Ask Your Question

Configure FeatureDetector.ORB using Java OpenCV 3.3.0

asked 2017-08-31 10:08:57 -0500

I would like to configure parameters of the ORB Feature Detector using OpenCV Java version 3.3.0. I saw an earlier post for an older version of OpenCV that the C++ java bindings might be broken for ORB in so much that it isn't possible to configure parameters when using java. Unless if I have a bug in my code, this appears to still be the case with the latest 3.3.0 version of OpenCV. Am I doing something wrong, or is there a viable workaround? Esstentially, the primary parameter that I want to configure is nFeatures. NOTE: featureDetector.write() produces an empty YAML file both before and after trying to the load modified parameters with;

edit retag flag offensive close merge delete

1 answer

Sort by ยป oldest newest most voted

answered 2017-08-31 10:28:37 -0500

berak gravatar image

no, sad as it is, nothing changed. you cannot pass params to the create() function, and ORB does not serialize anything.

edit flag offensive delete link more


That is, indeed, sad. I might take a look at the source and see if I can contribute a patch. Since I would like to increate nFeatures, I am pondering the following workaround: "gridding" my image by running featureDetector.detect(image, keypoints) and descriptorExtractor.compute(image, keypoints, descriptors) in each quadrant independently, then post-process the resulting "keypoints" and "descriptors" (for example, in the keypoints, adjust and to be translated from the quadrant coordinates to the full image coordinates; not sure if anything needs to be adjusted in the descriptors) combining all keypoints into a single MatOfKeyPoint and descriptors into a single Mat before calling matcher.knnMatch(descriptors1, descriptors2, knnMatches, 2); Would this work?

catchwoodynow gravatar imagecatchwoodynow ( 2017-08-31 19:38:43 -0500 )edit

if you want my 2 ct here: the whole features2d thing in java is still based on the opencv2.4 approach, where it needed manual wrapping cv::Ptr, it seriously needs a general overhaul. (already the division into featureDetector and descriptorExtractor classes is artificial(only in the java legacy), it's the same thing in c++, and you would rather use detectAndCompute(), for not wasting computation effort)

berak gravatar imageberak ( 2017-08-31 19:51:28 -0500 )edit

Thank you for the rapid responses. I am pretty green when it comes to OpenCV, so it is taking me a bit to parse your last reply... I gather that there is a desperate need for an overhaul in the implementation as it pertains to this issue; however, I can't tell if your comment regarding "detectAndCompute()" is directed at the efforts for the general overhaul, or if you are suggesting that I could somehow utilize "detectAndCompute()" from Java today to better solve my issue. If it wasn't clear, my immediate issue is that I have some images with "features" that I am interested in, but since the image has so many high frequency components and nFeatures is fixed at 500, I am not getting any features (keypoints/descriptors) in some regions of the image that I am interested in.

catchwoodynow gravatar imagecatchwoodynow ( 2017-08-31 20:13:06 -0500 )edit

aw, sorry, that was more a rant from a c++ perspective, not at all a constructive approach at your current problem ;(

what about using like AKAZE instead of ORB ? should give far more keypoints with the default settings (yea, you pay for that with computation time..)

berak gravatar imageberak ( 2017-08-31 20:18:09 -0500 )edit

I choose ORB at the time because it appeared to be a good/"free" alternative to SIFT...I need an algorithm that is somewhat scale invariant, and for the most part ORB is doing quite well. I will read up on AKAZE -- I don't imagine that it would take a lot to make the change from ORB to AKAZE in my code (however, the project is pretty far down a path to be making such a fundamental change at this point...but...if it is the right thing to do). Perhaps some prototyping is in order. Thanks again for your thoughtful responses.

catchwoodynow gravatar imagecatchwoodynow ( 2017-08-31 20:28:50 -0500 )edit
Login/Signup to Answer

Question Tools

1 follower


Asked: 2017-08-31 10:08:57 -0500

Seen: 685 times

Last updated: Aug 31 '17