Automatically determine image/tensor color mapping & need for sRGB decoding

[Wumpf]

What

Fixes #2274

• Image viewer normalizes (?) the colors and popup shows unnormalized RGB #2274

Will need to be further worked on via:

• Allow to specify a tensor's data range via component #2341

Introduces 3 new things actually:

• finite range, determined if the range of a tensor was in fact not finite
• heuristic for color mapping range
• heuristic for srgb
  • ⚠️ does this break things?

Range & sRGB:

This

import rerun as rr
import numpy as np
import matplotlib.pyplot as plt

rr.init("image color", spawn=True)
img = np.random.random((300, 300, 3)) * 0.2
rr.log_image("image", img)
print(img.min(), img.max())  # to check they are in range [0,1]

plt.imshow(img)
plt.show()

Gives now the expected image:

When setting a pixel to inf, we show the "finite range" as well, and behave correctly:

Checklist

• I have read and agree to Contributor Guide and the Code of Conduct
• I've included a screenshot or gif (if applicable)

PR Build Summary: https://build.rerun.io/pr/2342

Docs preview: https://rerun.io/preview/a1011c1/docs
Examples preview: https://rerun.io/preview/a1011c1/examples

[Wumpf]

python ./examples/python/api_demo/main.py --demo image_tensors
before:

after:

looks the same which is encouraging

[emilk]

python ./examples/python/api_demo/main.py --demo image_tensors before:

after:

looks the same which is encouraging

I find it surprising and discouraging that the changes in this PR seems to have no effect.

On of the central parts of the heuristic in this PR is that if the color values are in the [0,255] range, we assume they are in gamma-sapce. The image_tensors test takes a u8 image and then cast it to u16, u32, f32, etc. So the range should stay the same, and the heuristic should assume we need to apply the gamma correction, no?

Look at img_rgb_uint8 - it has the gamma-curve applied, giving that darker red look that is correct. Shouldn't we get that same correct color for all the unsigned and floating point types?

[emilk]

I think the heuristics are good (except for it being RGB(A) only), but the I would expect it to have some effect on our tests. In particular, I would expect it would make all float and unsigned types in ./examples/python/api_demo/main.py --demo image_tensors have the correct gamma.

[Wumpf]

Much much better now! Needed quite a few more changes, but I'm happy with those now

[emilk]

Looks good, but please make sure you test it with depth maps and tensor view before merging!

[Wumpf]

oh. yes. haven't done that again with the latest changes!

[Wumpf]

looked all good, didn't find any regressions related to this