SURF with CUDA is not faster by a noticeable amount

I wrote a program to compare the increase in performance between SURF with CUDA and without it. The implementation with CUDA only resulted in a speed up of 1.25x times.

The following is my code.

while (true) 
    {

        stream1.read(Frame_1);
        stream1.read(Frame_2);          

        cvtColor(Frame_1,Frame_1,CV_BGR2GRAY);
        cvtColor(Frame_2,Frame_2,CV_BGR2GRAY);
        float t1,t2;

t1 = getTickCount();

        Ptr<xfeatures2d::SURF> surf = xfeatures2d::SURF::create();
        Mat descriptor_1,descriptor_2;
        vector<KeyPoint> keypoint_1,keypoint_2;

        surf->detect(Frame_1,keypoint_1);
        surf->compute(Frame_1,keypoint_1,descriptor_1);
        surf->detect(Frame_2,keypoint_2);
        surf->compute(Frame_2,keypoint_2,descriptor_2); 
        Ptr<DescriptorMatcher> dscMatcher = DescriptorMatcher::create("BruteForce");
        vector< vector< DMatch > >  matches;

        dscMatcher->knnMatch(descriptor_1,descriptor_2,matches,2);

t1 = getTickCount() - t1;

t2 = getTickCount();
        cuda::SURF_CUDA surf_cuda ;

        cuda::GpuMat key_1_GPU,key_2_GPU,desc_1_GPU,desc_2_GPU,img_1,img_2;
        img_1.upload(Frame_1);
        img_2.upload(Frame_2);
        surf_cuda(img_1,cuda::GpuMat(),key_1_GPU,desc_1_GPU);
        surf_cuda(img_2,cuda::GpuMat(),key_2_GPU,desc_2_GPU);

        surf_cuda.downloadKeypoints(key_1_GPU,keypoint_1);
        surf_cuda.downloadKeypoints(key_2_GPU,keypoint_2);

        Ptr<cuda::DescriptorMatcher> dscMatcher1 = cuda::DescriptorMatcher::createBFMatcher();
        vector< vector< DMatch > >  matches1;       
        dscMatcher1->knnMatch(desc_1_GPU,desc_2_GPU,matches1,2);
        t2 = getTickCount() - t2;

        cout << "No cuda : " << t1/getTickFrequency() << "    With Cuda : "<< t2/getTickFrequency() << endl;

if (waitKey(30) >= 0)
        break;

    }

The output which is the time taken gave the following results on the average. Without CUDA it took 0.54 seconds and with CUDA it took 0.43 seconds. I am implementing the code on a NVidia Jetson TX2. The images that I am processing has a size of 900 x 1440.

Then I proceeded to do the same test with ORB which only resulted in a speed up of 1.6 times.

I am wondering whether the problem is with the code I wrote or is it with the hardware. Is there a standard test to see whether the CUDA is working properly on the board?

Some details

• OS : Ubuntu 16.04
• OpenCV : 3.4
• CUDA : 9.0

did you build OpenCV yourself? Are you sure the correct nvcc compiler options for your platform (compute level) were used here?

256 NVIDIA CUDA® cores do not really provide a lot of compute power, but it should be a bit faster than the ARM Cortex on the board (unless that software uses very specific ARM optimizations)

You could attempting to build this with a CUDA 8 toolkit also. I’ve seen some performance regressions using CUDA 9 on Pascal devices in my software.

Thank you for your reply. I did built the OpenCV myself. I tried the example tutorial to confirm that the CUDA was installed.

I used the following for compiling the program

nvcc -m64 -arch=compute_62 -code=sm_62 -x c++ <i>files</i> ......

were any optimization flags given, such as -O2 or -O3 ?

also note that giving any device debug flags such as -G would make the code extremely slow.

I hadn’t specified any optimization flags earlier. I compiled with -O2 flag, but the results were still the same. I haven’t specified any debug flags