Gaze360

The Pytorch Implementation of "Gaze360: Physically Unconstrained Gaze Estimation in the Wild". (updated in 2021/04/28)

We build benchmarks for gaze estimation in our survey "Appearance-based Gaze Estimation With Deep Learning: A Review and Benchmark". This is the implemented code of the "Gaze360" method in our benchmark. Please refer our survey for more details.

We recommend you to use data processing codes provided in GazeHub. You can direct run the method' code using the processed dataset.

Links to gaze estimation codes.

A Coarse-to-fine Adaptive Network for Appearance-based Gaze Estimation, AAAI 2020 (Coming soon)
Gaze360: Physically Unconstrained Gaze Estimation in the Wild, ICCV 2019
Appearance-Based Gaze Estimation Using Dilated-Convolutions, ACCV 2019
Appearance-Based Gaze Estimation via Evaluation-Guided Asymmetric Regression, ECCV 2018
RT-GENE: Real-Time Eye Gaze Estimation in Natural Environments, ECCV 2018
MPIIGaze: Real-World Dataset and Deep Appearance-Based Gaze Estimation, TPAMI 2017
It’s written all over your face: Full-face appearance-based gaze estimation, CVPRW 2017
Eye Tracking for Everyone, CVPR 2016
Appearance-Based Gaze Estimation in the Wild, CVPR 2015

Performance

The method is evaluated in three task. Please refer our survey for more details. Note that, we use the static model for fair comparsion in this benchmark.

License

The code is under the license of CC BY-NC-SA 4.0 license.

Introduction

We provide two projects for leave-one-person-out evaluation and the evaluation of common training-test split. They have the same architecture but different train.py and test.py.

Each project contains following files/folders.

model.py, the model code.
train.py, the entry for training.
test.py, the entry for testing.
config/, this folder contains the config of experiments for each dataset. To run our code, you should write your own config.yaml.
reader/, the data loader code. You can use the provided reader or write your own reader.

Getting Started

Writing your own config.yaml

Normally, for training, you should change

train.save.save_path, The model is saved in the $save_path$/checkpoint/.
train.data.image, This is the path of image, please use the provided data processing code in GazeHub.
train.data.label, This is the path of label.
reader, This indicates the used reader. It is the filename in reader folder, e.g., reader/reader_mpii.py ==> reader: reader_mpii.

For test, you should change

test.load.load_path, it is usually the same as train.save.save_path. The test result is saved in $load_path$/evaluation/.
test.data.image, it is usually the same as train.data.image.
test.data.label, it is usually the same as train.data.label.

Training

In the leaveout folder, you can run

python train.py config/config_mpii.yaml 0

This means the code will run with config_mpii.yaml and use the 0th person as the test set.

You also can run

bash run.sh train.py config/config_mpii.yaml

This means the code will perform leave-one-person-out training automatically.
run.sh performs iteration, you can change the iteration times in run.sh for different datasets, e.g., set the iteration times as 4 for four-fold validation.

In the traintest folder, you can run

python train.py config/config_mpii.yaml

Test

In the leaveout folder, you can run

python test.py config/config_mpii.yaml 0

bash run.sh test.py config/config_mpii.yaml

In the traintest folder, you can run

python test.py config/config_mpii.yaml

Result

After training or test, you can find the result from the save_path in config_mpii.yaml.

Citation

If you use our code, please cite:

@InProceedings{Kellnhofer_2019_ICCV,
	author = {Kellnhofer, Petr and Recasens, Adria and Stent, Simon and Matusik, Wojciech and Torralba, Antonio},
	title = {Gaze360: Physically Unconstrained Gaze Estimation in the Wild},
	booktitle = {The IEEE International Conference on Computer Vision (ICCV)},
	month = {October},
	year = {2019}
}

@article{Cheng2021Survey,
        title={Appearance-based Gaze Estimation With Deep Learning: A Review and Benchmark},
        author={Yihua Cheng and Haofei Wang and Yiwei Bao and Feng Lu},
        journal={arXiv preprint arXiv:2104.12668},
        year={2021}
}

Contact

Please email any questions or comments to yihua_c@buaa.edu.cn.

Reference

MPIIGaze: Real-World Dataset and Deep Appearance-Based Gaze Estimation
EYEDIAP Database: Data Description and Gaze Tracking Evaluation Benchmarks
Learning-by-Synthesis for Appearance-based 3D Gaze Estimation
Gaze360: Physically Unconstrained Gaze Estimation in the Wild
ETH-XGaze: A Large Scale Dataset for Gaze Estimation under Extreme Head Pose and Gaze Variation
Appearance-Based Gaze Estimation in the Wild
Appearance-Based Gaze Estimation Using Dilated-Convolutions
RT-GENE: Real-Time Eye Gaze Estimation in Natural Environments
It’s written all over your face: Full-face appearance-based gaze estimation
A Coarse-to-fine Adaptive Network for Appearance-based Gaze Estimation
Eye Tracking for Everyone
Adaptive Feature Fusion Network for Gaze Tracking in Mobile Tablets
On-Device Few-Shot Personalization for Real-Time Gaze Estimation
A Generalized and Robust Method Towards Practical Gaze Estimation on Smart Phone

yihuacheng / Gaze360