fanghannus / pimog-an-effective-screen-shooting-noise-layer-simulation-for-deep-learning-based-watermarking-netw Goto Github PK

In your article, 𝜆1, 𝜆2, 𝜆3, 𝜆4 are set as 0.5, 2, 3, 0.001 by default. however, In the code you provided, the mask corresponds to the Gradient mask generator with a weight of 0.5, but in the article it is 2, corresponding to v_ The pre trained edge mask generator corresponding to the mask has a weight of 2, but in the paper it is 0.5. Which one is correct?

In direction 2 3 4, the k in np.rot90() should be 1 2 3 but 1 1 1 now.

Hello, author. I noticed a significant difference between the edge masks produced by the BDCN mentioned in the paper and the examples you provided. Could you please share the code for generating the dataset? I am interested in training your model for comparative experiments. Thank you very much!

Hello, using the model provided by the author for the embedding test, the result obtained has five images in it, why is the perspective distortion not reflected in the noise image

方老师，您好。自从做这个方向以后也看了您的很多工作，启发很大。想问下您关于PIMoG这个工作您在训练的过程中有没有遇到噪声层计算速度慢，导致每一轮训练需要花费较多时间的情况，如果有，想问下您是怎样进行处理的。

Hello, using the model provided by the author for the embedding test, the result obtained has five images in it, why is the perspective distortion not reflected in the noise image

Thanks for your excellent work on screen-shooting resilient watermarking.

In this paper:

the detection and correction process of watermarked images are the same as the settings of StegaStamp.

So could you share the pre-trained segmentation model for testing?

line 189 in Noise_Layer.py
max_len = np.max([np.sqrt(x**2+y**2),np.sqrt((x-255)**2+y**2),np.sqrt(x**2+(y-255)**2),np.sqrt((x-255)**2+(y-255)**2)])

According to the paper, this line calculates max distance from a certain point to four corners of the image. The number 255 in this line should be the width and height of the image.