Gaze Estimation Diversity — Model Weights
Checkpoints for the models used in "Investigating Bias and Fairness in Appearance-based Gaze Estimation" (Akgül, Şahin & Kalkan, 2026).
Full code, training scripts, annotations, and documentation are on GitHub: akgulburak/gaze-estimation-fairness
What's in this repo
Checkpoints for 5 architectures (CrossGaze, GazeTR, L2CS-Net, MCGaze, PureGaze), each trained under 4 conditions:
- Baseline (no bias mitigation)
- Oversampling
- Resampling
- Loss reweighting
Each mitigation condition was trained separately targeting the ethnicity attribute and the gender attribute, all on the Gaze360 dataset.
File Naming
We repeated each run 5 times to account for training variance. The _1-_5 suffix in each filename indicates the repetition index - these are independent runs of the same configuration, not different models.
Usage
See the GitHub repo for training/inference code.
License
This project is licensed under CC BY-NC-SA 4.0 - see LICENSE for details.
Citation
@inproceedings{akgul2026biasingaze,
title = {Investigating Bias and Fairness in Appearance-based Gaze Estimation},
author = {Akg{\"u}l, Burak and {\c{S}}ahin, Erol and Kalkan, Sinan},
booktitle = {2026 IEEE 20th International Conference on Automatic Face and Gesture Recognition (FG)},
year = {2026},
address = {Kyoto, Japan},
publisher = {IEEE},
doi = {10.1109/FG67764.2026.11557007}
}