zhanggongjie / imfa Goto Github PK

hi, @ZhangGongjie , thanks for your enlightening work.

I have a simple question when reading your paper:

In Section 5, there is a sentence describing details:

A crucial implementation detail involves incorporating skip connections for encoded features between Transformer encoder layers, as motivated by [63] and [65, 66] to facilitate feature semantic alignment.

what does this mean? From my point of view, the skip connection is already in the encoder layer itself.

In fact, when I try to reproduce your proposed multiple stacked detection stages structure as shown in figure 2 on models like Deformable-DNDETR and Deformable RankDETR, I didn't get the same performance as table 4 which only gets 0.3 AP degradation, instead, it suffers 2-3 AP drop. I am confused by such difference and I suspect I made some mistakes.

Hope I can get your suggestions. ( And I can't wait to read your source code!)

I checked your state which shows busy, so Sorry to bother you.

How to get the keypoints in formula 1, that is, from mlp,nc(query)->nj*2 (keypoints)? Did it take j mlpS to get j keypoints?

Your paper is very useful for me and thank you for your work!!! @ZhangGongjie

@ZhangGongjie

Can't wait to see your code? How could there be such a perfect idea?

Hello, when will the code be released？

May I ask when the code could be release and we can't wait to try it

Can't wait to see your code? How could there be such a perfect idea?

Can't wait to see your code? How could there be such a perfect idea?