Bringing Creativity, Agility, and Efficiency with Generative AI in Industries 24th Edition | Page 127

Advancements in Synthetic Video Generation for Autonomous Driving
4.6 OBJECTIVE FUNCTIONS USED IN PROPOSED GAN
The objective functions in GANs consist of commonly used Loss functions known as Hinge Loss , feature extractor loss , and flow loss , like what was used in existing video-to-video synthesis methods [ 4 ] - [ 6 ]. Hinge loss is calculated as in ( 2 ). The generator tries to minimize , and the discriminator tries to maximize the objective function .
L Hinge ( G , D ) = E X
[ max ( 0 , ( 1 − D ( x ))] + E z [ max ( 0 , ( 1 + D ( G ( z )))]
( 2 )
Here Ex , Ez are expected value real inputs and random inputs , respectively , D ( x ) and D ( G ( z )) are the expected estimate of discriminator for real instance to be real and fake instance to be real , respectively .
Feature extractor loss , as in ( 3 ), is discriminator feature based on VGG19 Loss , which is calculated using the below formula . ∅ i ( G ) , ∅ i ( R ) are features of VGG19 layers of generated and real images .
L vgg = | ∅ i ( G ) − ∅ i ( R )| ( 3 )
Flow loss is based on Flownet2.0 21 architecture which contains :
• Flow L1 loss : L1 loss with respect to ground truth flow .
• Flow warp loss : L1 loss between warped and target image .
• Flow mask loss : Loss of occlusion mask .
Calculation of each flow loss is done by as by changing respective input values x and y in ( 4 ). L ( x , y ) = { L 1 , L 2 , … L n } T , L n = | x n − y n |
( 4 )
Combining all these objective functions , the generator aims to minimize , and the discriminator maximizes the function value to get a model with good generative capability .
21 https :// arxiv . org / abs / 1612.01925 122
March 2024