, there isn't a study out there that focuses specifically on mitigating these bias issues in gaze estimation yet. Third, a big reason for both of those is that the current datasets lack demographic labels; they only have the gaze label itself.
Meng: So if we can't get the labels, how can anyone even start to train something fair? That sounds like a major hurdle for any real-world deployment.
Jane: Exactly. The paper tackles this by suggesting some concrete improvements to the current process of training these models fairly, so let's look at what they propose next.
Title and authors: Tom: They suggest several ways to improve things, starting with data annotation for fairness itself. They propose annotating existing datasets like Gaze360 and GazeCapture with gender and ethnicity attributes using a deep learning model called Fairface to generate initial labels before human annotators validate them.
Lu: That makes sense for getting those initial labels, but what about fixing the actual training process once you have those annotations? What are they testing for mitigation strategies?
Tom: They systematically analyze bias in the gaze predictions from several state-of-the-art models—CrossGaze, MCGaze, L2CS-Net, PureGaze, and GazeTR—on both the Gaze360 and GazeCapture datasets to really highlight these severe fairness issues across ethnicity and gender.
Meng: Which specific techniques did they test for fixing this bias? Are we talking about simple tweaks or something more complex? Because on my end, I need to know what's feasible to deploy without needing massive retraining cycles.
Tom: They establish a baseline benchmark by evaluating several preprocessing and in-processing bias mitigation algorithms against the identified biases in gaze estimation regression problems. Then they apply oversampling—to make minority group sample probabilities equal to the majority group—and loss reweighting, where the loss value is increased for a minority group with respect to the inverse of the sample size.
Jane: Wait, so they found that standard bias mitigation techniques don't really work very well for gaze regression, which is a surprising result considering how much this technology has improved.
Lu: They actually note that "resampling and loss reweighting methods provide the most number of improvements across models" when looking at the results. And they even suggest prioritizing models like PureGaze when using resampling and loss reweighting on Gaze360 data, because it consistently benefits from resampling across both ethnicity as well as gender labels on that specific dataset.
Meng: If I'm thinking about deploying this, would I lean towards using PureGaze if I have to use that resamplng and loss reweighting strategy? What are the practical trade-offs there?
Title and authors: Tom: The paper highlights open issues too, though. It calls for research into developing more robust, equitable gaze estimators and designing more effective bias mitigation strategies because standard techniques just didn't provide the solution they were looking for.
Jane: So to wrap up what we’ve heard about this work on "Investigating Bias and Fairness in Appearance-based Gaze Estimation," it seems the paper provides a comprehensive look at where appearance-based systems are falling short regarding fairness across ethnicity and gender.
Tom: It shows that just having good accuracy isn't enough; you have to actively measure and mitigate these biases using things like resampling and loss reweighting, especially when dealing with datasets that don't have demographic labels.
Lu: The implication is that for any serious system in this area, we need to move past just training on the raw data and start incorporating these explicit fairness checks during the entire pipeline.
Meng: From a deployment standpoint, it’s clear that if we want to ship something reliable, we have to invest time into getting those demographic labels and then running those specific mitigation strategies.
Jane: It really boils down to this: as we move appearance-based gaze estimation into real applications, the focus has shifted from just how accurate the prediction is to how fair that prediction is for everyone.
Tom: Yeah, they clearly laid out that standard techniques fall short, so we have a clear direction on what needs more research in geometric or domain-specific solutions instead of just general fixes.
Lu: That opens up a lot of avenues for representation learning research focused specifically on how those sensitive attributes interact with the model's internal representations.
Meng: It’s interesting to hear that PureGaze showed consistent benefit, because if we can find a model architecture that naturally handles this better, it could save us a lot of work down the line.
Jane: That's what I wanted to say. The paper "Investigating Bias and Fairness in Appearance-based Gaze Estimation" gives us a clear roadmap for how to start tackling these fairness challenges head-on in gaze estimation research.
The paper's summary: Tom: So, to recap, this paper is looking at how gaze estimation models—those that guess where someone is looking just from their face—are failing different groups of people when we look at things like ethnicity and gender.
Jane: Right. It’s not just about whether the model guesses the direction correctly; it’s about whether its performance is fair across different populations.
Tom: Exactly. They found that these appearance-based systems have serious fairness problems, and they point out that this is a big deal because we're putting these things in life-or-death situations, like driver monitoring systems.
Lu: The real gap they’re addressing is the lack of labels on the data itself. Right now, most gaze datasets just tell you where the gaze is, but they don't tell you anything about who that person is.
Meng: So if we can't label it by gender or ethnicity in our training set, how can we even start to fix the bias?
Jane: They address that by suggesting a two-step approach. First, they propose using a deep learning model called Fairface to create initial labels for gender and ethnicity on existing datasets like Gaze360 and GazeCapture before human annotators check them.
