synchronize default scaling among ImageShow, ImageView, and Plots?

[JeffBezanson]

Given color components outside [0,1], currently we have the following:

• ImageShow/IJulia: clamp
• ImageView: uses minimum and maximum value as black and white
• Plots: errors

It would be nice to handle this consistently. I prefer the ImageView.jl behavior, since it generally lets you see as much data as possible. However, I will also suggest the following algorithm:

• Let lo, hi = extrema(image)
• Black = 0 <= lo <= 1 ? 0 : lo
• White = 0 <= hi <= 1 ? 1 : hi

The intuition is as follows:

• If all values are between 0 and 1, just use them directly as intensities.
• If the values are all over the place, use the min and max as black and white (what ImageView does now).
• If values are generally in [0,1] but some go outside the range in one direction, rescale around 0 or 1 to bring those values back into range.

The advantage of this is that when some values overflow or underflow a bit, 0.5 still looks "grayish" just like it would if you clamped the values, but without losing information.

[mkborregaard]

Why not just have Black = 0; White = 1 though and clamp? That is the most accurate way of depicting the underlying color information in the matrix of Grays. If it's to use this to visualize matrix values (so you need relative contrasts to be as large as possible) then possibly a heatmap is more appropriate?

[JeffBezanson]

Often a processed image has values in a different range, e.g. after convolution with certain kernels, or because it came from some data source with a strange representation. Now of course, in those cases you might eventually apply some custom scaling to get a real usable image out, but for development it's nice to be able to quickly look at the data with some reasonable default scaling.

I suppose if the Gray type means intensity from 0 to 1, then clamping is indeed the right thing, and it would be nice for all image-displaying packages to adopt that consistently. My problem is that I've been using Gray.(a) as a way to tell e.g. Plots that the data should be treated as an image. Do we have, or should we have, a way to label values as image-like without specifying a scale or color space? E.g. ScalarIntensity, or UnscaledGray or something like that?

[mkborregaard]

I agree completely on the consistency. But not sure I understand what you mean by telling Plots you want the data treated like an image. Isn't that what heatmap is for?

using Plots
a = 0.6rand(200,200).+0.2
plot(
          plot(Gray.(a)),
          heatmap(a, c = :grays),
          heatmap(a, c = :grays, clim = (0,1), legend = false),
          layout = (1,3), size = (800, 250)
          )

IIUC heatmap(a, c = :grays) does exactly what you'd want?

[JeffBezanson]

Ok, I agree. We should consistently render Gray by clamping to [0,1], and at least when using Plots I'll just use a heatmap.

[mkborregaard]

Great, I'll implement the clamp.