Skip to content

Code, logs, and final models for ACM ETRA '23: One Eye is All You Need: Lightweight CNN Encoders for Accurate Zero-Calibration Gaze Estimation.

License

Notifications You must be signed in to change notification settings

vytal-ai/One-Eye-is-All-You-Need

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

4 Commits
 
 
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

One Eye is All You Need: Accurate Gaze Estimation

This repository contains the code to reproduce the results of our paper One Eye is All You Need: Lightweight CNN Encoders for Accurate Zero-Calibration Gaze Estimation, submitted to the 14th ACM Symposium on Eye Tracking Research & Applications in Tübingen, Germany from May 30-June 2, 2023.

Table of Contents

About The Project

Many current gaze estimation models not only fail to utilize robust computer vision (CV) algorithms but also require the use of either both eyes or the entire face, where high-resolution real-world data may not be available. Thus, we propose a gaze estimation model that implements lightweight ResNet, Inception, and SqueezeNet models as encoders for eye-image data and makes predictions using only one eye, which prioritizes both accuracy and speed. We attain high performance on the GazeCapture dataset with these models without calibration; with two eyes, we achieve a base prediction error of 1.471 cm on the test set, and with just a single eye, we achieve a base prediction error of 2.312 cm on the test set. These results surpass those of other uncalibrated gaze tracking models and demonstrate that gaze predictors can still achieve exceptional results when only considering a single eye.

Getting Started

First, create a Conda environment, and then use the command conda install --file requirements.txt --channel conda-forge --channel comet_ml in the terminal to install the necessary packages and dependencies for this project. Note that this will only install the necessary libraries and dependencies required for this project. Due to the large size of the dataset, you will be unable to run this code on your machine.

Repository Structure

Folder Description
full Models and training files where both eyes are used in model training.
one_eye Models and training files where only one eye is used in model training.
utils Data loading files from the GazeCapture dataset. There are two files, based on what type of data the model requests.

Issues and Contact

For any issues pertaining to the code, please use the Issues tab. For questions regarding the research paper, please email any of the first authors of this paper (Rishi Athavale, Lakshmi Sritan Motati, Anish Susarla, and Rohan Kalahasty; email addresses in the paper), and we will get back to you to address your questions.

About

Code, logs, and final models for ACM ETRA '23: One Eye is All You Need: Lightweight CNN Encoders for Accurate Zero-Calibration Gaze Estimation.

Resources

License

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published

Languages