Verification-based pairwise gait identification

Convolution neural networks (CNNs) have been shown to be effective to produce state of the art results for gait recognition recently, which attracts more and more researchers' attention. However, small inter-class variations and large intra-class variations are still intractable to deal with. Inspired by the Siamese neural network, we propose a coupled deep neural network which combines gait verification with gait identification to improve the accuracy of gait recognition. In this way, the network parameters are learned by using the contrastive loss. Then, we predict whether the input GEIs pair belong to the same person or not and learn the network parameters by using the logistic regression loss. Finally, the gait identification and verification signals are combined with a proper weight to learn the whole network. Experimental results show that our method outperforms the current best approaches by a significant margin.