SURF feautures doesn't work with DenseGrid when grid sampling is pair

Transferred from http://code.opencv.org/issues/2031

|| Alexandre Benoit on 2012-06-08 11:29
|| Priority: Normal
|| Affected: None
|| Category: nonfree
|| Tracker: Bug
|| Difficulty: None
|| PR: None
|| Platform: None / None

SURF feautures doesn't work with DenseGrid when grid sampling is pair

Hi everyone,
I observe that on the current svn trunk and OpenCV2.4 release, the SURF descriptor returns only zeros when used with a Dense grid configured with a pair pixel sampling.
The problem seems to not appear when using SIFT

Code samples are below

History

Alexandre Benoit on 2012-06-08 11:45

Hi again, oops, my explanation was a bit wrong at the end, sorry for that, here are the good code samples :

I observe that on the current svn trunk and OpenCV2.4 release, the SURF descriptor returns only zeros when used with a Dense grid configured with a pair pixel sampling.
The problem seems to not appear when using SIFT

Typically when the dense grid is such like this :
<pre>
cv::Ptr<cv::FeatureDetector> _featureDetector = new cv::DenseFeatureDetector(1.f, 1, 2.5f, 6);)// descritor returns zeros when sampling is 6,8,10,12, however, it "works" with sampling like 5,7,9, etc.
</pre>

and the descriptor being like this :
<pre>
_descExtractor = cv::DescriptorExtractor::create("SURF" ); or "OppenentSURF"
</pre>

Then, running the descriptor like this :
<pre>
_featureDetector->detect(colorImage, imageKeypoints, mask);
_descExtractor->compute( colorImage, imageKeypoints, currentImageDescriptors);
</pre>
==> here currentImageDescriptors is the right length but it contains only zeros is dense features sampling is a mutiple of 2

Kirill Kornyakov on 2012-06-15 06:47

Alexandre, thank you for report, but please use wiki "syntax":http://code.opencv.org/help/wiki_syntax.html, and do not use High priority by default. All bugs are triaged, and the team will decide about the priority and Target version.
-   Target version deleted (3.0)
-   Priority changed from High to Normal
-   Description changed from Hi everyone, I observe that on the current
    svn trunk and OpenCV2.4 release, t... to Hi everyone, I observe that
    on the current svn trunk and OpenCV2.4 release, t... More
-   Category changed from features2d to nonfree

Start by reproducing the issue with cv::DenseFeatureDetector and the SURF or OpponentSURF descriptor using the detect and compute calls shown in the report. Compare descriptor output for even sampling values such as 6, 8, 10, and 12 against odd values, and consider the issue done when dense SURF descriptors are no longer all zeros for valid keypoints.