StyleGAN-LSRO Method for Person Re-Identification

Abstract

In this study, the StyleGAN-LSRO method has been developed for person re-identification (re-ID) tasks. This method applies the style-based generative adversarial network (StyleGAN) to generate new synthetic images from existing person re-ID datasets and the label smoothing regularization for outliers (LSRO) algorithm to process those newly produced unlabeled images by assigning them a uniform label distribution along with the definition of a loss function for the training process. A baseline model based on a convolutional neural network (CNN) was developed to learn the discriminative features to recognize a person's identity. The developed method has been tested on three datasets. These datasets are Market-1501, DukeMTMC-reID, and MSMT17. The experimental results show that the StyleGAN model achieved a Fréchet inception distance score of 12.67 and structural similarity score of 0.387, outperforming all the previous generative methods and demonstrating that the images generated by StyleGAN are of superior quality. Adding these StyleGAN-generated data significantly improves the person re-ID accuracy. The StyleGAN-LSRO person re-ID method achieved 98.5% rank-1 accuracy and 91.8% mean average precision (mAP) on Market-1501, 87.0% rank-1 accuracy and 83.8% mAP on DukeMTMC-reID, and 81.5% rank-1 accuracy and 60.9% mAP on MSMT17, respectively. These results show that the StyleGAN-LSRO method significantly outperforms most of the state-of-the-art person re-ID methods. The success rate for person re-ID increases when the images used are of high resolution and square matrix form. In other cases, the success rate decreases.