Giters

pengzhou1108 / RGB-N

Geek Repo:Geek Repo

Github PK Tool:Github PK Tool

the FP of authentic image

chengxius opened this issue · comments

commented

hello, peng
I have some questions while I learn your model. I tested fifty authentic coco images with your model , the confidence of the highest score box of twenty-four images is above 0.5, I think the situation as the FP of authentic image . In our application scenario, we hope the probability of authentic image detected as tampered image can be as low as possible. I have tried many methods to drop the FP, however all failed .I don't know if you have considered this problem，and how can I handle with the problem?
Looking forward to your reply ， thank you!

commented

Hi @chengxius

  1. On how any iterations you have trained the model ? I have also trained the model on 110k iterations.

  2. Have you used COCO synthetic dataset for training provided by the author ?

  3. Have you fine tuned on further datasets such as COVER, Columbia, CASIA etc ?

Thanks

commented

Hi @chengxius

  1. On how any iterations you have trained the model ? I have also trained the model on 110k iterations.
  2. Have you used COCO synthetic dataset for training provided by the author ?
  3. Have you fine tuned on further datasets such as COVER, Columbia, CASIA etc ?

Thanks

yes, I have tried iterations more than 110k and kinds of datasets ,but don't work out .

commented

Hi @chengxius ,
Have you solved the problem of false positive? Could you please share some solutions? I have also encountered similar problem in my application.
Thanks!