Vit's profile - activity

I use OpenCV3-beta in Eclipse under Win8.1 and have to build source from git since there are no minGW libs available for download. It was ok before, but today(just now) I updated all sources and now mingw32-make suddenly fails:

\git\opencv\modules\highgui\src\window_w32.cpp:2011:9: error: 'TBBUTTONINFO' was not declared in this scope

It is not the only error, there are some another errors which is similar to mentioned above:

\git\opencv\modules\highgui\src\window_w32.cpp:2023:81: error: 'BTNS_AUTOSIZE' was not declared in this scope

\git\opencv\modules\highgui\src\window_w32.cpp:2023:97: error: 'BTNS_BUTTON' was not declared in this scope

"You need permission. Ask the owner for access, or switch to an account with permission."

You can save 32 bits as CV_8UC4 image, but it is a little tricky. Use FileStorage , it is simple and can handle even large images.

If you want "just segmented image" then 4 colors are enough for any segmentation :)

OpenCV contains a thousands of functions. What do you mean by "OpenCV performance"? Do you know OpenCV and ITK functionality intersection to compare only common parts? ITK is for medical imaging, OpenCV is for near-realtime video and image processing of broad purpose -- but what is your task?

img8bit and CV_32FC3 ? 8 bit image is CV_8UC1 or CV_8UC3 (afair imwrite writes 8 bit 3 channel or 8 bit 1 channel images only). You have CV_32F (one channel) so try CV_8UC1:

wshed.convertTo(img8bit, CV_8UC1, 255.0);

Sure you will lose some information, but if you want to have just visualization it is ok. If you want store and then read all information use FileStorage class.

OpenCL version of Canny contains bug more than 1 year. Try to compare OpenCL and non-OpenCL (use cv::ocl::setUseOpenCL(false) ) results.

1. @pradeep_kb returning to the starting point: are you sure not all of your cores are calculating? Morphology is very parallelable and OpenCV uses parallel_for etc for such kind of calculation by default. 2. It is not question of OpenCV, just try do reduce morphology CPU-time using methods I mentioned above. 3. May be your picture is not very noisy and small morphology kernels is enough, or you can update another part of your code to reduce noise and avoid large kernels. It is hard to say what to do exactly without your video-example, your code and its output results.

@pradeep_kb please update you question with profiling information. Based on your manual profiling one can see the most calculation-intensive part is morphology, it takes ~80% of time. The common advise for morphology is to avoid large kernels, emulation for large kernel is simplest resize and small kernel or "small kernel N-times passing".

profiling

Short answer: no, afaik there are no out of the box 3D convolution for arbitrary kernel in openCV.

Long answer: some convolution kernels are separable i.e. 2d convolution could be presented as a sequence of two 1D-convolution in one direction and then 1D in another direction (see sepFilter2D). The same for 3D -- some 3D kernels could be presented as a sequence of 1D or 2D convolution. For instance Gauss kernel is always separable. 1D convolution is much faster than 2D or 3D that's why one should always try to use separable kernels. If your particular filter is not separable you can try to approximate it as a superposition of separable filters (using SVD etc). But you always should think twice is it really so necessary to use not separable 3D filtering in your case.

Where is your program bottleneck? Did you perform profiling?

In C++ it is cvtColor(raw, bgr, CV_BayerBG2BGR) (there are a lot of CV_Bayer... values), I hope in Java it looks very similar.

Are you sure your regions are too small? By default drawMatches just draw points, not full patches. Use DRAW_RICH_KEYPOINTS flag for drawMatches(...). See docs.

As I can see from source code, you can fill options for constructor AKAZEFeatures::AKAZEFeatures(const AKAZEOptions& options) and change its descriptor_pattern_size which is set by default to 10 ( see AKAZEConfig.h )

It is size in bits (not pixels), and it is set to max by default: descriptor_size Size of the descriptor in bits. 0 -> Full size

I mean 'create' not 'use' since afaik no OpenCL AKAZE in OpenCV yet.

As I can see from brief view to the linked article the main complexity comes from matrix inversion and this matrix size is ~number of reference points. So one can assume that complexity is O(n^3). Sure there some other calculations but large matrix inversion is the most intensive. You can easily check this assumption plotting 'calculation time vs number of points' chart and fitting it by polynomial curve (or plotting it in semi-log scale and fitting by line). Anyway big-O notation is asymptotic estimation only, so for small N it can significantly deviates from 'ideal' polynomial form.

Guess the best way to reduce execution time is to create its OpenCL version (afaik there is only CPU version of algorithm). KAZE by its nature is parallel so performance could be boosted by OpenCL approach.

Usually every complicated method in OpenCV docs links to science article, where complexity always mentioned. This science articles often has open access. And sometimes complexity mentioned in .h files.

