OpenCV ORB GPU implementation slower than CPU

Question

I am trying to run the ORB OpenCV algorithm to the frames of a video and I noticed the CPU version performs a lot faster than the GPU version. Here is the code:

#include <iostream>
#include "opencv2/core/core.hpp"
#include "opencv2/features2d/features2d.hpp"
#include "opencv2/highgui/highgui.hpp"
#include "opencv2/gpu/gpu.hpp"
#include <fstream>
#include <sstream> 
#include <math.h>
#include <omp.h>

#include <algorithm>
#include <vector>
#include <string>

using namespace cv;
using namespace std;
using namespace cv::gpu;

void process_cpu(string vid, int start_frame, int end_frame)
{
VideoCapture myCapture(vid);
Mat frame, gray_frame;
ORB myOrb(400);
Mat descriptors;
vector<KeyPoint> keypoints;

myCapture.set(CV_CAP_PROP_POS_FRAMES, start_frame);

for (int i=0; i<end_frame-start_frame; i++) {
    myCapture.read(frame);
    cvtColor(frame, gray_frame, CV_RGB2GRAY);
    myOrb(gray_frame, Mat(), keypoints, descriptors);
}
myCapture.release();
}

void process_gpu(string vid, int start_frame, int end_frame)
{
VideoCapture myCapture(vid);
Mat frame, gray_frame;
GpuMat gpu_frame;
ORB_GPU myOrb(400);
GpuMat keypoints, descriptors;

myCapture.set(CV_CAP_PROP_POS_FRAMES, start_frame);

for (int i=0; i<end_frame-start_frame; i++) {
    myCapture.read(frame);
    cvtColor(frame, gray_frame, CV_RGB2GRAY);
    gpu_frame.upload(gray_frame);
    myOrb.blurForDescriptor = true;
    myOrb(gpu_frame, GpuMat(), keypoints, descriptors);
}
myCapture.release();
}

int main (int argc, char* argv[])
{
int n = 4;
VideoCapture myCapture(argv[1]);
double frameNumber = myCapture.get(CV_CAP_PROP_FRAME_COUNT);
myCapture.release();

double TimeStart = 0;
double TotalTime = 0;
TimeStart = (double)getTickCount();

process_gpu(argv[1], 0, frameNumber);

TotalTime = (double)getTickCount() - TimeStart;
TotalTime = TotalTime / getTickFrequency();
cout << "Gpu Time : " << TotalTime << endl;

TimeStart = (double)getTickCount();

process_cpu(argv[1], 0, frameNumber);

TotalTime = (double)getTickCount() - TimeStart;
TotalTime = TotalTime / getTickFrequency();
cout << "Cpu Time : " << TotalTime << endl;

return -1;
}

After running this on a video with 3000 frames and 720x480 resolution, the GPU time is 54 sec and the CPU time 24 sec. I get similar results with other videos (not HD). PC specs:

• i7-4770K CPU 3.50 GHz

• NVIDIA GeForce GTX 650

• 16 GB RAM

Other feature detection/description algorithms like SURF perform faster with the GPU implementation on my machine.

Has anyone compared the two implementation of ORB on his machine?

Answer 1

Taken from this post:

cv::ORB applies a GaussianBlur (about 20 lines from the end of orb.cpp) before computing descriptors. There is no way to control this through the public interface.

cv::gpu::ORB_GPU has a public member bool blurForDescriptor, which by default constructs as false. When I set it instead to true, I find that min/avg/max hamming distance drops to 0/7.2/30 bits, which seems much more reasonable.