search

er-muyue / DeFRCN

MIT License
181 stars 43 forks source link

Importance of w,w/o FPN in DeFRCN #26

Closed buncybunny closed 1 year ago

buncybunny commented 2 years ago

Hey there! Thanks a lot for the quality codes.

I'm wondering why you guys have taken out FPN from FRCN. Have you guys done the same experiments with FPN? (wondering if GDL or PCB were afftected by it or all that jazz...)

I couldn't find any details about this in the paper.

MartinYYYYan commented 2 years ago

@buncybunny I have tried to add 3 more Affine layers for each channel of RPN. When I FT the model, I have frozen the backbone(5-levels) and unfrozen the FPN, RPN, ROI-HEAD. The performance is similar to RES5ROI model. (1% difference in Base Class and Novel class mAP50).

ssyze commented 2 years ago

@MartinYYYYan I'm sorry that could you share your FPN config file with me? I also add some more Affine layers for the FPN model, but my performance is lower than the RES5ROI model by about 2-5% in any shot. I'm quite confused about this. Thx a lot!

er-muyue commented 2 years ago

Hi @buncybunny, I think you have the similary question with this issue. Actually, FPN is not a necessary module for verifying the performance of DeFRCN. We directly adopted Res5ROIHeads in our paper. If there is enough time next, I will experiment on FPN and provide the experimental results in this repo.