"Certain items" can be detected with certain approaches ) Show your picture and mark item on it to be very specific.

The reason could be RANSAC which uses random generator to find best matches, so it is not deterministic approach.

guess link is better 1. it shows that this question is already exists and answered 2. link has confirmation from recovered OpenID, it gives more faith )

fixed 2 days ago

If bees are always moving and there are no forever-static bees, you may slightly modify your approach and use MOG only for creating binary mask of static image parts. Then just overlap this static parts from your frames and you will have "street without moving people".

Yes, it works. My GoogleID-account (this) is now restored. Thank you.

How many photos of fish do you have in your database and how many photos do you have to process every year (?) to find it in this database? How many photos per fish? How large is acceptable level of "false positive" and "false negative" results?

Can you show what is "scale pattern" on your pictures? And also show a few possible patterns.

@axk Your images scales differs, so "some size normalization should be performed (based on lines interval, letters size etc) or brute force like build size pyramid using small steps like (1.5)^(1/8) = 1.05". Phase correlate works good for dx+dy shifting only, not for scale change.

@FooBar It is the same page photos, hence it does not matter how many rows in your descriptors -- they will mostly match each other, the bigger descriptors area the smaller false positive match rate.

I see a few possible solutions for your particular task:

• it seems that simple and fast phase correlate should be enough, see phaseCorrelate() function docs; may be some size normalization should be performed (based on lines interval, letters size etc) or brute force like build size pyramid using small steps like (1.5)^(1/8) = 1.05
• text is self-similar on the letter size scale and not self-similar on the word size scale, hence you can try to increase zone for key-point descriptor if you want to stay in salient points approach.

@Antonio, thanks for you informative report on OpenCV bug-tracker. I hope this will activate works on that issue. According to attached "cmake-gui-out.txt" file there I have 32-bit version.

Windows 8 has a lot of security restrictions by default, including for system partition root path, so you'd better try to write to you project directory "./thevideo.avi" .

I have seen on November somewhere on opencv.org about February as the month for OpenCV3.0 release. But looking to the bunch of reported bugs.. guess it will be on summer-autumn 2015.

@Antonio, thanks for your answer. I tested Canny and it is ok in 2.4 with this test picture, your code works perfect. The problem is actual for OpenCV3-beta only. Post updated by soft- and hardware info.

As for threshold, from my point of view it is more convenient to have normalized output i.e. threshold should not be a function of another parameters, this increases the code hidden-errors possibility.

Simple code for simple image result is not connected contour as it is expected. Top or bottom edge missed. Result printscreen is square-bug.png . The picture itself is square.png. Code is:

   void test(){
        UMat img = imread("./square.png", IMREAD_COLOR).getUMat(ACCESS_READ);
        UMat imgGray, imgGrayEdges3, imgGrayEdges5;
        cvtColor(img, imgGray, CV_BGR2GRAY);
        Canny(imgGray, imgGrayEdges3, 50, 150, 3);
        Canny(imgGray, imgGrayEdges5, 300, 2000, 5); // why so large thresholds?..
        cout << imgGray.getMat(ACCESS_READ) << "\n" << imgGrayEdges3.getMat(ACCESS_READ) << "\n" << imgGrayEdges5.getMat(ACCESS_READ) ;
        imshow("img", img);
        imshow("canny3", imgGrayEdges3);
        imshow("canny5", imgGrayEdges5);
        waitKey();
    }

Am I doing something wrong?

I use OpenCV3-beta in Eclipse under Win8.1 and have to build source from git. Source updated and compiled today (last canny.cpp change was 7 weeks ago). I do have openCL so I guess UMat Canny function uses OCL_RUN version of function

UPDATE: device and system info

• Driver Packaging
• Version 14.501.1003-141120a-178000C
• Catalyst Version 14.12 AMD Catalyst
• Omega Software 2D Driver
• Version 8.01.01.1443 Direct3D
• Version 9.14.10.01080 OpenGL
• Version 6.14.10.13283 Mantle Driver

• Version 9.1.

  • Graphics Chipset AMD Radeon HD 7900 Series
  • BIOS Version 015.012.000.004
  • BIOS Part Number 113-C3860100-100
  • BIOS Date 2011/12/07

Just guess (never tried this): it is worth do use ROI(1x1px), copy it to a new 1x1 oclMat and download this small matrix. Accoring to docs if you try download a part of oclMat then all matrix is transferred: "Data transfer between Mat and oclMat: ROI is a feature of OpenCV, which allow users process a sub rectangle of a matrix. When a CPU matrix which has ROI will be transfered to GPU, the whole matrix will be transfered and set ROI as CPU’s. In a word, we always transfer the whole matrix despite whether it has ROI or not." -- row(i).col(j) is a kind of ROI and this is your case.