Review of Optimization Strategy in Pose Transfer

Optimization strategy in Generative Adversarial Network (GAN) for pose transfer is crucial for improving the model performance, enhancing the quality of the person image generated, and achieving better evaluation metrics. Three optimization strategies are widely applied in GAN improvement, i.e. (a) architecture, (b) loss function, and (c) optimizer. Many variants of architecture, loss function, and optimizer have been proposed to further enhance the model performance focusing on mode collapse, vanishing gradient, and image quality. In this paper, the commonly used optimization strategies such as architecture, loss function, and optimizer in GAN for pose transfer are reviewed. A comparative study of the quantitative performance obtained by various GAN architectures with the respective optimization strategies on the DeepFashion dataset is summarized in chronological order. The optimization strategies in the Pose with Style, SCA-GAN, BiGraphGAN, and XingGAN are the current state-of-the-art methods for pose transfer based on the quantitative results obtained. The potential future development of optimization in GAN for pose transfer is concluded to provide insights into GAN development.