Improving face verification and person re-identification accuracy using hyperplane similarity

The standard framework for using a convolutional neural network (CNN) for face verification is to compare the feature vectors taken from the penultimate network layer of a CNN trained to classify the identity of an input face using a softmax loss over identities. Feature vectors are typically compared using the simple L2 distance. We demonstrate that the L2 distance is not the best distance to use in this scenario, and propose the hyperplane similarity as a more appropriate similarity function that is derived from the softmax loss function used to train the network. We demonstrate that hyperplane similarity improves verification results especially for low false acceptance rates which are usually the most important operating regimes for real applications. We also propose a fast algorithm for finding the separating hyperplanes needed to compute hyperplane similarity.