International Research Journal of Engineering and Technology (IRJET)
e-ISSN: 2395-0056
Volume: 09 Issue: 06 | June 2022
p-ISSN: 2395-0072
www.irjet.net
Creating Objects for Metaverse using GANs and Autoencoders Aditya Singh1, Tejas Patil1 1Student,
Dept. of Computer Engineering, Datta Meghe College of Engineering, Navi Mumbai, Maharashtra, India ---------------------------------------------------------------------***---------------------------------------------------------------------
Abstract – Metaverse refers to online spaces which allow
The following is the general loss function of GAN called Minimax loss there are many other variations of this loss function.
people to interact with one other in a more captivating way than a website. This can be achieved through virtual reality (VR) and augmented reality (AR). To make the Metaverse more connected to reality, we make use of real-life objects that translate into the Metaverse, that's where the need to have state of an art model to achieve this goal. In this paper, we have proposed a way to create such objects using Generative Adversarial Networks (GANs) coupled with an Autoencoder. GANs will create a new object, the object could be anything from animals to human faces, cartoon figures etc. With the creation of such objects, the goal is basically achieved here. However, the output given by the GAN model is not of very high-quality that's where Autoencoder comes in with the use of Autoencoder the object is upscaled (superresolution).
Ex[log(D(x))] + Ez[log(1-D(G(z)))] In the above minimax loss equation, D(x) is the discriminator’s estimate of the probability that the present data is real. Ex is the expected value over all real data instances. G(z) is the output given by the generator from noise G. D(G(z)) is the estimate of the probability that the fake instance is real
1.2 Autoencoders Autoencoder [3] is an interesting variant of a Fully connected neural network, It consists of three parts the encoder, the bottleneck and the decoder. The encoder takes the input and its compression is performed in order to store the spatial data in the bottleneck which only consists of three neurons, Therefore the bottleneck stores the spatial data in a lower dimension. this data is further fed into the decoder which is again a fully connected layer and the original image is generated. The bottleneck is forced to learn important information in order to compress the data into lower dimensions such that using the same data the original image can be generated. Autoencoder just like GAN has many variations depending upon the fully connected layer sometimes it's replaced by pooling layers. Depending upon the variation of Autoencoder the loss function may vary, mentioned below is general a loss function:
Key Words:
Metaverse, Generative Adversarial Networks (GANs), Autoencoder
1. INTRODUCTION Metaverse [1] is like ad immersive world where the user can interact with others, attend a concert etc. which is achieved with AR and VR technologies. In this VR world, we may make our model and react with such different models this paper aims to make this goal more realistic by creating objects based on real-life entities with which these models (people) can react.
1.1 Generative Adversarial Networks (GANs) Generative Adversarial Networks [2] also known as GANs consist of a pair of neural networks namely generator and discriminator that compete with each other. As the name suggests, the generator has to generate images while the discriminator has to detect whether the image generated is real or fake. The discriminator is fed at random real images from the data set it's trained on and images generated by the generator, it has to successfully identify the real image. On the other hand, the generator has to generate a betterquality image such that the discriminator fails to discriminate between the real image and generated image. The basic idea of a GAN model is to create a fake never seen image with help of existing data. The model stops training when the Nash equilibrium is achieved.
© 2022, IRJET
|
Impact Factor value: 7.529
min E(A, B) = min E(xt) = min ∆(A ◦ B(xt), xt)
2. PROPOSED WORK In this paper, we propose a model which will generate realworld objects including human faces with help of a Generative Adversarial Network model and upscale the pixel quality of the generated object with help of Autoencoders, this generated object will be further used in AR/VR application to serve the need of brining real-life feel in Metaverse. For the GAN model, we made use of the Deep Convolutional Generative Adversarial Network (DCGAN) [4] which consists of the deep convolutional layer in the generator and discriminator. In the case of the autoencoder, we made use of both the convolutional layer and pooling layer for both encoder and decoder. the GAN
|
ISO 9001:2008 Certified Journal
|
Page 1595