Low-cost gaze detection with real-time ocular movements using coordinate-convolutional neural networks

Abstract

Detection of ocular-movements unfolds various possibilities in computer vision but requires large datasets, expensive hardware and computational power. Prior research substantiates the belief that Convolutional Neural Network provides the highest recognition rate compared to traditional techniques, but they begin to overfit after achieving a certain accuracy due to the coordinate-transform-problem. Different image conditions like variation-in-viewpoint or illumination can be pragmatic for image processing and require on-device calibration. This paper proposes a framework that works with low-computational-complexity in varied environmental conditions to provide efficient gaze estimations that points out screen coordinates in real-time. We use a depth-wise convolution, an expansion and a projection layer along-with coordinate-channels to improve classification. The model is experimented against different environmental conditions, multiple subjects, image augmentation and different data sizes in real-time to estimate the coordinate classes using eye-movements on a standard web camera, yielding better accuracy and preventing overfitting of model with fewer hardware requirements.