Face Identification with Bilinear CNNs

The recent explosive growth in convolutional neural network (CNN) research has produced a variety of new architectures for deep learning. One intriguing new architecture is the bilinear CNN (BCNN), which has shown dramatic performance gains on certain fine-grained recognition problems~\cite{lin2015bilinear}. We apply this new CNN to the challenging new face recognition benchmark, the IARPA Janus Benchmark~A~(IJB-A)~\cite{IJBA}. This is the first widely available public benchmark designed specifically to test face identification in real-world images. It features faces from a large number of identities in challenging real-world conditions. Because the face images were not identified automatically using a computer face detection system, it does not have the bias inherent in such a database. As a result, it includes variations in pose that are more challenging than many other popular benchmarks. In our experiments, we demonstrate the performance of the model trained only on ImageNet, then fine-tuned on the training set of IJB-A, and finally use a moderate-sized external database, FaceScrub~\cite{ng265data}. Another feature of this benchmark is that that the testing data consists of collections of samples of a particular identity. We consider two techniques for pooling samples from these collections to improve performance over using only a single image, and we report results for both methods. Our application of this new CNN to the IJB-A results in gains over the published baselines of this new database.