Advancements in Synthetic Video Generation for Autonomous Driving
Video-to-video synthesis proposed by T . C . Wang et al . [ 5 ] discusses an approach using Generative Adversarial Networks ( GANs ) 16 . The study employs sequential generator architecture to achieve spatial-temporal coherence . It converts input-segmented video to realistic video . For lower-resolution images , inputs are previously generated output frames and semantic images . The label maps are combined to form intermediate high-level features that are processed from various residual blocks , as shown in Figure 2-3 . It applies similar residual-based processing for the previous images . Then , after combining intermediate layers , it is again processed by two different residual networks , which gives an output as the intermediate image , the mask , and the flow map .
Figure
2-3 : Architecture for low resolution outputs . [ 5 ]
As explained earlier , it uses similar configurations above the low-resolution network for higherresolution results . It first down-samples the inputs and feeds them into the low-resolution network .
Figure 2-4 : Architecture for high resolution outputs G1 corresponds to lower architecture network and G2 corresponds to higher resolution network . [ 5 ]
16 https :// arxiv . org / abs / 1406.2661 Journal of Innovation 111