Pages: 14-20

Abstract

Nowadays, deep learning networks have drawn a lot of attentions in the area of person re-identification. It is attempted in the present article to improve a deep learning neural network so that it can gain more distinct and better characteristics and reach a higher accuracy level. The network simultaneously deals with feature learning and subsequent feature comparison. The use of generalized pooling layer in lieu of ordinary pooling layer has been an innovation proposed in this article. There are two types of generalized pooling used in this network. Tree pooling has been the method of choice in the initial layers and the proposed network is trained with pooling filters and a combination of the generalized pooling and pooling filters so that the network could be responsive. In the final layers, we have used gated max-average pooling with which the network is trained via a gating mask during the learning process so as to finally reach a relative composition of the two types of pooling, i.e. max pooling and average pooling. The network has succeeded in acquiring very much better results on such large datasets as CUHK03 and CUHK01 in comparison to its preliminary state; the network also offers higher accuracy on such smaller datasets as VIPeR in contrast to its mainstream method but the accuracy enhancement is rather trivial due to the few numbers of the data.