Tom: That’s the data annotation fix. But then they move into testing the models themselves, checking state-of-the-art ones like CrossGaze or PureGaze against these biases using standard fairness metrics.
Lu: They established a baseline for fairness by seeing how those existing models behave when you introduce gender and ethnicity as factors in their predictions across both datasets. It highlights how severe the disparities are.
Meng: And then they test mitigation strategies, like oversampling or loss reweighting, to see if we can actually reduce that bias during training.
Jane: They found that oversampling and loss reweighting are the most effective methods for reducing bias in these gaze estimation problems across all the models they tested.
Tom: And they even gave a specific recommendation: when using those techniques on Gaze360 data, you should prioritize the PureGaze model because it consistently shows improvement across both ethnicity and gender labels.
Lu: The implication here is that we can't just train these models and hope for the best; we have to actively inject fairness checks into the data labeling process and the training algorithm itself.
Jane: So, for someone listening who isn't in a lab, it means that any system built to look at someone’s eyes needs a plan to make sure it doesn't unfairly misinterpret what different people are looking at.
Tom: It’s about moving beyond just accuracy and thinking about what kind of outcome is fair for every single person interacting with the system.
Lu: This points toward needing more research into how we can structure the representations inside the AI so that gender and ethnicity don't accidentally get baked into how it interprets gaze.
Jane: It sets up a big question for us: if we can identify these biases this clearly, what does that mean for the next generation of vision models?
The paper's improvements: Tom: We just talked about how these models are biased, and now we’re looking at what they actually suggest we do to fix it in practice.
Jane: So, what’s the roadmap for getting this fair? What specific steps are they recommending for people trying to build or use these gaze estimation tools?
Tom: They outline a few concrete improvements. First, they suggest using that Fairface model we mentioned earlier to get better initial labels on the existing datasets.
Lu: That’s the data preparation step again, making sure our starting point isn't inherently biased before we even start training anything new.
Jane: Then they focus on in-processing techniques like oversampling and loss reweighting during training to actively pull those biases out of the model's learning process itself.
Tom: They found that this combination of resampling and loss reweighting actually gave the most consistent improvements across all the different models they tested.
Meng: So, if I’m trying to deploy something fast, does that mean I just run one of these training routines and it works reliably? What about cost?
Lu: Well, they did suggest prioritizing a model like PureGaze when using those specific resampling and loss reweighting methods on the Gaze360 dataset because it showed better fairness gains across both gender and ethnicity labels.
Jane: That’s a practical hint for anyone trying to pick which AI architecture to start with when you're dealing with sensitive attributes.
Tom: The main implication is that we need a structured pipeline: label the data fairly, then train the model using these specific weighting strategies, and maybe even choose a better base model like PureGaze if you can.
Lu: They’re not just tweaking one number; they’re looking at how preprocessing and in-processing methods interact to get the best fairness outcome for a regression problem like gaze estimation.
Jane: This means that the next step for researchers isn't just building a more accurate model, but building a more systematically fair one from the ground up.
Tom: It shifts our focus from just getting high accuracy numbers on a test set to ensuring those numbers hold up when you split the data by ethnicity or gender.
Lu: And that opens up a lot of questions about how we can design representations inside the AI so that these sensitive attributes don't accidentally get baked into how it interprets what someone is looking at.
Jane: It’s clear this paper gives us a practical set of tools to start fixing these problems right now, and it also points toward the bigger research question of representation learning itself.
Conclusion: Tom: So we've covered how appearance-based gaze estimation has serious fairness issues across ethnicity and gender, and what methods they propose to fix them in practice by using things like oversampling and loss reweighting.
Jane: It boils down to this: accuracy isn't enough; we have to actively measure and mitigate these biases during the entire training pipeline.
Lu: The main implication is that we can't just train these models and hope for the best; we have to inject fairness checks into the data labeling process and the training algorithm itself.
Meng: It means for production, if we want something reliable, we have to invest time into getting those demographic labels and then running those specific mitigation strategies, which adds complexity to our deployment pipeline.
Lalam: I’m confident this paper is solid; they clearly showed that standard techniques fall short for gaze estimation regression problems and pointed toward a specific path forward.
Tom: Exactly. They laid out a clear roadmap for tackling these fairness challenges head-on in gaze estimation research, showing that we need to look beyond just the prediction accuracy score.
Jane: It really boils down to this: as we move appearance-based gaze estimation into real applications, the focus has shifted from just how accurate the prediction is to how fair that prediction is for everyone.
Tom: Yeah, they clearly showed that standard techniques fall short, so we have a clear direction on what needs more research in geometric or domain-specific solutions instead of just general fixes.
Lu: That opens up a lot of avenues for representation learning research focused specifically on how those sensitive attributes interact with the model's internal representations.
