About the CNN+FPN

[gaiyi7788]

Excuse me. I have tried the code about this work. However, what confused me is that the paper mentions to use CNN+FPN structure for feature extraction, but I didn't find the FPN in the code section. In fact, after I added FPN to the code myself, I found that the quality of the generated pseudo-labels decreased. Can you help me to explain this, thank you very much!

[briliantnugraha]

Hi, CMIIW, just a conjecture on my part, but it seems that this repo is still in progress/still partially shared (as I've different error, you can see the other issue in this repo).

On the other hand, I've also tried to implement it myself when the repo is still not yet available. What I did (following the paper) was using anchor-detr, and add exemplar feature+feature aggregation at stage 1 as shown below.
CMIIW, However, in this repo, the stage 1 seems to be object detection training, neither exemplar features nor feature aggregation appears in the code. Here are the refs: backbone stage1, overall flow stage1, backbone stage2.

Will be glad to hear more discussion from the author(s) @NguyenVanThanhHust to shed some light for us, thank you~