Person Re-Identification: A Lightweight Feature Extraction Architecture for Image Based Methods

Person re-identification network is widely used in person search method where these networks find correct person match from a gallery of images. Person re-identification has gained more attention among the researchers in recent years because of its use in security and other purposes. Existing methodologies in this area focuses on dealing with various challenges of person re-identification such as occlusion, lighting, camera viewpoint, clothing changes, etc. from different perspectives. In this article, we have used the Siamese model architecture in person re-identification. Different feature extractor models for the twin architecture have been implemented for learning to generate relevant feature vectors. An existing model has been used to improve its performance and proposed a new lightweight model architecture for feature extraction. Additionally, both model's performances have been analyzed by varying different hyper-parameters. Further, pre-trained models have also been used in the architecture and their performances comparison with other models is presented. In the whole process, A subset of publicly available MARS dataset is used while ensuring that it resembles all the qualities of the original dataset. Finally, it is presented that increasing parameters of the model can provide good results although almost similar outcome can be achieved with a decent smaller architecture too.