About the role of prior file for clevr dataset and how to get them
Ganlin-Yang opened this issue · 1 comments
Hello, I find that for clevr dataset, you don't get the translation t in the transformation {s,t,R} through random sampling, while you get s and R through random sampling during training process. Instead, you sample the translation t from a prior file, which consists of a large number of coordinate value to sample from. I have 2 questions about this: firstly, what is the role of this prior file? Why can't we obtain t just like s and R? secondly, I see that the prior file is attached with the images when download the dataset, but if we only have image dataset, how can we obtain the prior file from this? In other words, where is such a prior file come from?
We saved this during dataset generation. Yes, you are right, we could also just sample these from prior distributions. For multiple objects, you need to take care of a few things (e.g. checking for collisions, etc) such that we simply used the file which acts as an approximation for a continuous distribution.