Question number 6000: Your hopes and dreams about OpenCV?
So, since its the 6000th question, I guess it is time to celebrate with a question where everyone could say what they like the most about OpenCV, what features are missing and should be implemented, and opinions about its future.
Congrats answers.Opencv! You've helped me a lot!
generic for_loop, generic parallel_for_loop like TBB. cv::ParallelLoopBody is too verbose. There are still a long time to wait for the c++ to support those parallel algorithms or make openMP as part of the standard of c++.
I suggest implementing more object detection algorithms, to be able to use openCV with some state-of-the-art techniques instead of using older frameworks :)
I agree. I suggest the Integral Channel Features algorithm, which is a very good upgrade to Haar and even HOG ( http://pages.ucsd.edu/~ztu/publication/dollarBMVC09ChnFtrs_0.pdf ), as a generic object detector (with excelent results in pedestrian detection). It was developed in further publications to the point it runs at 30 FPS. Its core is the Viola & Jones detection algorithm, so I guess much of t he work is done.
First of all, OpenCV is a great library. And what it makes so good is that there exist wrappers to Python, Java, C#. Furthermore that the speed is improved impressively with GPU/OCL/parallel_for_ . This effort should be continued and the missing wrappers should be added. From all the modules, imho the ml-part needs to be revised the most. It is outdated, partially awkward to setup and functionality is missing (e.g. a better predict-function that you can give a matrix of features which you would like to predict instead only one single sample). However, since opencv works in general nicely with python,scikit-learn does a great job here for me(unfortunately not the fastest library). Steven also made a good point with the object-detection-module (maybe parts of vlfeat could be incorporated?).
Great analysis @Guanta. Lets not forget the best advantage of all, which is the fact that it is open source and has a huge collaborative community that helps improving the library. I agree with that the ML module needs serious improvement. I'd also suggest an automatic design of boosted decision cascades algorithm (although this is not only applicable to computer vision), so it is possible to use the fast decision method with any features the user wants (not only haar hog and lbp).
It will add that it would be nice to have better support for network cameras and mobile video encoding. Some kind of SLAM module (I think some work is being done here), or state of the art object detection/tracking.