Appearance-based Gaze Estimation using Attention and Difference Mechanism

Abstract

Appearance-based gaze estimation problem received wide attention over the past few years. Even though model-based approaches existed earlier, availability of large datasets and novel deep learning techniques made appearance-based methods achieve superior accuracy than model-based approaches. In this paper, we proposed two novel techniques to improve gaze estimation accuracy. Our first approach, I2D-Net uses a difference layer to eliminate any common features from left and right eyes of a participant that are not pertinent to gaze estimation task. Our second approach, AGE-Net adapted the idea of attention-mechanism and assigns weights to the features extracted from eye images. I2D-Net performed on par with the existing state-of-the-art approaches while AGE-Net reported state-of-the-art accuracy of 4.09 <sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">◦</sup> and 7.44 <sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">◦</sup> error on MPI-IGaze and RT-Gene datasets respectively. We performed ablation studies to understand the effectiveness of the proposed approaches followed by analysis of gaze error distribution with respect to various factors of MPIIGaze dataset.