Boosted verification using siamese neural network with DiffBlock

On face recognition, person and vehicle re-identification tasks, different networks and losses have been proposed to learn better features, which further maximizes the decision margin in the feature space. Despite the promising progress having been made, it still remains a challenge to discriminate the different but similar targets while recognizing the same but dissimilar objects, which results from the contradiction between the information retention and the intra-/inter-class distance optimization in the static feature representation methods. The similarity of the static features is insufficient to represent the relationship between diverse images. In this paper, a novel DiffBlock module is proposed to compare the pairwise intermediate features and amplify the difference between the samples. Then SNND (siamese neural network with DiffBlock) is proposed to progressively dig out the discriminative information and judge the relationship between the samples precisely. Extensive experiments on multiple benchmarks for face, person and vehicle verification show that our proposed SNND significantly outperforms previous state-of-the-art methods.