GAN Based Try-On System: Improving CAGAN Towards Commercial Viability

In the era of disruptive innovations where technological leaps in internet-based applications along with advanced Artificial intelligence and Computing are combining to produce consumer-based applications, there is a notable shift of the fashion retail industry towards e-commerce. However, it still falls behind the traditional experience of shopping which offers customers not just a more visual and tactile stimuli, but also assurance of their purchase. This gap in experience can therefore be attributed to the lack of online try on solutions with online retail industry giants. This generates a disconnect between the retailers and consumers, making them hesitant and hence, skeptical about ordering via e-commerce. Further, consumer misjudgment of products results in returns, which are undesirable for online retailers with thin margins. We in this paper propose a prototypical solution to this problem by introducing a photorealistic virtual try-on solution while ensuring optimization and minimalistic use of resources as required by industrial standards. We present a Generative Adversarial Network (GAN) built upon the existing CAGAN architecture, with an improved generator based on a reworked U-Net generator architecture and a combination of instance and batch normalization for generator and discriminator respectively. Further, a combination of ReLu and Leaky ReLu activation functions have been utilized for the generator along with hyperparameter optimization of the network. In our visual analysis of the generated images, comparing with the CAGAN approach on Zalando dataset, we have achieved a superior quality in shorter training duration of 13 hours on a smaller dataset of 10,167 images.