How to use virtual images?

[Loydian]

Why do monocular datasets use virtual images while nvidia datasets don't? What is the difference between these datasets? And we find that virtual images are crucial to the kid-running case, which is not mentioned in the paper. It would be of great help if you can solve my question.

[zhengqili]

Hi, we described virtual source views in section 4 of the paper and basically virtual views provides stronger geometry information support for moving objects that prevents model stuck in bad local minimal. We believe one reasons is that the camera-object motions relations from real monocular videos offer more ambiguous cues for moving objects when doing volumetric feature aggregation compared with the camera-object motions existing in the Nvidia dataset.

[handsomeZhuang]

how to produce the video using virtual views, and could you share codes with me?