Uncalibrated Eye Gaze Estimation using SE-ResNext with Unconstrained Head Movement and Ambient Light Change

Abstract

Abstract Technological advances in smartphones, tablets, computer games, virtual reality, metaverse, and other fields have made gaze estimation (GE) using standard hardware more necessary than ever before. It can also be used in other areas such as psychology, increased driving safety, and advertisement. This paper proposes a structure based on convolutional neural networks (CNNs). In this structure, several well-known CNNs are implemented and trained with a section of the GazeCapture dataset for acceleration. The SE-ResNext network, which has the best results in initial training, is selected in the end. The test error for the designated structure is 1.32 cm in training with the entire dataset. The ambient light is an effective factor in GE accuracy. It clearly affects different GE methods. The dataset is divided into low-light and bright-light environment sets to find a solution. The bright-light environment samples are much more abundant than the low-light ones, something which causes a bias in gaze estimator training. Therefore, standard data augmentation methods are employed to increase the number of low-light samples and retrain the gaze estimator. As a result, the GE error is reduced from 1.20 to 1.06 cm for bright-light environments and from 3.39 to 1.87 cm for low-light environments. To examine resistance of the gaze estimator to head movement, the test dataset is manually and intuitively classified into five subsets based on head positions. In this classification, test errors of 1.27, 1.427, 1.496, 1.952, and 2.466 cm are respectively obtained for the frontal, roll to right, roll to left, yaw to right, and yaw to left head positions.