PROPOSAL AND IMPLEMENTATION OF AN EFFECTIVE CNN BASELINE FOR PERSON RE- IDENTIFICATION

Abstract

Person re-identification is a challenging task due to the critical issues of human pose variation, human body occlusion, camera view variation, etc. To deal with this, most of the state-of-the-art methods based on the deep convolutional neural networks have strong feature extraction and classification capacity. However, there are not enough studies about building an effective CNN baseline model. There are three good practices are followed in this work for building and effective CNN architecture. These practices are adding batch normalization after the global pooling layer, use only one fully connected layer for classification and use Adam optimizer. Using these three techniques in the implementation, the performance of a simple pre-trained CNN model have been enhanced without making any high level changes and experimental results supports this argument.