This repository contains the implementation to develop a gaze estimator for 128x128 full face images. It also contains the error analysis performed on the predictions of the same.
This code was developed and tested on Ubuntu 18.04.5 LTS, it requires GPUs and CUDA support. The python version used was 3.8.0 . To run the code you would need Jupyter Software. You can also configure custom kernels for it in VS code.
All other requirements can be installed with:
pip install -r requirements.txt
GazeCapture.h5, MPIIGaze.h5 are the training and testing sets used respectively. To obtain these please follow steps mentioned here to pre process the original datasets.
Redirected_samples1.h5, is the augmented dataset please use this to generate the augmented dataset.
For doing the error analysis, the predictions are loaded from loss_2303196.pkl pickle file. Please use the code in train_gaze_estimator.ipynb to generate predictions on the augmented dataset.
The notebook train_gaze_estimation.ipynb contains the code to train/ test the model, the cells are organized so that one may perform both separately.
The resultant plots are saved in plots folder.
The notebook error_analysis.ipynb contains the code to perform error analysis on the predictions.