Li-Qingyun
commented
1 year ago I notice that AP of DINO-4scale using r50 is 49.0% in table 1, while DINO (ours, Row5+contrastive DN) in table 4 is 47.9%. Which setting or model design is modified? It seems the outcome for this project is 47.9%， while mmdetection appears to have a result of 49.0%. I am not sure if the code is different in these projects.
you can consider it as a typo
in fact, a bug about init MSDA weight was contained in early DINO implementation, hence the early result is lower. The authors then found the bug and achieved better performance. They seem forget to modify at some places.
In mmdetection, we followed the new implementation. The code is different, but the theory is the same.
I notice that AP of DINO-4scale using r50 is 49.0% in table 1, while DINO (ours, Row5+contrastive DN) in table 4 is 47.9%. Which setting or model design is modified? It seems the outcome for this project is 47.9%， while mmdetection appears to have a result of 49.0%. I am not sure if the code is different in these projects.