Disparity in the numbers of features detected with SIFT in comparison with OpenCV SIFT implementation

[Bikraman Karmakar]

Hi,

First of all thank you for building BoofCV. I am currently building a feature in our application where I have to find if a person is taking the photographs of the same object multiple times in order to reduce spoofing. I have already achieved this using OpenCV in android. The only issue is that using OpenCV is really increasing the size of APK by a lot of MBs. We want to try to keep the size low. 

That is what brought me to BoofCV. The issue now is that BoofCV is not giving the same similarity scores using SIFT. It is outputting way less of keypoints as compared to OpenCV. 

If OpenCV SIFT implementation finds 9893 and 5675 keypoints for images A and B respectively, BoofCV finds only 3790 and 3729 keypoints for the same images. 

ConfigSiftDetector conf = new ConfigSiftDetector();
conf.extract.radius = 3;
conf.extract.threshold = 2;
conf.maxFeaturesPerScale = 120;
return conf;

The above is my SIFT Config.

Can someone kindly help me out. I would really appreciate it. 

Best,

Bikraman

[Bikraman Karmakar]

Okay turns out it was my dumb mistake, there is a homography estimation code which was affecting the final score. BoofCV SIFT is pretty consistent with OpenCV scores actually. My bad.