add dynamic window in aruco cornerRefinement #24355
Conversation
2438a78 to
6b1e3b0
Compare
| * In the case of markers located far from each other, it may be useful to increase the value of the parameter to 0.4-0.5. | ||
| * In the case of markers located close to each other, it may be useful to decrease the parameter value to 0.1-0.2. | ||
| */ | ||
| CV_PROP_RW float minCornerRefinementWinSizeInPin; |
There was a problem hiding this comment.
I propose to name the field as relativeCornerRefinmentWinSize. And add information about avarage Aruco pin size and logic to the description.
There was a problem hiding this comment.
renamed to relativeCornerRefinmentWinSize:
/** @brief minimum allowed window size for the corner refinement process (in pins of ArUco marker) (default 0.3).
*
* cornerRefinementWinSize is recalculated for each marker as:
* cornerRefinementWinSize = min(cornerRefinementWinSize, averageArucoModuleSize*relativeCornerRefinmentWinSize),
* where averageArucoModuleSize is average module size of ArUco marker in pixels.
* (ArUco marker is composed of black and white modules)
* In the case of markers located far from each other, it may be useful to increase the value of the parameter to 0.4-0.5.
* In the case of markers located close to each other, it may be useful to decrease the parameter value to 0.1-0.2.
*/
CV_PROP_RW float relativeCornerRefinmentWinSize;
6b1e3b0 to
663f4ba
Compare
663f4ba to
fd4af21
Compare
|
@AleksandrPanov what is the peak at 10px? |
|
@victor1234, all markers detected with an error of 10 pixels or more are marked as undetected. These markers make a peak of 10 pixels. The error calculate as the Chebyshev distance between the gold marker and the detected marker. |
I would say that mean error (=1px) on synthetic data for subpixel method is high. And outliers with error > 1 are also look strange. Do you know the reason? |
|
Several groups of images are blured. Also the benchmark adds color noise on top of the image like pepper-solt augmentation. Current hypothesis is related to this. We make experiments right now and analyze the detector code to prove it. |
Fixes #24113
cornerRefinementMethod=CORNER_REFINE_SUBPIX does not work correctly with markers that are too close to each other.
Now
cornerRefinementWinSizeis recalculated for each marker as:cornerRefinementWinSize = min(cornerRefinementWinSize, averageArucoPinSize*relativeCornerRefinmentWinSize)relativeCornerRefinmentWinSize=0.3 by default and could be changed
Also the method has been tested by the benchmark, 289 aruco boards like this were tested:
command line to generate and run (also need set cornerRefinementMethod: 1 and cornerRefinementWinSize: 10 in aruco_params.yml):
tested with cornerRefinementWinSize=10
error in pixels with old cornerRefinementMethod:
mean detected error in pixels 5.912010732442689
error in pixels with new cornerRefinementMethod:
mean detected error in pixels 0.991642490414958
command line to show histogram:
new.json.txt
old.json.txt
Pull Request Readiness Checklist
See details at https://github.com/opencv/opencv/wiki/How_to_contribute#making-a-good-pull-request
Patch to opencv_extra has the same branch name.