Hello, I am an HPC(high perfromance computing) Engineer which is interested in and good at visual descriptor/extraction algorithm in order to comtribute to other Visual Alogrithm 's Research, meaning that most of the visual backend algorithm could not concern too much about the fontend speed , I could I contribute on your BAD project.

Great work! I have achieved satisfying performance in my own dataset using BAD.
But i can't figure out what 'cropping_scale' and 'sigma' mean for HASHSIFT. What impact does adjusting these two parameters have on the results?

Thanks for your excellent work, I've tried your BAD and HashSIFT descriptors and it works very well! And with opencv greater than 4.0, we can use the parallel_for_ module to speed up the computation of descriptors, it works in my project when in PC. But when I transfer it to my project in mobile device(both ios and android, opencv version > 4.0, by built static lib,), I found the parallel module cannot make any improvement.
I would be thankful for any help!

Hello, I'm trying to reproduce the results of the paper in terms of computational cost using my own computer (intel i7), at the moment I'm able to obtain the same description time in ms using opencv 4.5.5 and oxford dataset.

The issue comes when I arrive to reproduce the matching results, the paper says that detecting 2000 ORB keypoints, descripting them with BAD-256 and matching them using BF Matcher in 100 runs the mean average computational cost is 4.65 ms but I'm obtaining 210 ms.

For the moment I'm using the code that is provided on the repo, called "stereo_demo.cpp", is there something that I'm doing wrong?

Thanks in advance

Hi, I've tried your HashSIFT descriptor and it works very well! But we found that the time cost in computing hashsift descriptors（with no parallel_for_） is much slower than opencv SIFT(1000 SIFT keypoints, 261ms vs 80ms in my PC). Is this normal?
Here is part of my sample code:

`

auto sift = cv::SIFT::create(1000);
std::vector<cv::KeyPoint> kpts1, kpts2;
cv::Mat sift_descs1, sift_descs2, hs_descs1, hs_descs2;

sift->detect(img1, kpts1);
sift->detect(img2, kpts2);

auto start = system_clock::now();
sift->compute(img1, kpts1, sift_descs1);
sift->compute(img2, kpts2, sift_descs2);
auto end = system_clock::now();
std::cout << "SIFT compute descs time: " << duration_cast<milliseconds>(end - start).count() << " ms." << std::endl;

auto hashsift = upm::HashSIFT::create(6.75, upm::HashSIFT::SIZE_256_BITS);

start = system_clock::now();
hashsift->compute(img1, kpts1, hs_descs1);
hashsift->compute(img2, kpts2, hs_descs2);
end = system_clock::now();
std::cout << "HashSIFT compute descs time: " << duration_cast<milliseconds>(end - start).count() << " ms." << std::endl;`

Any help will be grateful!