RAB implementation

Question

RAB implementation

Opened this issue 3 years ago · 1 comments

Hi!
In the paper you say about using RAB as the refine net which described in equation 17.
But in current (deleted) implementation you use 5 stacked conv+relu blocks instead of RAB.
Is there any particular reason for that? Does stacked conv-relu block perform better?

Answer 1 · 2021-05-24T02:15:29.000Z

Hi,
sorry, the current version is not our final implementation. Since we construct many ablation studies, the code may be confused. In fact, compared with RAB, using the 5 stacked conv+relu blocks can obtain similar reconstruction performance.