3 ms·

Local tone mapping works just wine without object recognition. It's a dumb algorithm that goes over the image and normalizes pixel values relative to their surr

Local tone mapping works just wine without object recognition. It's a dumb algorithm that goes over the image and normalizes pixel values relative to their surroundings to cram the higher dynamic range of the image into the lower dynamic range of the screen. Phone camera apps had it for at least 10 years, and people did it manually with exposure bracketing ever since digital cameras became mainstream.

astrange 4y ago
Yes, but any local area algorithm benefits from segmentation because it automatically becomes "smarter". Upscaling is the most obvious one - the upscaled version of a blue object on a red background is not the average of the two colors - but it applies all over the place. I don't know of any implementations of this, but an interesting one would be auto white balance. It's typically done as a global slider, but if the image has multiple light sources this doesn't always look good. And actual straight-up "knowing what's in the scene" AI can help too; people shouldn't look sickly in your photos just because they're under a low-CRI yellow light. You probably want to know what color skin tones actually are.
- grishka 4y ago
  > Upscaling is the most obvious one I don't want any upscaling at all in my photos. One pixel on the sensor must correspond to one pixel in the output.
  - astrange 4y ago
    Camera sensors don't have pixels: https://en.wikipedia.org/wiki/Bayer_filter https://en.wikipedia.org/wiki/Bayer_filter JPEG files also don't have pixels: https://en.wikipedia.org/wiki/Chroma_subsampling https://en.wikipedia.org/wiki/Chroma_subsampling Then again, OLED displays don't either: https://en.wikipedia.org/wiki/PenTile_matrix_family https://en.wikipedia.org/wiki/PenTile_matrix_family
    - grishka 4y ago
      I'm aware of all these things but there still is a 1:1 correspondence between camera sensor pixels and JPEG [luminance] pixels. Yes, color information isn't complete, but it's the right balance. Enlarge it and you aren't adding any new information. Shrink it and you're losing information.
      - astrange 4y ago
        Definitely not true unfortunately. In particular it's not true for areas colored red or blue, because the green channel doesn't have the luminance information - that's why I mentioned them. Lightroom and at least one current phone camera have ML Bayer demosaicing for this reason and it's visibly sharper. (Note sharpening and enlarging are the same operation.)

astrange 4y ago

Yes, but any local area algorithm benefits from segmentation because it automatically becomes "smarter". Upscaling is the most obvious one - the upscaled version of a blue object on a red background is not the average of the two colors - but it applies all over the place. I don't know of any implementations of this, but an interesting one would be auto white balance. It's typically done as a global slider, but if the image has multiple light sources this doesn't always look good. And actual straight-up "knowing what's in the scene" AI can help too; people shouldn't look sickly in your photos just because they're under a low-CRI yellow light. You probably want to know what color skin tones actually are.

grishka 4y ago
> Upscaling is the most obvious one I don't want any upscaling at all in my photos. One pixel on the sensor must correspond to one pixel in the output.
- astrange 4y ago
  Camera sensors don't have pixels: https://en.wikipedia.org/wiki/Bayer_filter https://en.wikipedia.org/wiki/Bayer_filter JPEG files also don't have pixels: https://en.wikipedia.org/wiki/Chroma_subsampling https://en.wikipedia.org/wiki/Chroma_subsampling Then again, OLED displays don't either: https://en.wikipedia.org/wiki/PenTile_matrix_family https://en.wikipedia.org/wiki/PenTile_matrix_family
  - grishka 4y ago
    I'm aware of all these things but there still is a 1:1 correspondence between camera sensor pixels and JPEG [luminance] pixels. Yes, color information isn't complete, but it's the right balance. Enlarge it and you aren't adding any new information. Shrink it and you're losing information.
    - astrange 4y ago
      Definitely not true unfortunately. In particular it's not true for areas colored red or blue, because the green channel doesn't have the luminance information - that's why I mentioned them. Lightroom and at least one current phone camera have ML Bayer demosaicing for this reason and it's visibly sharper. (Note sharpening and enlarging are the same operation.)