bcmi / dirl-inharmonious-region-localization Goto Github PK

在运行程序的过程中，我发现评估是采用把图像尺寸规定到224*224来计算评估分数的，请问为什么不将图像还原到原本图像的大小来计算呢？如果说网络的输入输出不是224，而是其他的话，这样的评估是不是可能会产生一定的误差影响？

May I ask where can I find the results on four sub-datasets in Supplementary in the Inharmonious Region Localization by Magnifying Domain Discrepancy article on AAAI?

Hello @jimleungjing,

Thanks for providing the implementation. I'm trying to re-train DIRL network by changing the backbone from Resnet34 to Resnet50 but it didn't work. Could you please comment on this on how to change the backbone. It seems we have to make some changes in the code in order for it to work?

Thanks

Hi,

Thanks for wonderful work and implementation. I've a question about the model behavior. I trained the model on cloud using a good GPU. The f1 score is good (0.7) which is acceptable for me. When I try to run the trained weights (encoder and decoder) on different System (CPU only and GPU only too). The performance drops significantly and the model is not giving same segmentation results as compared to the system where it was trained. Can you comment on this issue, why the performance decreasing on different system???

Thanks!
Shan

Hello,

Thanks for providing the implementation. Your work is very interesting. I'm interested to re-train this network to add object detection head at the end of this network in order to predict (mask, bounding box and class label) in the single network from tampered images. Can you please point where should I begin from? I want to use existing resnet features for detection + classification.

Any advice would be appreciated.
Thanks.

Very interesting work, some doubts as shown in the title.

It also makes me wonder that you said "the ground-truth mask M is resized to match the resolution of each A_{k}" in your paper, but I found that the code seemed to do the opposite.

Looking forward to your reply!

I am try DIRL with my own little test dataset. In 48 total here is only 2 success and around 45 false positive alarms.

For example: (fake, mask, DIRL result)