Thank you very much for open-sourcing the relevant code, datasets, and pre-trained models.

I used your open-source pre-trained model to infer directly on the public datasets including CASIAv1, COVERAGE, NIST16, IMD20, Columbia, and there is a big gap between the evaluation results and the metrics reported in the paper, and I also tried to retrain the dataset you provided, but generalization is hard to guarantee. Hope you can check the uploaded pretrained model, or provide the code for metric evaluation.

Hello and thank you for your very interesting work,

While I was experimenting with the provided pre-trained models I found out that the results were worse when the images were utilized on their original size compared to when they were limited to 256x256 (by retaining aspect ratio). An example dataset to reproduce the issue is Columbia.

So, I would like to ask you what is the input size that leads to the best results? What was the one utilized during the evaluation reported in the paper?

Hello author, I used the code you provided for testing, but the coverage AUC was only 65, which is far from the article. Is there something wrong with my AUC evaluation method?

I want to retrain your network, but after the loss enters hundreds of pictures, the loss does not change. What is the reason for this problem?

您好，想问一下您代码中给出的checkpoint中的参数是只预训练过，还是已经微调过的？我用原参数计算的f1分数只有四十多，auc达到80多，但感觉我计算的不对劲，方便的话可以给评价指标的代码吗？

How soon can the source code be released? Could the pre-trained model and evaluation code be released first? Looking forward to your reply. Best regards.

Hi, can you please tell how you converted the tricolor groundtruth masks to binary masks for Columbia dataset?

The datasets(splice) you provided doesn't exist the floder of mask,Can you provide it ?Thanks!!

Hello, why is the result like this when I run the train file?

=> loading HRNet pretrained model models/hrnet_w18_small_v2.pth
HRNet weight-loading succeeds: ./checkpoint/HRNet_checkpoint/HRNet.pth
NLCDetection weight-loading succeeds: ./checkpoint/NLCDetection_checkpoint/NLCDetection.pth
DetectionHead weight-loading succeeds: ./checkpoint/DetectionHead_checkpoint/DetectionHead.pth
length of traindata: 500
previous_score 0.9871
authentic_ratio: 0.25 fake_ratio: 0.75
resuming FENet by loading epoch 30
resuming SegNet by loading epoch 30
resuming ClsNet by loading epoch 30

Process finished with exit code 0

1

作者你好，不知道是不是你更新了新的代码以后把之前的环境配置相关信息给覆盖掉了，目前这个代码没有环境配置的具体信息，请问你可以提供一份这个代码的相关环境配置吗？
感谢！

I am really working on this code and trying to further contribute but have some issue with the following values.

pscc_args.val_num = 200
pscc_args.train_num = 100000

authentic, splice, copymove, removal

pscc_args.train_ratio = [0.25, 0.25, 0.25, 0.25]

May I know why you make the train_num as 100000 as I checked. The copy-move list size is 100000 while for authentic, the size is 81910. So I am confused, what the train_num value should be.

Next, on what basis you chose val_num as 200?

Furthermore, I am trying to make only two classes instead of the 4 you mentioned. so what will be train_ratio? for instance, you used [0.25, 0.25, 0.25, 0.25] for authentic, splice, copymove, removal. Does it mean, it will take .25 % of sample from these? Will be good if you elaborate this too.

Many Many thanks in Advance.

I come across the same problem with issue3I tried to retrain, but it doesn't work. How to solve the problem？

good

Hello! I would like to ask on which dataset the pre-trained model released in the repository is trained? Also can you release the test code used to calculate the evaluation metrics?

this paper is very good .Excuse me, can you publish the heat map, F1, testing and fine-tuning code on small data sets in the paper?

My AUC on coverage is only 68%, which is far from your 84.7%. What could be the reason for this significant difference? Have you adjusted other parameters or anything? I trained the model using the original parameters.
expect for you answer！

我在coverage上的auc只有68%，和你的84.7%相去甚远，请问这是因为什么原因？请问你有调整其他参数之类的吗？我是按照原始参数训练出来的模型

Thank you for opening the source code of your work. This work is excellent. I downloaded the code and parameter weights you provided. Because I didn't see your evaluation code, I used my own evaluation indicator code, and the test results are somewhat different from your results. The following is my evaluation index method and test results. Can you share the code of your evaluation or point out my mistakes

The CopyMove.zip folder in the Baidu Cloud Link is corrupted. Cannot extract the Copymove images. Can you fix the CopyMove images?

Hi, @proteus1991 , I find there are some undefined variables in the test.py and could you please check it? thank you !

Hi,

I wanted to download the datasets that you provided in Baidu Cloud, however, I don't have an account with it and registration requires a cell number based in China. Could you provide me other link having the same contents?

Thank you.

您好，我有两个问题想请教您一下：
1.代码中评估用的“score”计算方法似乎没有在论文中体现，请问这个评价指标如何去理解，pth模型在各个数据集上的“score”值的测试能否提供官方数据；

2.论文中F1和auc的值的计算，您是否能提供一下测试程序，计算结果似乎有一定偏差。

Hello, I have two questions that I would like to ask you:

1.The "score" calculation method used for evaluation in the code does not seem to be reflected in the paper. How to understand this evaluation indicator, and can official data be provided for testing the "score" values of the PTH model on various datasets;

2.Can you provide a testing program for the calculation of F1 and AUC values in the paper? The calculation results seem to have some deviation.