ASGaze

This is the repo for SenSys 2022 paper: "Gaze Tracking on Any Surface with Your Phone".

Authors: Jiani Cao, Chengdong Lin, Yang Liu, Zhenjiang Li

Project website: ASGaze

Demo video:

Requirements

The program has been tested in the following environment:

• Python 3.7
• Numpy 1.21.4
• Pytorch 1.10.3
• torchvision 0.11.1
• opencv-python 3.4.2.17
• dlib 19.17
  

Project Structure

|-- iris_boundary_detector                   
    |-- data_sources                      
        |-- detection.ipynb               // detect and save eye regions using dlib library 
        |-- cvdata                        // facial landmarks used in "detection.ipynb"
        |-- ASGaze_data.ipynb             // Pytorch Dataset prepared for train and inference
        |-- transform.ipynb               // data augmentation
    |-- graph       
        |-- vgg_unet.ipynb                // backbone of segmentation network
        |-- losses.ipynb                  // loss functions used to train segmentation network
    |-- utils
        |-- load_model.ipynb              // helper functions used to load and save model
        |-- metrics.ipynb                 // metrics used to evaluate segmentation network
        |-- refinement.ipynb              // leverage temporal relationship to refine iris boundary
    |-- configs
        |-- segmentation_train.json       // config parameters for training network
        |-- gaze_inference.json           // config parameters for output inferenced iris boundary
    |-- train.ipynb                       // main workflow of train
    |-- inference.ipynb                   // main workflow of inference

|-- gaze_ray_estimator    
    |-- cone_model.ipynb                  // cone model used to establish the relationship between 3D circle and 2D ellipse
    |-- estimator.ipynb                   // main workflow of gaze ray estimator

|-- mapping    
    |-- mapping_principle.ipynb           // mapping principle
    |-- shape_constrained.ipynb           // proposed mapping method

|-- Database                              // conclude data sample, pretrained model and camera matrix

|-- setup.ipynb                           // remove ambiguity and calculate offsets (one-time effort)

|-- main.ipynb                            // main workflow of ASGaze

Quick Start

• Download and unzip the Database folders. Detailed descriptions are in Database.md.

• Change the "dir", "runs_dir" of data and pretrained model in gaze_inference.json to the path on your machine.

• Run the main.ipynb script and you can visualize the tracking process, just like the demo video.

  ---

Citation

If you find our work useful in your research, please consider citing:

@inproceedings{cao2022gaze,
title={Gaze Tracking on Any Surface with Your Phone},
author={Cao, Jiani and Lin, Chengdong and Liu, Yang and Li, Zhenjiang},
booktitle={Proceedings of the 20th ACM Conference on Embedded Networked Sensor Systems},
pages={320--333},
year={2022}
}