A YOLO Detector Providing Fast and Accurate Pupil Center Estimation using Regions Surrounding a Pupil

Eye-tracking technology has many useful applications, including Virtual Reality (VR) devices, Augmented Reality (AR) devices, and assistive technology. The main objective of eye-tracking technology is to detect eye position and track eye movements. It is possible to determine the eye position when the pupil center is detected. In this paper, a deep learning-based approach to the detection of pupil centers from webcam images is presented. As opposed to all previous approaches to object detection based on training the detector with objects to be detected, our object detector was trained with both the region surrounding a pupil and the region between an eye and the region surrounding a pupil. The latter set of regions has been found to increase the overall detection accuracy. A novel post-processing algorithm is also presented to estimate the pupil center from all the detected regions. To achieve real-time performance, we have adopted the tiny architecture of YOLOv3, which has 23 layers and can be executed without requiring a GPU accelerator. To train the detectors, different variations of regions covering a pupil and an eye were used, as well as expanding regions surrounding a pupil and an eye. The PUPPIE dataset was used as the primary input for training the detector. The setting with the best detection accuracy was applied to all publicly available datasets: I2Head, MPIIGaze, and U2Eyes. In terms of accuracy, the results indicate that pupil center estimation is comparable to the state-of-the-art approach. It achieves pupil center estimation errors below the size of a constricted pupil in more than 98.24% of images. Furthermore, the detection time is 2.8 times faster than the state-of-the-art approach.