An Efficient Siamese Network for Pet Identity Verification
Shun Li, Chenyang Ge · 2022
Vision-based pattern identification (such as face, fingerprint, iris etc.) has been successfully applied in human biometrics for a long history. However, pet identification is still a challenging problem due to the lack of labeled data and the existence of imbalanced sample classes in our available dataset. This paper overcomes the above difficulties and presents an efficient method for the 1 vs 1 pet identity verification mission launched by CVPR 2022 pet biometric. We use ArcFace method to select some hard samples at first. Then we make use of the hard samples obtained to construct positive and negative sample pairs. Finally, a Siamese network has been proposed to classify the sample pairs and trained to classify the image pairs of test sets. Meanwhile, in order to solve the problem that each class only with few samples in the training set, we propose an offline data augmentation strategy, which greatly improved the classification performance. Our methods achieve 88.3% AUC on test set 1 and 81.6% AUC on test set 2, outperforming some SOTA Re-ID methods and winning the thirteenth place in the challenge.