Re: fine-tune the control points from command line

[JohnPW]

I don't know, but I assume you have been here?

http://wiki.panotools.org/Panorama_scripting_in_a_nutshell

I see no command listed for it, but this doesn't look definitive.

I'm sure a knowledgable person around here can give you a definite answer.

John

On Monday, October 8, 2012 6:22:43 PM UTC-5, Willeke wrote:

Hi,

Is it possible to fine-tune the control points from the command line?

Willeke

[TB]

Close from 2 years after previous post...

I have been to the link http://wiki.panotools.org/Panorama_scripting_in_a_nutshell
and confirm that it is still not a command (as of today, September of 2014).

As per what I saw from Hugin source code, the "Fine-tune all Points" menu is included in Hugin itself and is not a command called by Hugin...

I also tried some Python scripts but I did not find any way to call from command line...

if someone sees another way, please respond here so we would be updated...

Thanks in advance!

[Terry Duell]

On Mon, 15 Sep 2014 10:21:48 +1000, TB <tbaeck...@gmail.com> wrote:

> Close from 2 years after previous post...
> I have been to the
> link http://wiki.panotools.org/Panorama_scripting_in_a_nutshell
> and confirm that it is still not a command (as of today, September of
> 2014).
> As per what I saw from Hugin source code, the "Fine-tune all Points" menu
> is included in Hugin itself and is not a command called by Hugin...
> I also tried some Python scripts but I did not find any way to call from
> command line...
>
> if someone sees another way, please respond here so we would be
> updated...
>

Am I right to assume that you are finding your control points via command
line, using cpfind and wish to have a command line method of fine tuning
those control points?
In case you haven't looked in detail at the cpfind options, you can use
--fullscale which may provide "better" control points.
There are also a number of parameters for fine-tuning, such as
--sieve1width, --seive1height etc which may help.
In addition, the command line tool cpclean has options to remove any
generated control points with errors bigger than a specified value.
Does any of this help?

Cheers,
--
Regards,
Terry Duell

[T. Modes]

Am Montag, 15. September 2014 02:21:48 UTC+2 schrieb TB:

Close from 2 years after previous post...

I have been to the link http://wiki.panotools.org/Panorama_scripting_in_a_nutshell
and confirm that it is still not a command (as of today, September of 2014).

As per what I saw from Hugin source code, the "Fine-tune all Points" menu is included in Hugin itself and is not a command called by Hugin...

 I don't see the point of fine-tune cp from the command line. Cpfind and align_image_stack set the cp already with sub-pixel accuracy. So fine-tune is not helpful.
Fine-tune has its application when setting cp by hand and then fine-tune the rough position.

[TB]

Hello / Hallo

> Am I right to assume that you are finding your control points
> via command line, using cpfind and wish to have a command
> line method of fine tuning those control points?

Yes, you're right, my exact command is:
cpfind --fullscale --celeste --multirow -o ptoFileName.pto ptoFileName.pto

And my idea was to drop CPoints that had a questionable correlation value.

> In case you haven't looked in detail at the cpfind options, you
> can use  --fullscale which may provide "better" control points.

Totally right. I confirm: It takes few more seconds (minutes with older CPU and less memory) but gives better results. I was using this one also.

> There are also a number of parameters for fine-tuning, such
> as --sieve1width, --seive1height etc which may help.

Good idea. I left those options aside, but I wonder if I would get more trustful CPoints by using smaller --sieve1* values...

> In addition, the command line tool cpclean has options
> to remove any generated control points with errors bigger
> than a specified value.

I use it also. But I understood that it removes points based on points "too far" from their position. Ideally, I would like to use it as a "second pass". The "first pass" would the correlation.

> Does any of this help?

Yes, always. By confirming my thoughts and giving some ideas... (sievel*)

> I don't see the point of fine-tune cp from the command line.
> Cpfind and align_image_stack set the cp already with sub-pixel
> accuracy. So fine-tune is not helpful.

I don't know how cpfind works in details but I always got CPoints with correlation below 0.5 (generally in some "blurry" areas, in the middle of a bigger pattern). I believe that this might be because cpfind works on bigger areas to find similarities while fine-tune option of hugin works on smaller search area.
My idea would be to drop the points that cannot be aligned at pixel-level (with high-contrast details in a small region).

(Regarding align_image_stack, I did not use it as I am focusing on panorama now, but useful to know anyway).

> Fine-tune has its application when setting cp
> by hand and then fine-tune the rough position.

Partially agree:
Fine-tune is extremely useful and saves a huge time when adding manual points.
But
It is also extremely useful to automatically estimate how "trustful" are the control points that cpfind created automatically. Especially with the "Edit > Fine-tune all Points" option.
When proceeding manually, right after having cpfind created the points, I "Fine-tune all Points" and then delete all the CPoints that are below 0.6 "distance" (that is correlation value at this point). I prefer creating my own extra CPoints if needed.

----------

Some more details on my target: I often take panoramas and people request to have those quickly on the web. So my aim here is to automate the creation of panorama from a set of pictures.
Sunday afternoon, I created the Python script attached to this message. That does most of the work.
The same is also there:
http://t.baeckeroot.free.fr/createPanorama.py
(I'll update on my site)
(feel free to adapt and/or revert to me and/or re-use it in any way you want. I build this for me: Python 2.7, Linux, Hugin 2011.4)

In brief my idea was to:
A- load the pictures composing a panorama
B- create Control-Points (cpfind)
C- Fine-tune them and drop CPoints that had correlation < 0.6
D- optimise alignment (autooptimiser and pano_modify)
E- drop CPoints that were statistically "too far" (cpclean)
F- create "final draft image" (nona and enblend)

Result is not perfect but enough for web. And fully automated.

The only missing part is C- to get the correlation value of the CPoints (to drop the less trustful).
If I can "Fine-tune all Points" and get the correlation value (in place of distance in hugin), then I believe I would be able to delete the CPoints from Python.

In my script, this is the part:
"
######## ( File > Preferences > Control Points Editor > Correlation Threshold: lower to 0.4 ) ########
######## Edit > Fine-tune all Points ########

######## Drop the CPoints that have correlation (distance) < 0.6 ########
"



Thanks Terry and Thomas

[T. Modes]

Hello,

Am Dienstag, 16. September 2014 02:47:31 UTC+2 schrieb TB:

I don't know how cpfind works in details but I always got CPoints with correlation below 0.5 (generally in some "blurry" areas, in the middle of a bigger pattern). I believe that this might be because cpfind works on bigger areas to find similarities while fine-tune option of hugin works on smaller search area.

The correlation is not the only measure for the "quality" of a cp. Cpfind is using a different "quality" criterion (and also on other scales). So this is not per se better or more worse than the correlation.
So I would not use the correlation as the only measure for the quality of a cp.

 

In brief my idea was to:
A- load the pictures composing a panorama
B- create Control-Points (cpfind)
C- Fine-tune them and drop CPoints that had correlation < 0.6
D- optimise alignment (autooptimiser and pano_modify)
E- drop CPoints that were statistically "too far" (cpclean)
F- create "final draft image" (nona and enblend)

D and E are in the wrong order.

Without the correlation criterion, where I don't agree, this can already done by PTBatcherGUI (queue assistant projects, then automatically stitch).