Awesome

pl_gaze_estimation

Training code for gaze estimation models using MPIIGaze, MPIIFaceGaze, and ETH-XGaze.

https://user-images.githubusercontent.com/25161192/143683988-06ab6495-ae97-4201-b3f1-39dacf4bc69c.mp4

A demo is available in this repo.

Installation

• Linux (Tested on Ubuntu only)
• Python >= 3.9

pip install -r requirements.txt

For docker environment, see here.

Usage

The basic usage is as follows:

python train.py --configs /path/to/your/config.yaml

Multiple config files can be specified to set new variables or overwrite previously defined values. You can also overwrite previously defined values by listing the key/value pairs after the --options flags.

usage: train.py [-h] --configs [CONFIGS ...] [--options ...]

optional arguments:
  -h, --help            show this help message and exit
  --configs [CONFIGS ...]
                        Paths to config files.
  --options ...         Variables to overwrite. (optional)

MPIIGaze

test_id=0
python train.py --configs configs/examples/mpiigaze_lenet.yaml \
                --options EXPERIMENT.TEST_ID ${test_id} \
                          EXPERIMENT.OUTPUT_DIR exp0000/$(printf %02d ${test_id})

You need to download and preprocess the dataset before training by running the following commands:

bash scripts/download_mpiigaze_dataset.sh
python scripts/preprocess_mpiigaze.py --dataset datasets/MPIIGaze -o datasets/

(Additional packages are required to preprocess the dataset. See the code.)

MPIIFaceGaze

test_id=0
python train.py --configs configs/examples/mpiifacegaze.yaml \
                --options SCHEDULER.WARMUP.EPOCHS 3 \
                          EXPERIMENT.TEST_ID ${test_id} \
                          EXPERIMENT.OUTPUT_DIR exp0000/$(printf %02d ${test_id})

You need to download and preprocess the dataset before training by running the following commands:

bash scripts/download_mpiifacegaze_dataset.sh
python scripts/preprocess_mpiifacegaze.py --dataset datasets/MPIIFaceGaze_normalized -o datasets/

(Additional packages are required to preprocess the dataset. See the code.)

ETH-XGaze

python train.py \
    --config configs/examples/eth_xgaze.yaml \
    --options \
        VAL.VAL_INDICES "[1, 23, 24, 35, 38, 46, 58, 63, 70, 78]" \
        SCHEDULER.EPOCHS 15 \
        SCHEDULER.MULTISTEP.MILESTONES "[10, 13, 14]" \
        DATASET.TRANSFORM.TRAIN.HORIZONTAL_FLIP true \
        EXPERIMENT.OUTPUT_DIR exp0000

Docker Environment

docker-compose 1.29.2 is required.

Build

docker-compose build train

Train

docker-compose run --rm -u $(id -u):$(id -g) -v /path/to/datasets:/datasets train python train.py --configs /path/to/your/config.yaml

Results on ETH-XGaze dataset

All the results in the table below are of a single run. In these experiments, the data of subjects with id 1, 23, 24, 35, 38, 46, 58, 63, 70, and 78 were used as the validation data. The models were trained for 15 epochs with a multi-step learning rate schedule. The learning rate was multiplied by 0.1 at epochs 10, 13, and 14.

Related repos

• https://github.com/hysts/pytorch_mpiigaze
• https://github.com/hysts/pytorch_mpiigaze_demo

References

• Cai, Xin, Boyu Chen, Jiabei Zeng, Jiajun Zhang, Yunjia Sun, Xiao Wang, Zhilong Ji, Xiao Liu, Xilin Chen, and Shiguang Shan. "Gaze Estimation with an Ensemble of Four Architectures." arXiv preprint arXiv:2107.01980 (2021). arXiv:2107.01980, GitHub
• Zhang, Xucong, Seonwook Park, Thabo Beeler, Derek Bradley, Siyu Tang, and Otmar Hilliges. "ETH-XGaze: A Large Scale Dataset for Gaze Estimation under Extreme Head Pose and Gaze Variation." In European Conference on Computer Vision (ECCV), 2020. arXiv:2007.15837, Project Page, GitHub, Leaderboard
• Zhang, Xucong, Yusuke Sugano, Mario Fritz, and Andreas Bulling. "Appearance-based Gaze Estimation in the Wild." Proc. of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2015. arXiv:1504.02863, Project Page
• Zhang, Xucong, Yusuke Sugano, Mario Fritz, and Andreas Bulling. "It's Written All Over Your Face: Full-Face Appearance-Based Gaze Estimation." Proc. of the IEEE Conference on Computer Vision and Pattern Recognition Workshops(CVPRW), 2017. arXiv:1611.08860, Project Page
• Zhang, Xucong, Yusuke Sugano, Mario Fritz, and Andreas Bulling. "MPIIGaze: Real-World Dataset and Deep Appearance-Based Gaze Estimation." IEEE transactions on pattern analysis and machine intelligence 41 (2017). arXiv:1711.09017
• Zhang, Xucong, Yusuke Sugano, and Andreas Bulling. "Evaluation of Appearance-Based Methods and Implications for Gaze-Based Applications." Proc. ACM SIGCHI Conference on Human Factors in Computing Systems (CHI), 2019. arXiv, code