Synthetic-gaze-data-augmentation-for-improved-user-calibration

Model weights and network implementation for reproducibility of the paper "Synthetic gaze data augmentation for improved user calibration"

Modules

We used Tensorflow as framework to build the architecture. Specifically, we used tensorflow.keras. An additional module necessary to build the model is the classification_models module, from which we used the pretrained Resnet-18 over imagenet. To install this module, please refer to: https://github.com/qubvel/classification_models

Pretrained model over U2Eyes

The pretrained model needs a custom object in order to be loaded by keras.

model = keras.models.load_model('path_to_file_h5', custom_objects={'regression_loss':regression_loss}

The regression_loss function is defined as:

import tensorflow as tf
@tf.function
def regression_loss(y_true, y_pred):
    dif_x = tf.math.square(y_true[:,0] - y_pred[:,0])
    dif_y = tf.math.square(y_true[:,1] - y_pred[:,1])
    dist_vector = tf.math.sqrt(dif_x + dif_y)
    loss = tf.math.reduce_mean(dist_vector)
    return loss

Because of space limitation for github (100MB), the weights are under releases section (check tags)