Single-Image Pupil Localization via Implicit 3D Eye Reconstruction

We propose a pupil localization algorithm that directly estimates the pupil region in a single image based on geometric 3D priors and implicit 3D eye reconstruction using selfsupervised learning. First, we develop a 3D eye reconstruction network that implicitly constructs a biologically inspired eye model from a single image by estimating geometric eye priors. Then, we project the reconstructed 3D eye model back onto the original image plane by developing an inverse ray tracing technique to localize the 2 D pupil region. Since this projection is non-differentiable and 3D annotations are unavailable, we develop a self-supervised learning strategy that generates 3D pseudo-annotations from 2 D pupil ground truths to train the reconstruction network. Experimental results demonstrate that the proposed algorithm achieves better performance than state-of-the-art algorithms in both quantitative and qualitative evaluations.