Gaze Duration Error

Gazerr is an application for estimating the expected error in a gaze duration measurement derived from repeated application of a point of gaze model. It is particularly applicable to machine learning models that work with device cameras to predict a stream of gaze fixation points from facial images.

The method requires an input dataset of validation points from the point of gaze predictive model. This data is used to generate the probability distribution of true gaze durations given a measured gaze duration.

Installation

Install from source code or from PyPi

Usage

The application can be used from the command line by passing in a path to the calibration file and the parameters for the duration measurement that will be bounded. Note: that the penultimate two parameters should be comma separated sets of integers that depict x,y coordiantes in pixels. The measurement length and session lenth should be expressed in milliseconds.

The final parameter is a path to a directory in which to store the raw results

gazerr <CALIBRATION> <MEASUREMENT> <SESSION> <TARGET TOP LEFT> <TARGET BOTTOM RIGHT> <RESULTS>

To use the application without installing it you can employ the runner script. Example below, using the supplied calibration data:

mkdir results/MREC_MAE_50
python ./gazerr-runner.py data/validation_50_MAE.csv 400 1000 40,40 340,290 results/MREC_MAE_50

Alternatively, you may inspect the code and use the library functions directly inside your own application.

Experiments

All experiments for the research paper can be executed via a series of scripts.

Create synthetic calibration data by running

python scripts/generate_datasets.py

Then execute the gazerr exeriments with the following two commands:

scripts/RUN_EXPERIMENTS.sh
scripts/RUN_BIAS_EXPERIMENTS.sh

Finally, analyse the results and generate the plots with

scripts/ANALYSE_RESULTS.sh

Documentation

Additional documentation to be made available on Read the Docs

If you use gazerr in your research please cite the following article

@article{hawkins2022,
   author = {John Hawkins},
   year = {2022},
   title = {Estimating Gaze Duration Error from Eye Tracking Data},
   journal = {TBC}
}