LWSeg: Lightweight Eye Segmentation Model for Gaze Estimation

Gaze estimation is increasingly becoming an integral part of head-mounted devices (HMD). Eye segmentation with ellipse fitting under near-infrared illumination is one of the current methods of eye tracking techniques for augmented reality (AR) HMD. However, most of the existing methods work well under laboratory environment without considering the effects of environmental changes, such as wearing posture and lighting. Meanwhile, The task of accurate iris/pupil segmentation remains challenging. To solve these problems, we first constructed a dataset with multiple wearing postures for different subjects, containing binocular images, segmentation truth labels, and gaze data, to support the subsequent segmentation and gaze estimation tasks. In addition, high-performance gaze estimation usually relies on accurate and fast eye segmentation. To improve the computational speed of the segmentation model, we proposed a lightweight eye segmentation model, which incorporated structures such as bottleneck layer and parallel dilated convolution on the basis of U-Net to achieve fast and accurate eye segmentation. Experiments show that the algorithm exhibits good segmentation performance on both self-constructed and public datasets, with 6% reduction in computational complexity, 43.7% reduction in the number of parameters, and 24% reduction in the inference speed of a single image on the CPU, compared to state-of-the-art (SOTA) segmentation networks.