Influence of Training Times and Sample Size on Results in Visual Try-On with GAN

Abstract

With online shopping becoming more and more trendy these days, Virtual Try-On Approach is proposed. It is a method of Generative Adversarial Network (GAN) composes of two models: Geometry Matching Module (GMM) and Try-On Module (TOM). It is proposed to solve the inconvenience of customers for not being able to see the fitting effect before receiving the goods. This paper is about finding the relationship between dataset and the effect of the experiment in the GAN experiment. After implementing the models and gaining the result images, we do some observations on the GMM one and try to analyze it and improve the effect. Comparison is made on different aspects: training times of the model and sample size of the training input. Similar trends are shown on the decrease of loss value when the training times or sample size rises. However, the visual effect is not necessarily improved all the time. Compared with the times it has been trained, the improvement from training times is far less than that from the sample size. Finally, we draw the conclusion that the effect is gradually improved as the training times or sample size increases at the beginning. Nevertheless, the improvement cannot be recognized evidently when the training times and sample size have grown large enough. Additionally, the sample size seems to be a more significant factor in the improvement of the visual effect than the training times.