Not as good as you said

[gz-happy]

I used gray image, channel: 1, size: 1920 x 1080.
1. These are some my OpenCV code:

 cv::Ptr<cv::MSER> mser_opencv = cv::MSER::create(5, 25, 1500, 0.5);
 timer->Tic();
 mser_opencv->detectRegions(grayimg, contours, bboxes);
 timer->Toc();
 std::cout << "[opencv] detect time: " << timer->Elapsed() << "ms" << std::endl;

2. These are some you methods code:
(1). config:

 alg->set_duplicated_variation(0.2f);
 alg->set_stable_max_variation(0.5f);
 alg->set_stable_delta(5);
 alg->set_min_max_point(25, 0.000723f);

(2). Four algorithms were tested separately：

 timer->Tic();
 mser->extract(mser_res, gray_image);
 timer->Toc();
 std::cout << "[FastMser] detect time: " << timer->Elapsed() << "ms" << std::endl; 

3. Results：
(1). OpenCV: 227ms
(2). img_ms_mser: 158ms
(3). img_linked_linear_mser: 183ms
(4). img_fast_mser_v1: 188ms
(5). img_fast_mser_v2: 160ms

4. So is there a problem with where I set it up? How can we achieve the speed you said.
Thank you very much!

[mmmn143]

Hi, thank you for interest in our work.

With the configuration of TextDetection used in paper:

alg->set_duplicated_variation(0.1f);
alg->set_stable_max_variation(0.5f);
alg->set_min_max_point(20, 0.25f);
alg->set_stable_delta(1);

For a grayscale image with the size of 200M:
In visual studio 2012 (used in paper):

• img_fast_mser_v1 thread-1: 241ms
• img_fast_mser_v1 thread-4: 90ms
• cv_mser thread-1: 5820ms
• cv_mser (comment code in img_opencv3_mser.cpp from 84 to 108) thread-1: 220ms

In visual studio 2019:

• img_fast_mser_v1 thread-1: 250ms
• img_fast_mser_v1 thread-4: 98ms
• cv_mser thread-1: 706ms
• cv_mser (comment code in img_opencv3_mser.cpp from 84 to 108) thread-1: 230ms

With your configuration:

alg->set_duplicated_variation(0.2f);
alg->set_stable_max_variation(0.5f);
alg->set_stable_delta(5);
alg->set_min_max_point(25, 0.000723f);

For a grayscale image with the size of 200M:
In visual studio 2012 (used in paper):

• img_fast_mser_v1 thread-1: 211ms
• img_fast_mser_v1 thread-4: 81ms
• cv_mser thread-1: 668ms
• cv_mser (comment code in img_opencv3_mser.cpp from 84 to 108) thread-1: 224ms

In visual studio 2019:

• img_fast_mser_v1 thread-1: 228ms
• img_fast_mser_v1 thread-4: 91ms
• cv_mser thread-1: 256ms
• cv_mser (comment code in img_opencv3_mser.cpp from 84 to 108) thread-1: 237ms

Conclusion:

1. img_fast_mser_v1 is stable in both visual studio 2012 and 2019 However, cv_mser in visual studio 2019 is much faster than that in visual studio 2012. This may be because new compiler in visual studio 2019 optimize the code in img_opencv3_mser.cpp from 84 to 108. The code uses a stl vector to store pixels extracted from MSERs. Method of push_back in stl vector may be fully optimized;
2. img_fast_mser_v1 is stable in different configurations because it uses an efficient pixel extraction method. However, cv_mser is unstable in different configurations. When the number of pixels in MSERs is large (for example, use small delta and large max_point_ratio), img_fast_mser_v1 will have a obvious speed-up compared to cv_mser;
3. img_fast_mser_v1 or img_fast_mser_v2 is a parallel algorithm. OpenMP may not work on your computer. I notice that the speeds of img_fast_mser_v1 and img_linked_linear_mser are similar. You may use omp_set_num_threads(4) and alg->set_parallel_thread_number(4) to accelerate img_fast_mser_v1 or img_fast_mser_v2;
4. All experiments in paper are evaluated using visual studio 2012. In the configuration of TextDetection, for a grayscale image with the sie of 200M, img_fast_mser_v1 thread-1 is 24 times faster than cv_mser thread-1. The spped-up in visual studio 2019 reduces to 2.8 times.

[mmmn143]

By the way, I only have time on weekends to track down problems.

[gz-happy]

Thank you for your reply, I will try again.