Potential Bugs for the Point Postion Prompting

[DerrickXuNu]

Hello Meta Research Team,

First of all, thanks for open-sourcing such great work, which is really beneficial for the whole community! When I read the code line by line, I found one place a bit confusing. When I go through the amg.py, I found when you generate the coordinates, you are multiplying the normalized coordinates with the new image size without padding (e.g. 1024, 972). Then when you generate the point embedding, you normalize the transformed coordinates with the image size that has padding (1024, 1024). Would this slightly influence the performance? Look forward to your reply!