Accurate CNN-based pupil segmentation with an ellipse fit error regularization term

Abstract

Semantic segmentation of images by Fully Convolutional Neural Networks (FCN) has gained increased attention in recent years as FCNs greatly outperform traditional segmentation algorithms. In this paper we propose using Ellipse Fit Error as a shape prior regularization term that can be added to a pixel-wise loss function, e.g., binary cross entropy, to train a CNN for pupil segmentation. We evaluate the performance of the proposed method by training a lightweight UNet architecture, and use three widely used real-world datasets for pupil center estimation, i.e., ExCuSe, ElSe, and Labeled Pupils in the Wild (LPW), containing a total of ∼230.000 images for performance evaluation. Experimental results show that the proposed method gives the best-known pupil detection rates for all datasets.