I am looking for a patent-free local image features for general image classification. I found ORB features implemented in OpenCV, but not sure about their performance. What would you suggest?
I think local gradient histogram performs very nice in both accuracy and robustness. If realtime performance is required, I prefer local binary pattern for fast speed. Of course, there are not only one type for each. Variants can be found in many papers. But, I don't know whether they are patent free or not.
You can also use other kind of local binary features, such as LDB or M-LDB. These descriptors are more recent than other similar approaches such as BRIEF, BRISK, ORB or FREAK. Furthermore, LDB has demonstrated that it can be more effective than the previous binary descriptors in tasks such as place recognition or CBIR. I attach you some very recent papers where we have employed these descriptors and where you can see its effectiveness with respect to other approaches.
Good luck with your research!
Conference Paper Fast and Effective Visual Place Recognition using Binary Cod...
Conference Paper Bidirectional Loop Closure Detection on Panoramas for Visual...