sew some random twenty lines into faces

[ickby]

there may also be some wording confusion, sometimes it is called circle, but more often cycle basis: https://en.wikipedia.org/wiki/Cycle_basis

[teobo]

-is a circle a vertex candidate?
-found by a window, feature detection method for "corners"?

so you find the best bet vertex candidates among your findings_of_all_vertex candidates(=circles?)
Is it that what is proposed? And it does include hough lines? No?

Maybe let's say let's start with a heuristic, which is is "take the first, leave the last, it would bring prototype results yet. The original question is to be changed to: are there samples or demos among the other users and doers of the tool opencv - or the like.
-And I might address that I still can not even name the circles-to_face-gap to be understood and to be pointed elsewhere.


-As a pioneer one would collect the testimonials of the coexperimenters, right? So this rather goes to be a demo script rally among opencv. (Only here?) They would have maintained a personal, cheap recognition best of management in their former production lines. So knowing 10, one copies such a code and uses it for the production line of the present case of lets say of probability limits of 97% and/of 100% successfully recognized compound.-> OK...

Will look for demotests in the meanwhile, waiting for new incentives , too, ...

[microelly2]

My idea:

1. prepare the image in a filter chain
2. create hough lines
3. analyse hough lines: in a perspective photo only the lines to the vanishing points
(they correspond to the main directions of my target house)

4. analyse "perspectivic parallel" lines to get exact coordinates/directions

when you have only one picture you need intervention because there is a meaning of the picture
which only a human can understand - or there are methods to find windows, doors
like nowerdays there are patterns for faces/humans/cars

(the case of multiple pictures is at the moment not in my focus)

Now I can interactive build my box-models.

5. find "perspective" circles == ellipses and arcs of them.
If there is a knowledge in which known plane they live, their real dimensions
can be approximated.

So I can get my model interactive with best possible data support.
Thats all. A human can select from a set of 100 - 300 objects the right one without
extra calculation.

I have done some experiements with realy bad data to find circle arcs

bn_924.png (667.33 KiB) Viewed 970 times

https://www.youtube.com/watch?v=V9GapOYxrXk

all steps automated after fine tuning the filters by hand
and the results were acceptable

opencv filtering of the background grid
collecting the connected points to pathes,
creating FreeCAD polygons and Bsplines
and finding circle arcs with ransac (hough circle does not work in this case
because there are to many objects on the page and I had no idea how to separate them automaticly)


(this is still time consuming but I cannot code without an idea where to do the tuning,
I think people can learn and then they can automate their knowledge
I want not go to deep theoretical - because I do not have the time)