Is there an example on how to use mosaicing?

[Nakilon]

https://libvips.github.io/libvips/API/current/libvips-mosaicing.html

I have a thousand of images that all have different amount of matching so I probably can't come up with universal numeric arguments.for vips_match. Also I can't tell I understand them at all. When I had a look on docs previously I believe I saw a function that produces a 2d bw map of level of matching between two images so I could then decide what do to with it. But I probably can't find what function that was and the...

irb(main):005:0> im11.match(im12)
Vips::Error: unable to call match: you supplied 2 arguments, but operation needs 10.

just tells me I need something different.

UPD: something like this: https://stackoverflow.com/q/36688439/322020
I believe I saw such function...

Or maybe that was match but I imagines it to work simplier when I was reading docs the last time.

UPD2: Or I guess it was spcor. After scaleimage (btw, is it per band?) gives me such image:

but takes eternity to correlate two 1280x720 images while I have a thousand of them...

UPD3: oh, and I just realised that it's only trying one quarter of relative shifting so I need to do it 4 times, right?

rd = img11res.fastcor(img12res)
ru = img11res.flipver.fastcor(img12res.flipver).flipver
ld = img11res.fliphor.fastcor(img12res.fliphor).fliphor
lu = img11res.fliphor.flipver.fastcor(img12res.fliphor.flipver).fliphor.flipver
Vips::Image.arrayjoin([lu,ru,ld,rd], across:2).scaleimage.write_to_file "temp.png"

does not look right...

Maybe I don't understand how it's supposed to work. It makes no sense to me.

a = Vips::Image.new_from_array [[0,0,0,0,0],[0,1,0,0,0],[0,0,0,0,0],[0,0,0,0,0]]

b = Vips::Image.new_from_array [[0,0,0,0,0],[0,0,0,0,0],[0,0,0,1,0],[0,0,0,0,0]]

Vips::Image.arrayjoin([lu,ru,ld,rd], across:2).scaleimage.write_to_file "temp.png"

[jcupitt]

Yes, spcor is very slow -- it computes the full normalized spatial correlation, so 1.0 is perfect correlation and 0.0 is perfect mismatch.

There's vips_fastcor() -- it just computes sum of squares of differences, so you get 0 for a perfect match and some huge number for a perfect mismatch.

[Nakilon]

Yep, I tried both. But I don't understand why those 4 results don't assemble smoothly. And absolutely don't understand the last example with two 5x4 images that should match when shifted by (2,1).

[Nakilon]

I guess I tried the wrong function. phasecor seems to be what I need:

 img12res.phasecor(img11res).stats.to_a.drop(1).transpose[8,2]
=> [[[0.0], [0.0], [0.0]], [[23.0], [23.0], [23.0]]]

These as I understand are the shifting candidates per band.

What I don't yet understand is why it's #<Image 160x90 double, 3 bands, b-w> -- being b-w but with 3 bands makes the .write_to_file save both dark and light pixels as white. How do I switch the interpretation? .copy(interpretation: :srgb)

P.S.: and it's weird that when I call it on two similar (i.e. no shift needed) or even identical images it's all a noise and no bright pixel in the left-top corner.
P.P.S: also there is a true point and three false points and they are I don't know how to chose one because they all are of the same intensity. For example, here the true one is the left bottom:

[Nakilon]

So I decided to just chose the one of 4 shifts by the lowest (rect1-rect2).abs.avg.
Then I used embed and bandrank and got this:

The new problem is that taking "median" among several images using bandrank may take the black color while I would like to discard it. Like if here https://www.rubydoc.info/gems/ruby-vips/Vips/BlendMode was such blendmode as "discard black and pick the one closest to the average". Otherwise the median results in this for now:

Or if I chose the index:images.size-1 it results in a mess:

[Nakilon]

To illustrate the result of "discard black and pick the one closest to the average", here is a rough implementation on Ruby level (that is obviously ~100 times slower i.e. unusable):