Siamese networks are able to give a lot of information on the embedding space by comparing the features of the anchor with an image from the same class and a negative class.
The paper introduced UniNet, an architecture that uses MaskNet to identify the regions of the image that do not belong to the iris, and FeatNet, which is a Fully Convolution Network for discriminative Iris feature extraction.
The main contribution of the paper is the Extended Triplet Loss. There are two modifications to the Triplet Loss, which is used for training Siamese networks. If the Iris network is rotated, the normalized segmented image is bit-shifted. So, we need to compare the features when there is minimum bit-shift between the two images. The loss considers the features when the L2 loss between them is minimum. This ensures that the images are rotated by approximately the same amount. The other contribution is to consider only the necessary features. The modified loss masks the output features such that only those belonging to regions of the iris are considered in the loss calculation. This ensures that only valid features are taken for updating the weights!
The MaskNet is a pre-trained model that had been trained on 500 images annotated by the authors. The Featnet is trained using the extended triplet loss but the weights of MaskNet are kept frozen during the training procedure. The method yielded state-of-the-art results on datasets like NDIRIS, IITD Iris Dataset etc.
The paper modified the Triplet Loss and added domain knowledge for IRIS detection to increase the accuracy of the system! The original paper is available at this link.