PERSON RE-IDENTIFICATION USING DENSELY CONNECTED CONVOLUTIONAL NEURAL NETWORK

Abstract

A successful system for person re-identification inspired by [67], a Densely Connected Convolutional Neural Networks (DenseNet) have been developed. This architecture was proposed by Huang et el. [67] (2017) for object recognition. We are using this model for exploring the network configurations and settings for getting a better solution for person re-identification tasks. We will train and test different person re-identification datasets, search for optimal settings and other factors affecting the result. In this work, we are using two different network configurations of DenseNet model i.e., DenseNet-121 and DenseNet-161 with growth rate of 32 and 48 respectively. The model is trained and tested on different RE-ID datasets which are CUHK01 [31], MARS [44], VIPeR [27] and Market1501 [29].