sikuli-driver team mailing list archive

[RaiMan <question224151@xxxxxxxxxxxxxxxxxxxxx>]

Question #224151 on Sikuli changed:
https://answers.launchpad.net/sikuli/+question/224151

    Status: Open => Answered

RaiMan proposed the following answer:
current Sikuli uses a 3 level pyramid matching algorithm.

This was a big step from version rc2 to rc3 (average 3 to 5 times faster
and optimal images (little uniform background, concentration on key
visual aspects) yield scores above 0,95)).

So if we are talking about matchTemplate with some pyrDown/pyrUp, then
we have it already.

The problem with matchTemplate is not really the size of images, but the difference in size between image and probe:
having an image of 1000x1000 (1M pixels) and the probe same size, matchTemplate has to compare 1M pixels once.
But if you have the same image, but a probe of 800x800, the matchTemplate has to compare 640K pixels 360K times.

So with the current implementation of Sikuli the best optimization for both speed and accuracy is:
- minimal possible search regions
- optimal probe images (little background, concentration on key aspects)

There are surely some optimization possibilities (handling of the
screenshot down to the VisionProxy, check images in their last match
regions, ...), but these mostly are independant of image size  and in
the range of some ten to few hundred milliseconds.

Regarding accuracy, I think we are already there for optimal images.

But feel free, to have a look into the native sources.

-- 
You received this question notification because you are a member of
Sikuli Drivers, which is an answer contact for Sikuli.