Are the images truly `png` images?

I see the typical jpg checkerboard in many of the png images. Is this intentional?

If the images were first saved in jpg and later converted to png that would only result in a higher file size.

F.e. 00a2e221a62f5c98c116f977ee4a447d.png in 0.tar

@ironbar How are you viewing these? That isn’t at all what the file looks like when I download 0.tar fresh and open up images/00a2e221a62f5c98c116f977ee4a447d.png.

Can you make sure your files haven’t been changed from the ones you downloaded? Here’s what I see when I hash that particular image file:

❯ md5sum images/00a2e221a62f5c98c116f977ee4a447d.png 
fb362578bd7662e8bcea250a838c8c4b  images/00a2e221a62f5c98c116f977ee4a447d.png

Do you get the same result?

Yes, I get the same md5sum result. The fille is correct.

I’m using ubuntu default image viewer. But if I open the image with GIMP and zoom to the center of the image I see the same checkerboard pattern.

Have you made zoom to the center of the image?

Ah, I didn’t realize from the initial message that your screencap was a close zoom.

To answer your question, these artifacts are present in the raw source images and were not converted from .jpg, but I can reach out to the subject matter experts to see if they have anything to add.

Hi @ironbar - the SME at NASA already got back to me so I can provide a quick update. Several of the backgrounds used to generate these images may have residual artifacts, but the vast majority should not.

Generally speaking, you can ignore this, as there is no benefit to overfitting the training data, and different (but broadly similar) subsets of background images were used for the private test data.

1 Like