Meng: If we can find an architecture that naturally handles this better, it could save us a lot of work down the line when we have to manage those labels at scale.
Lalam: I’m sure the next steps will involve applying these mitigation strategies to see if they actually translate into real-world fairness gains on live traffic data.
Dept. of Computer Eng. and ROMER Robotics Center, Middle East Technical University
cs.CV
Submitted: 2026-04-12
Updated: 2026-04-12
Journal ref: 2026 IEEE 20th International Conference on Automatic Face and Gesture Recognition (FG), 2026
DOI: 10.1109/FG67764.2026.11557007
Code: https://github.com/akgulburak/gaze-estimation-fairness
License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/
Importance score: 81/100
The gist: Appearance-based gaze estimation methods suffer from significant fairness issues across different demographic groups and this paper presents the first extensive evaluation of fairness in…
Key concepts
- Appearance-based Gaze Estimation
- This is the process where an AI system tries to guess where a person is looking in an image by analyzing their physical appearance, such as facial features. The system learns patterns from images to predict gaze direction, but these predictions can be unfair if the training data lacks diversity.
- Fairness Metrics
- These are mathematical measures used to quantify how fair or unfair a model's predictions are across different demographic groups like gender and ethnicity. The paper uses these metrics to systematically identify where the gaze estimation models perform poorly for certain groups.
- Bias Mitigation Strategies
- These are techniques applied during the training of an AI model to reduce unfairness. The study tested oversampling, which balances data samples by increasing representation of minority groups, and loss reweighting, which penalizes errors on underrepresented groups more heavily.
Terminology
Summary
Appearance-based gaze estimation methods suffer from significant fairness issues across different demographic groups and this paper presents the first extensive evaluation of fairness in appearance-based gaze estimation focusing on ethnicity and gender attributes. p. 0
Research Gaps Addressed
The paper addresses several research gaps concerning bias and fairness in this domain. The identified gaps include:
-
Many high-stake applications, e. p. 0g., driver monitoring systems, can lead to life threatening outcomes owing to biases across demographic groups. p. 0
-
There is no study focusing on mitigating the bias issues in gaze estimation. p. 0
-
One fundamental reason for these two gaps is the lack of demographic labels for the currently published gaze estimation datasets where only the gaze label is available. p. 0
Data Annotation and Bias Analysis
The authors establish a fairness baseline by analyzing state-of-the-art models using standard fairness metrics after annotating existing datasets with sensitive attributes.
Green — the paper says it, and the page is printed. Red — no passage carries it, so treat that sentence as unverified. The same check runs on your own text: check a summary against its source, $5.
Improvements for AI systems
-
Data Annotation for Fairness: Annotate existing gaze estimation datasets, namely Gaze360 and GazeCapture, with gender and ethnicity sensitive attributes using a deep learning model like Fairface to generate initial labels before human annotators validate them.
-
Bias Analysis across Models: Systematically analyze bias in gaze predictions of state-of-the-art models (CrossGaze, MCGaze, L2CS-Net, PureGaze, GazeTR) on both Gaze360 and GazeCapture datasets to highlight severe fairness issues across ethnicity and gender.
-
Baseline Fairness Benchmarking: Establish a baseline benchmark for fairness algorithms by evaluating several preprocessing and in-processing bias mitigation algorithms against the identified biases in gaze estimation regression problems.
-
Targeted Bias Mitigation Strategy: Apply oversampling (to make minority group sample probabilities equal to the majority group) and loss reweighting (where loss value is increased for a minority group with respect to the inverse of the sample size) during training to reduce bias, noting that
resampling and loss reweighting methods provide the most number of improvements across models.
-
Model Selection for Fairness: Prioritize models like PureGaze when using resampling and loss reweighting on Gaze360 data, as they
consistently benefit from resampling across ethnicity as well as gender labels on Gaze360
and show improvement in fairness metrics.
Sources
- Normalise for Fairness: A Simple Normalisation Technique for Fairness in Regression Machine Learning Problems
- Faces of Fairness: Examining Bias in Facial Expression Recognition Datasets and Models
- Demographic Parity: Mitigating Biases in Real-World Data
- Model-agnostic bias mitigation methods with regressor distribution control for Wasserstein-based fairness metrics
- mixup: Beyond Empirical Risk Minimization
Related papers
- TEA: Text Encoder Alignment for Robust Concept Erasure in Text-to-Image Models
- Do Egocentric Video-Language Models Capture Both Hand- and Object-Centric Cues?
- Aura: Consistent Multi-Subject Video Generation via VLM-Grounded Semantic Alignment
- PVRF: All-in-one Adverse Weather Removal via Prior-modulated and Velocity-constrained Rectified Flow
- DLWM: Diverse Latent World Models for Efficient Multimodal Reasoning
- Out of Sight, Still in Mind: Token Compression for Omni-LLMs