search

google / butteraugli

butteraugli estimates the psychovisual difference between two images
Apache License 2.0
1.96k stars 137 forks source link

Question about butteraugli's threshold value #24

Open NanyangYe opened 7 years ago

NanyangYe commented 7 years ago

Dear all:

There is something we are not certain about. We test butteraugli on this image. Original:

artificial

Compressed JPEG compression quality 90 artificial_90

According to butteraugli's result, the distmap's value is above good threshold and even bad threshold, so the difference can be visible. But according to our experiment, human cannot observe the difference at all. And the butteraugli' results keep the same even when the quality reached 98. Is there any problems with our understanding. Many thanks.

jyrkialakuijala commented 7 years ago

How do you conduct the human rating experiment?

I see some difference in the darker areas where the pattern disappears into the black. There is higher saturation and the pattern disappears more quickly into the dark in the lower images. Also, the lower image has some ringing in the background pattern.

NanyangYe commented 7 years ago

@jyrkialakuijala The images were compared in a side-by-side experiment on a calibrated sRGB display with the peak luminance of 220 cd/m^2, dark room. The viewing distance was controlled so that the angular resolution was 60 pixels per visual degree.

mqudsi commented 2 years ago

The issue with the samples you posted isn't jpeg compression artifacts; there's a very noticeable difference in the gamma or the color range.