I have a question.

[Hanhanhan11]

Hi, I have a question that the code directly choose the random image and its previous four frames, where does the code control that the previous images are exactly in the datasets?
For example, if the selected image's frame number is 5, its previous frames are 1,2,3,4, maybe the 1 frame is not in the datasets.

[lxtGH]

Cityscape dataset have frames range in (0,29) 19 is current frame with fine annotation.

[lxtGH]

@D-Nilsson Hi! In your paper, STGRU modules are trained with continous frame? Here I see the above code is using like frame pair.