We propose to use unlabelled eye image data for domain adaptation of an iris segmentation network. Adaptation allows the model to be less reliant on its initial generality. This is beneficial due to the large variance exhibited by eye image data which makes training of robust models difficult. The method uses a label prior in conjunction with network predictions to produce pseudo-labels. These are used in place of ground-truth data to adapt a base model. A fully connected neural network performs the pixel-wise iris segmentation. The base model is trained on synthetic data and adapted to several existing datasets with real-world eye images. The adapted models improve the average pupil centre detection rates by 24% at a distance of 25 pixels. We argue that the proposed method, and domain adaptation in general, is an interesting direction for increasing robustness of eye feature detectors.
|Titel||ETRA '20 Full Papers: ACM Symposium on Eye Tracking Research and Applications|
|Forlag||Association for Computing Machinery|
|Status||Udgivet - 2020|