Large-Scale Bisample Learning on ID Versus Spot Face Recognition

Abstract

In real-world face recognition applications, there is a tremendous amount of data with two images for each person. One is an ID photo for face enrollment, and the other is a probe photo captured on spot. Most existing methods are designed for training data with limited breadth (a relatively small number of classes) and sufficient depth (many samples for each class). They would meet great challenges on ID versus Spot (IvS) data, including the under-represented intraclass variations and an excessive demand on computing devices. In this paper, we propose a deep learning based large-scale bisample learning (LBL) method for IvS face recognition. To tackle the bisample problem with only two samples for each class, a classification-verificationclassification (CVC) training strategy is proposed to progressively enhance the IvS performance. Besides, a dominant prototype softmax (DP-softmax) is incorporated to make the deep learning scalable on large-scale classes. We conduct LBL on a IvS face dataset with more than two million identities. Experimental results show the proposed method achieves superior performance to Fig. 1: The ID versus Spot (IvS) data, each identity has one ID photo and one spot photo. previous ones, validating the effectiveness of LBL on IvS face recognition.

DocFace+: ID Document to Selfie Matching

“…Additionally, our method is not designed for any specific degradation type and could be applied to any type of ID document photos. Concurrent with our prior work [13], Zhu et al [14] also worked on deep CNN-based ID-selfie matching systems. With 2.5M ID-selfie pairs, also from a private Chinese ID card dataset, they formulated it as a bisample learning problem and proposed to train the network in three stages: (1) pre-learning (classification) on general face datasets, (2) transfer learning (verification) and (3) fine-grained learning (classification).…”

“…Hence, the earlier published results on ID-selfie matching are now obsolete. To the best of our knowledge, our prior work [13] is the first to investigate the application of deep CNN to this problem, concurrent with Zhu et al [14].…”

“…Our work, on the other hand, proposes a special optimization method to address the slow convergence problem of classification-based embedding learning methods on ID-selfie datasets, which does not require multi-stage training. Compared with our prior work [13], the differences of this version are as follows: (1) a larger IDselfie dataset (over 50,000 papers), which is a combination of the two small private datasets in [13] and another larger ID-selfie dataset, (2) a different loss function, namely DIAM-Softmax, to learn the face representation, (3) more comprehensive experiments to analyze the effect of each module and (4) an evaluation of the proposed system as well as other face matchers on a new ID-selfie benchmark, Public IvS, released by Zhu et al in [14].…”

“…Public IvS is a dataset released by Zhu et al [14] for evaluation of ID-selfie matching systems. The dataset is constructed by collecting ID photos and live face photos of Chinese personalities from Internet.…”

“…We use a large dataset of Chinese Identity Cards with corresponding selfies 2 to develop the system and to evaluate the performance of (i) two Commercial-Off-The-Shelf (COTS) face matchers, (ii) open-source deep network face matchers, and (iii) the proposed method. We also compare the proposed system on the open benchmark established by Zhu et al [14]. The contributions of the paper are summarized below:…”