Advancements in Synthetic Video Generation for Autonomous Driving
The Multi SPADE residual block is explained in detail in Figure 2.7 . Each block is subdivided into multiple SPADE Residual blocks ( 3 in this case ).
A . Mallya et al . [ 4 ] significantly improves spatial and temporal consistency , but it lacks various user perspective applications , scenario generation and different AI rendering aspects .
High-Resolution Image Synthesis and Semantic Manipulation with Conditional GANs proposed by T . C . Wang et al . [ 12 ] present an image-to-image translation method that uses a conditional GANbased approach to generate high-resolution , realistic images from semantic labels . This study helps understand various scenario-changing possibilities , high-quality outputs , and an interactive environment , which would be helpful from the user ' s perspective . It uses instance and label mapping both as input to get sharper as well as easily distinguisable outputs .
Figure 2-8 : Instance as well as label used for encoding . [ 12 ]
We will use a similar approach for our video synthesis approach , along with previous work on video-to-video synthesis .
Park et . al . [ 13 ] proposed a method also known as GauGAN , which shows how Spatially Adaptive Normalization ( SPADE ), with the help of GAN , can help us remove the ‘ wash away ’ problem in normalization layers and finally help us to reduce the problems in dealing with morph images and consistency with respect to images .
114 March 2024