The performance of the retinanet

[WangJian981002]

@yhenon-nextdroid @yhenon @xu1718191411 @mimoralea @rvandeghen Thanks to your great works! But i got some training issues in this repo. The original retinanet using SGD and finetune 12epochs on the ImageNet pretrained Resnet. I change the training settings follow the paper, after solving some small convergence problems, my recurring performance is much lower than the paper (ap30.5% vs 34.3% in 600 px). For this I checked all the relevant details in the code, but did not find any problems, do you know how to solve this problem？Looking forward to your reply.

[yhenon]

I don't really have time to work on this project, but I believe the gap in performance is due to the limited data augmentation in this implementation (only flipping). It is currently lacking:

• any kind if size/scale augmentation
• any kind of color/brightness augmentation There could also be more advanced augmentation schemes applied. I'd love to see a PR on this, but I can't make time for it right now.