Eye-gaze estimation in PyTorch

Docker

Pull the image from Docker Hub. It contains all the required packages.

docker pull kroniidvul/pytorch_mpiigaze:latest

Run the container interactively.

docker run -it --rm kroniidvul/pytorch_mpiigaze /bin/bash

Download the dataset and preprocess it

$ wget http://datasets.d2.mpi-inf.mpg.de/MPIIGaze/MPIIGaze.tar.gz
$ tar xzvf MPIIGaze.tar.gz
$ python preprocess_data.py --dataset MPIIGaze --outdir data

Usage

$ python -u main.py --arch lenet --dataset data --test_id 0 --outdir results/00
$ python -u main.py --arch lenet --dataset data --test_id 0 --outdir results/lenet/00 --batch_size 32 --base_lr 0.01 --momentum 0.9 --nesterov True --weight_decay 1e-4 --epochs 40 --milestones '[30, 35]' --lr_decay 0.1

Project

This work explores various parameters, lr schedulers, deep neural architectures, ensembling, and a mask-based approach of using upsampled gaze vectors for appearance based gaze estimation on the MPIIGaze dataset.

Results

References

• https://github.com/hysts/pytorch_mpiigaze Original Git repo
• Xucong Zhang and Yusuke Sugano and Mario Fritz and Bulling, Andreas, “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
• Xucong Zhang and Yusuke Sugano and Mario Fritz and Bulling, Andreas, “MPIIGaze: Real-World Dataset and Deep Appearance-Based Gaze Estimation,” arXiv:1711.09017