Advancements in Synthetic Video Generation for Autonomous Driving
Figure
2-9 : SPADE normalization . [ 13 ]
The ‘ wash away ’ problem is losing semantic information after the normalization layer . Figure 2-9 defines how segmented mask is not directly fed to batch normalization ; instead , it is convolved and gives modulation parameters to save semantic information . The problem with this approach is that it lacks different styling information . Thus , we have used a modified version of SPADE in our proposed method .
Zhu et al . [ 3 ] describes different styling aspects and the shortcomings of GauGAN [ 13 ] which is a Generative Adversarial Networks having spatially adaptive normalization layer . It is mainly related to the quality of output image and styling corresponding to region of interest . SEAN helps us to improve the quality of synthesized images and region-based style encoding . We will use Semantic Region Adaptive Normalization ( SEAN ) based approach , as shown in Figure 2-10 , to get region-based styling information in our proposed solution . SEAN seems to be a prime solution to use it on our proposed method . We are using its style feature to get diverse quality outputs .
Journal of Innovation 115