OpenCV recompile with CuDNN for Win10 undeclared identifier error

I wrote a python script using opencv to compare 7k pictures amongst themselves and flag duplicates, but even on my PC:

  • 6 core AMD Ryzen 5 3600X (12 logical ones),
  • Win10,
  • 16GB RAM,
  • 1TB SSD,

it takes too long. I added threading to make it go faster but even so it took like 10 hours to check 49 pictures against the 7k images. So now I am onto the next idea to speed it up by recompiling opencv to use my nvidia GTX 1660 Ti cause it has more cores. I am using this guide here but with more recent versions: CUDA Toolkit 11.8 and OpenCV 4.7.

I got to step 12.6, after using cmake to do the config and building the ALL_BUILD:
12.6. Right-click “ALL_BUILD ” and click build .

This step fails with the error DIFFERENT_SIZES_EXTRA: undeclared identifier.

INSTANTIATE_TEST_CASE_P(CUDA_Warping, Remap, testing::Combine(
    DIFFERENT_SIZES_EXTRA, // this line here!!!
    testing::Values(MatType(CV_8UC1), MatType(CV_8UC3), MatType(CV_8UC4), MatType(CV_32FC1), MatType(CV_32FC3), MatType(CV_32FC4)),
    testing::Values(Interpolation(cv::INTER_NEAREST), Interpolation(cv::INTER_LINEAR), Interpolation(cv::INTER_CUBIC)),
    testing::Values(BorderType(cv::BORDER_REFLECT101), BorderType(cv::BORDER_REPLICATE), BorderType(cv::BORDER_CONSTANT), BorderType(cv::BORDER_REFLECT), BorderType(cv::BORDER_WRAP)),

I am wondering if I can just comment out these two test cases?

I wouldn’t as it looks like you are mixing up versions of the OpenCV main and contrib repositoriess.

DIFFERENT_SIZES_EXTRA was added to the main repository in

so if you have the test cases from

but not the defines you are using a more recent commit from the contrib than the main repo.

I would rebuild from the tip of the master of both the main and contrib repositories.

maybe you dont even need gpu support for your task, but a less naive algorithm !
say, you have 40 new images, and want to find out, if those are already in your 7k database:

  • use phash, to generate 64bit (8 byte) signatures from your database images. (this will take a while, but you only need to do that once) put those into a dict with sig as key, filepath as value & serialize to disk.
  • for new images, get a phash signature again, compare that to those in the dict (a matter of microseconds), if it’s already there, discard, else save image to database folder & append sig/filepath to dict (and serialize dict again)

see, i’m regularily checking my webbrowser cache for images (fodder for machine learning), extracting like 1k images from there, and check those against a ~200k (and ever growing !) image database, works like a charm !

note, that phash will also work with cropped / resized / gray / flipped / rotated images, unlike your naive subtraction !

oh nice thanks, yeah the 7k is just the sample, i got a much bigger set to compare with later :slight_smile: