SinGAN icon indicating copy to clipboard operation
SinGAN copied to clipboard

Calculations about sifid

Open 15732031137 opened this issue 5 years ago • 10 comments

Hello! Congratulations on this paper getting the best ICCV 2019 paper! I recently reproduced your program. I do n’t understand how to calculate the SIFID score. I trained with balloons.png and got 50 pictures. I want to know that your code calculates one picture at a time. 50 times, or do you calculate all 50 at a time, and what do you mean by saying that the real image file name and the fake image file name are the same? In the beginning, I thought your program was to calculate the sifid of the real picture and the 50 pictures it generated at one time, but these 50 pictures were put into a folder, and I could not change the file names of the 50 fake pictures to be the same. Yes, if it is sifid that only calculates one real picture and one fake picture generated by it at a time, should I change the file names of these two pictures to be the same?

wish you a happy life! Looking forward to your reply!

15732031137 avatar Mar 05 '20 08:03 15732031137

@tamarott I understand

15732031137 avatar Mar 05 '20 10:03 15732031137

The sifid function compares a single real image with a single fake one

tamarott avatar Mar 05 '20 22:03 tamarott

@tamarott Thank you for your busy reply! I have understood that I successfully reproduced the SIFID scores of 0.05 and 0.09 in your paper. I would like to ask you how to select a fake picture corresponding to each real picture to calculate the SIFID. Pick any one of the 50 fake images generated by the real picture? Thank you! Your paper is really great! It has helped me a lot!

15732031137 avatar Mar 06 '20 07:03 15732031137

@15732031137 Hello! I see that you successfully reproduced the SIFID scores. Do you mean the images generated in RandomSample? I get the SIFID scores of 0.05 and 0.09 with the images in /Downloads/userstudy folder too, but I have problem with the images generated by myself. I used /Downloads/userstudy/real/1.jpg to generate 50 images, then calculate the sifid of the real and each generated image, but the SIFID scores are all around 8.34

XuXShirley avatar Mar 07 '20 06:03 XuXShirley

@XuXShirley I used to calculate only one picture and sifid of each fake picture it generated, and the results were particularly small, as if only the sifid between different real pictures and the fake pictures they generated was calculated according to the author To get the desired result

15732031137 avatar Mar 08 '20 02:03 15732031137

@15732031137 @tamarott I calculate the sifid of one picture and the fake picture generated, when the image suffix is "png", the result is particularly small, like 2.9502307e-06. Then I convert them to "jpg", the result turns out to be 0.038818467.It's weird.

XuXShirley avatar Mar 08 '20 06:03 XuXShirley

@XuXShirley Wow! This finding is very interesting. I started with png and got very small results, but I changed the program to the default output format as jpg and see what happens! Where do you come from?

15732031137 avatar Mar 08 '20 08:03 15732031137

@15732031137 I come from China.

XuXShirley avatar Mar 08 '20 10:03 XuXShirley

@XuXShirley 哈哈 我也认为我们都是中国人 我微信是15776557066 有空可以加一下互相交流吗

15732031137 avatar Mar 08 '20 12:03 15732031137

Thank you for noticing this. The format should be jpg. Code and readme file were updated.

tamarott avatar Mar 11 '20 09:03 tamarott