"BUSY BUG" on face detection---- once we initialize the Detection, 100% CPU will be used

[Harley]

Dear Yermalayeu Ihar and other experts on SIMD project，
Ｉ am trying to use SIMD for face detection as Yermalayeu Ihar metioned the SIMD makes face detection 2~3 times faster than opencv version.

But I got a "BUSY BUG" as shown in below code, once we initialize the Detection, then 100% CPU will be used, no matter whether we do face detection or not.

Can you fix this bug ? Because we want do face detection at any time and release the CPU when stop it.

Best regards.

Harley@HK

---

#include "Simd/SimdDetection.hpp"
#include "Test/TestUtils.h"
#include "windows.h"

int main()
{
typedef Simd::Detection<Simd::Allocator> Detection;
Detection::View image;
Test::Load(image, "../../data/image/lena.pgm");
Detection detection;
detection.Load("../../data/cascade/lbp_face.xml");

detection.Init(image.Size()); //###### BUSY BUG HERE !!! #####

//Detection::Objects objects;
//detection.Detect(image, objects, 3, 0.2);
while(1)
{
    Sleep(1000);
    printf("Sleep to show the  ###### BUSY BUG ##### \n");
}

}

[Yermalayeu Ihar]

It is a short sample of using Simd::Detection and it doesn't check correctness of input data. I would check the result of execution of functions Test::Load() and Detection::Load(). Might be the paths to the files are not correct.

[Harley]

Dear Yermalayeu Ihar, thanks for your prompt reply.
Waiting for your solution :)

[Yermalayeu Ihar]

I can't check path to files on your computer. At my computer the example works fine.

[Harley]

Dear Yermalayeu Ihar, I set a sleeping here:

        void Run()
        {
            while (_run)
            {
                if (_tasks.size())
                {
                    _tasks.front()->Run();
                    std::unique_lock<std::mutex> lk(_mutex);
                    _tasks.pop();
                    _cv.notify_one();
                }
                else
                {
                    Sleep(1);/// + [8/26/2016 harley]
                }
                std::this_thread::yield();
            }
        }

[Yermalayeu Ihar]

My current platform is Windows 10 (VS 2015). Could you tell me about your platform?
Because this bug depends on used OS or Compiler.
Your solution of the problem isn't general because it uses Windows' specific function ::Sleep().

[Yermalayeu Ihar]

I have reproduced this bug. There is a problem: if we insert sleep(1) it will lead to decreasing of performance (about 20%). Now I am seeking better solution.

[Yermalayeu Ihar]

I have fixed this bug. Thank you for your bug report.

[Harley]

Dear Yermalayeu Ihar, I tested your updated version, and it runs really faster than simply adding a Sleeping, Well done! My platform is Windows 10 (VS 2012 Express).
By the way, could you let me know why SIMD does not support TreeBased .xml model file, only support StumpBased .xml model file ?

Harley@HK,China

[Yermalayeu Ihar]

Because there isn't any effective way to optimize (with using SIMD) calculation of these cascades. Also because using of stump based cascedes is default behevior of trainig utiltities in OpenCV (and there isn't any tree based cascades in standard OpenCV samples of HAAR and LBP cascades).