PureGaze: Purifying Gaze Feature for Generalizable Gaze Estimation

Description

Our work is accpeted by AAAI 2022.

Picture: We propose a domain-generalization framework for gaze estimation. Our method is only trained in the source domain and brings improvement in all unknown target domains. The key idea of our method is to purify the gaze feature with a self-adversarial framework.

Picture: Overview of the gaze feature purification. Our goal is to preserve the gaze-relevant feature and eliminate gaze-irrelevant features. We define two tasks, which are to preserve gaze information and to remove general facial image information. The two tasks are not cooperative but adversarial to purify feature. Simultaneously optimizing the two tasks, we implicitly purify the gaze feature without defining gaze-irrelevant feature.

Performance: PureGaze shows best performance among typical gaze estimation methods (w/o adaption), and has competitive result among domain adaption methods. Note that, PureGaze learns one optimal model for four tasks, while domain adaption methods need to learn a total of four models. This is an advantage of PureGaze.

Feature visualization: The result clearly explains the purification. Our purified feature contains less gaze-irrelevant feature and naturally improves the cross-domain performance.

Usage

This is a re-implemented version by Pytorch1.7.1 (origin is Pytorch1.0.1).

We provides an Res50-Version PureGaze. If you want to change the backbone to Res18, you could use the file in Model/Res18.

Resourse

Model/: Implemented code.
Masker/: The masker used for training.

Get Started

1. You could find data processing code from this link.

2. modifing files in config/ folder, and run commands like:

   Training:python trainer/total.py -c config/train/config-eth.yaml

   Test:python tester/total.py -s config/train/config-eth.yaml -t config/test/config-mpii.yaml

   Visual:python tester/visual.py -s config/train/config-eth.yaml -t config/test/config-mpii.yaml

Pre-trained model.

We provide a pre-trained model of Res50-version PureGaze. You can find it from this link.

Citation.

@article{cheng2022puregaze,
  title={PureGaze: Purifying Gaze Feature for Generalizable Gaze Estimation},
  author={Yihua Cheng and Yiwei Bao and Feng Lu},
  journal={Proceedings of the AAAI Conference on Artificial Intelligence},
  year={2022}
}

Contact

Please email [email protected].