Learning Low-shot facial representations via 2D warping
share
Face recognition has seen a significant improvement by using the deep convolutional neural networks. In this work, we mainly study the influence of the 2D warping module for one-shot face recognition. To achieve this, we first propose a 2D-Warping Layer to generate new features for the novel classes during the training, then fine-tuning the network by adding the recent proposed fisher loss to learn more discriminative features. We evaluate the proposed method on two popular databases for unconstrained face recognition, the Labeled Faces in the Wild (LFW) and the Youtube Faces (YTF) database. In both cases, the proposed method achieves competitive results with the accuracy of 99.25% for LFW and 94.3% for YTF, separately. Moreover, the experimental results on MS-Celeb-1M one-shot faces dataset show that with the proposed method, the model achieves comparable results of 77.92% coverage rate at precision = 99% for the novel classes while still keeps top-1 accuracy of 99.80% for the normal classes.
READ FULL TEXT