A Study of 3D-GANs and their Implementation Challenges

  • Vrushali Ghodnadikar , Aditi Thopte, Prabhdeep Singh Gandhi, Siddhant Shah, Shailesh Bendale


Computer vision is undoubtedly the most researched field in Artificial Intelligence, the introduction of GANs was intended for applications in this domain as well. General Adversarial networks as a paradigm are intriguing as it takes the old adversarial learning concept and brings new perspectives on it. From human faces to mortgages, GANs are hired to produce a wide range of images. Speaking of the emergence of GANs, since its inception in 2014, various approaches have been developed by bright minds in this field to gain convergence. It has a number of challenges that need to be addressed before we can acquire the same ability as other well-known methods of deep learning. In this research paper, we focus on the generation of 3D models using GANs. They sample the distribution of sound evenly while producing these models. Because the strength of the GANs is based on a random sample, it becomes more difficult to make the desired result and reach it over the spectrum. And, since human understanding mainly benefits from triple-digit numerical data, our focus will be on using two-dimensional drawings to produce and rebuild 3D characters. Finally, as noted earlier, GAN training is a way to punish a wild horse! So, basically we will learn different ways to achieve the same. To conclude, this paper will contribute to a deeper understanding of the philosophy and engineering behind Generative Adversarial Networks, their unique challenges, and will end with a focus on 3D modeling.