Inv marker contour detection improved

[szk1509]

This pull request improves the accuracy of the white markers

Resolves #2055.

• An accuracy improvement of the white marker corners was achieved.
• A slight improvement for the precision of the black markers corners was achieved.
• The achieved accuracy remains stable along different noises.

This pull request modifies the contour selection for the white marker contours, by modifying the selection process of the corners and contour candidates, in _filterTooCloseCandidates.

candidatesSet: The candidates and the contours are saved in a structur, for future improvements.

[alalek]

Please squash patch into one commit (remove unrelated commits from the patch and remove merge commits (use rebase instead of merge if necessary)).
After that rebase patch onto 3.4 branch: https://github.com/opencv/opencv/wiki/Branches

[alalek]

I rebased PR changes onto 3.4 branch (please fetch commit from your fork, do not use merges, cherry-pick new commits if any).

[szk1509]

Thank you @alalek
🤔 fetch commits, use cherry pick and do not merge (use rebase instead of merge) are instructions for future PR? or do i have to fetch, then commit and squash the commits if so?
Sorry!! i am a bit (too) confused 😭

and again, thank you for your help 😃

[alalek]

Thank you for the contribution!

[alalek]

vector< vector< vector<

There are really two many nested vectors.
Two values are expected in outer dimension.

Perhaps we should split this parameter into two:

vector< vector< Point2f > > &candidatesOut,
vector< vector< Point2f > > &candidatesInvertedOut,

[alalek]

if we observing lack of performance in this code, then we can try to utilize smart pointers:

-const vector< vector< Point2f > > &candidatesIn
+typedef vector< Point2f > CandidatePoints;
+typedef cv::Ptr<CandidatePoints> CandidatePointsPtr;
+const vector<CandidatePointsPtr> &candidatesIn

[alalek]

Looks good to me! Thank you 👍

[hamddan4]

I have a question regarding how it works, posted here: https://stackoverflow.com/questions/58302536/how-to-properly-detect-inverted-white-markers-with-opencv-aruco-module

[szk1509]

Hi @hamddan4,
The good news is that it works for me, so i think i know where the problem is :)
the bad news is that it comes soon on the next PR...

if you like i can give you some more details, so that you can keep working ;)

[hamddan4]

Hi! Yes, i would like to know how to add this functionality to my code. Afaik, detectInvertedMarkers option of parameters is on the source code, but it doesn't seem to work.. what I am missing here?

[szk1509]

@hamddan4 🤔 well ...
The good news is that after trying 3.4 ad 4.1, it worked for me, so it works... my bad, I was in a hurry and couldn't check it properly.
The bad news is that I can't reproduce your error 🤔 so let's see...

import cv2
from cv2 import aruco
dic = aruco.getPredefinedDictionary(aruco.DICT_APRILTAG_16h5)
par = aruco.DetectorParameters_create()
par.detectInvertedMarker = True
image = cv2.imread("error/MNSGxZT.png")
dMarkers = aruco.detectMarkers( image, dic, parameters=par )
cv2.imshow("d", aruco.drawDetectedMarkers(image, dMarkers[0]))
cv2.waitKey(0)

i don't actually see what are you missing 🤔

[hamddan4]

So if it works with you, it has to be some problem with the my installed libraries? Maybe the headers are good but is not implemented in the installed version of opencv. I've done this by pip install opencv-contrib-python. Don't know what's wrong with it though. Is there anything to test it? opencv.__version__ yields 4.1.1.

[szk1509]

@hamddan4 well, i use (and recommend you) the official git repo... it includes some tests and tutorials :)
https://docs.opencv.org/