When comparing the CbCr compressed image and the original image, I was hardly able to tell the difference
between the two; between the Y' compressed image and the original, the difference was quite apparent. The fact that significant compression can be achieved without significant loss of image quality is cool, since until now I never understood how compressed images could maintain their overall structure/color so well.
ethanweber
Are there many issues of training computer vision models (for classification, etc.) on JPEG images? I've heard it's best to use PNGs but I'm curious if the effects of JPEG compression have been quantified in image processing pipelines other than our perceptual interpretation.
rubywerman
@ethan it depends on the dataset. This thread talks about it: https://stats.stackexchange.com/questions/440144/which-image-format-is-better-for-machine-learning-png-jpg-or-other
For video streaming, JPEG is fine, for sharper lines, PNG
somaniarushi
I went back and forth between this image and the original — crazy that I can't pick up a single difference! Wondering if anyone else can. Very cool that these file types use human biology to be more efficient!
shreyaskompalli
I wonder what evolutionary advantage this behavior of the human eye provided to humans. What I mean is, why do our eyes choose to prioritize noticing differences in luminance over chromaticity? I'd imagine that it maybe has something to do with seeing in the night/not being caught unaware in dimly lit situations, but I'd love to hear any concrete theories about this.
When comparing the CbCr compressed image and the original image, I was hardly able to tell the difference between the two; between the Y' compressed image and the original, the difference was quite apparent. The fact that significant compression can be achieved without significant loss of image quality is cool, since until now I never understood how compressed images could maintain their overall structure/color so well.
Are there many issues of training computer vision models (for classification, etc.) on JPEG images? I've heard it's best to use PNGs but I'm curious if the effects of JPEG compression have been quantified in image processing pipelines other than our perceptual interpretation.
@ethan it depends on the dataset. This thread talks about it: https://stats.stackexchange.com/questions/440144/which-image-format-is-better-for-machine-learning-png-jpg-or-other For video streaming, JPEG is fine, for sharper lines, PNG
I went back and forth between this image and the original — crazy that I can't pick up a single difference! Wondering if anyone else can. Very cool that these file types use human biology to be more efficient!
I wonder what evolutionary advantage this behavior of the human eye provided to humans. What I mean is, why do our eyes choose to prioritize noticing differences in luminance over chromaticity? I'd imagine that it maybe has something to do with seeing in the night/not being caught unaware in dimly lit situations, but I'd love to hear any concrete theories about this.