Train on custom datasets

[primecai]

Hi,

Many thanks for releasing the awesome work! I would like to try GSN on some real world scenes, e.g. RealEstate10K, to see whether GSN can at lease produce some results. However, I am a bit struggling to understand whether some processing regarding the camera poses provided by RealEstate10K is needed---is there any essential difference between habitat-sim camera poses, and RealEstate10k style poses? A few doubts I'm having:

1. I see the principal points are all [0, 0] for both Vizdoom and Replica datasets. From the code I guess [0.5, 0.5] is always assumed?
2. Are the camera transformations world 3D coordinates to 2D projection, or camera local coordinates to 2D projection? From the code I get that it should be w2c transformation, is it correct? Any insight regarding how should the parameters be modified will be very helpful.

Best.