MSER slower in 3.0.0 than 2.4.9

asked 2015-07-06 23:42:17 -0500

huckw gravatar image

updated 2015-07-07 11:10:32 -0500

I have a project that I originally developed with opencv2.4.9 that used mser like so...

cv::MserFeatureDetector mser(delta, mnArea, mxArea, maxVariation, minDiversity);
std::vector< std::vector< cv::Point > > ptblobs;
mser(img, ptblobs);

It works just like I expect. I switched to opencv3.0.0 and so had to change the above to this...

cv::Ptr<cv::MSER> mser = cv::MSER::create(delta, mnArea, mxArea, maxVariation, minDiversity);
std::vector< std::vector< cv::Point > > ptblobs;
std::vector<cv::Rect> bboxes;
mser->detectRegions(img, ptblobs, bboxes);

It is MUCH MUCH slower in version 3.0.0 than 2.4.9. If I comment out the detectRegions() line everything else runs same speed as I know its the detectRegions() call and not some other thing or things. What am I doing wrong?

Thanks in advance for your help

EDIT/ADDITION I've come up with an uber simple project that isolates and exhibits the issue. With 2.4.9 I get 1.5sec and with 3.0.0 its 4.4sec. I'm posting the code here just in case its useful...

#include "opencv2/highgui/highgui.hpp"
#include "opencv2\features2d\features2d.hpp"
#include "opencv2\imgproc\imgproc.hpp"
#include <iostream>
#include <chrono>
#include <ctime>

#define USE249
using namespace std;

int main(int argc, char** argv)
    std::cout << "OpenCV version: "
            << CV_MAJOR_VERSION << "." 
            << CV_MINOR_VERSION << "."
            << CV_SUBMINOR_VERSION
            << std::endl;
    cv::Mat im = cv::imread("C:/example.jpg", 1);
    if (im.empty())
        cout << "Cannot open image!" << endl;
        return -1;
    cv::Mat gray;
    cv::cvtColor(im, gray, cv::COLOR_BGR2GRAY);
    int mnArea = 40 * 40;
    int mxArea = im.rows*im.cols*0.4;
    std::vector< std::vector< cv::Point > > ptblobs;
    std::vector<cv::Rect> bboxes;
    std::chrono::time_point<std::chrono::system_clock> start, end;

    start = std::chrono::system_clock::now();
#ifndef USE249
    cv::Ptr<cv::MSER> mser = cv::MSER::create(1, mnArea, mxArea, 0.25, 0.2);
    mser->detectRegions(gray, ptblobs, bboxes);
    cv::MserFeatureDetector mser(1, mnArea, mxArea, 0.25, 0.2);
    mser(gray, ptblobs);
    end = std::chrono::system_clock::now();

    std::chrono::duration<double> elapsed_seconds = end - start;
    std::time_t end_time = std::chrono::system_clock::to_time_t(end);

    std::cout << "finished computation at " << std::ctime(&end_time)
        << "elapsed time: " << elapsed_seconds.count() << "s\n";

    cv::namedWindow("image", cv::WINDOW_NORMAL);
    cv::imshow("image", im);

    return 0;
edit retag flag offensive close merge delete


Are you sure that both versions are built with the same CMAKE flags? It could be due to not using a specific optimization.

StevenPuttemans gravatar imageStevenPuttemans ( 2015-07-07 07:45:09 -0500 )edit

Thats a good question....I used defaults in cmake (the gui version) for both builds. But I went back and checked and as near as I can tell they are essentially the same. I did notice that in 3.0.0 "with IPP" is checked (enabled) by default as I hear there is a subset of IPP for my 2.4.9 build it was disabled completely. Just for giggles I rebuilt 3.0.0 with IPP disabled but it did not change the performance at all. I also tried using a UMat instead of a Mat but it too had no effect.

huckw gravatar imagehuckw ( 2015-07-07 10:40:07 -0500 )edit

Do I have to do anything to specify I want the greyscale version of MSER rather than the color version? The image I run detectRegions on is greyscale.

huckw gravatar imagehuckw ( 2015-07-07 11:11:43 -0500 )edit

No experience with MSER so not really sure there ...

StevenPuttemans gravatar imageStevenPuttemans ( 2015-07-09 04:19:23 -0500 )edit

Maybe this one will help you:

manfeyn gravatar imagemanfeyn ( 2015-11-16 06:27:09 -0500 )edit