Advancements in Synthetic Video Generation for Autonomous Driving
generation , we used pre-trained network DeepLabV3 + 26 . We use FlowNet 2.0 [ 21 ] architecture to extract the optical flow as the ground truth flow . For measuring the closeness of output to the real world , we will be using the following :
• Fréchet Inception Distance ( FID ) -Inception-v3 network 27
• Kanade-Lucas-Tomasi ( KLT ) -Based Score .
5 RESULTS AND CONCLUSIONS
The outputs of scenario generation which include changing environment , extended road and adding realistic objects ( like pedestrians , car , bike etc .) are shown in Figure 5-1 . This is achieved by converting color maps ( semantic maps ) as required . In extended road scenario generation , the proposed system will read color map for road and map the same into surroundings to generate corresponding semantic map and thus generate output . Similarly , for adding the objects as required , a masked area with the right flow and contextual information is added to the color map to get required scenarios .
5.1
SAMPLE FRAME OUTPUT
26 https :// arxiv . org / abs / 1706.05587
27 https :// arxiv . org / abs / 1512.00567 124
March 2024