Advancements in Synthetic Video Generation for Autonomous Driving
4 METHODOLOGY
Our methodology to generate realistic data is a GAN ( Generative Adversarial Networks ) -based approach in which , given an input image , the generator first synthesizes a fake image . Then , the fake image is combined with the input image and put into the discriminator , which should be able to determine that it ’ s fake . On the other hand , when a real image is used , the discriminator should be able to give output as it is real . So , during generator training , the generator ’ s goal is to fool the discriminator and make its output real .
4.1 GENERATIVE ADVERSARIAL NETWORKS
Generative adversarial networks ( GANs ) are the models that generate new data instances , which is like a GAN-based approach to generating synthetic outputs ; that is why we call these models ‘ generative .’ For example , GANs can create fake outputs that are like the real world , even when these outputs do not belong to any real-world scenario . GANs reduce the reality gap by combining a generator , that trains the model to produce an output , with a discriminator , that trains that model to differentiate real data from the generator ’ s fake output . The generator trains to fool discriminators , and the discriminator tries to discriminate as much as possible to not be fooled .
• Generative models can generate new data instances by taking joint probability .
• Discriminative models discriminate distinct kinds of data instances by using conditional probability .
4.2 TRAINING OF GENERATIVE ADVERSARIAL NETWORKS
GANs training requires both generator and discriminator training simultaneously over every epoch . It follows a below pattern : 1 . Training of discriminator for n epochs . 2 . Training of generator for n epochs . 3 . Repeat the training process as in 1 and 2 .
We usually keep the generator constant during the training of the discriminator and vice versa .
4.3 OBJECTIVE FUNCTIONS IN COMMONLY USED GANS
The objective function in GANs consists of the commonly used Loss function known as Minimax Loss . The generator tries to minimize the objective function value and the discriminator tries to maximize it . min max L ( G , D ) = E X [ log ( D ( x ))] + E Z [ log ( 1 − D ( G ( z )))]
( 1 )
118 March 2024