S3PL : Self-Supervised Segmentation for Robust Pupil Localization

The task of pupil localization is an imperative step for any gaze estimation and eye tracking systems. Due to the challenges induced by diverse factors like illumination, reflections and noise, researchers have proposed several classical and learning based approaches addressing this problem. Even though learning based methods have shown superior accuracy compared to classical methods, they require extensive annotations for training and are prone to poor cross-dataset performance on new unseen scenarios. In this work, we formulated pupil center estimation as an unpaired image-to-image translation task and introduced a novel loss function to the existing CycleGAN framework to achieve robust and accurate pupil localization using self-supervised learning paradigm. Further, we evaluated our method on four publicly available datasets, Labelled pupils in the wild (LPW), OpenEDS, Świrski, and NVGaze, and reported the state-of-the-art pupil localization performance. Świrski, and NVGaze are used for cross-data evaluation to understand the generalizability of the proposed method. The proposed method achieved an improvement of around 16% on LPW, 21% on OpenEDS, 5% on Świrski, and 14% on NVGaze datasets in terms of pupil detection rate over the conventional methods.