xfeatures2d BriefDescriptorExtractor not implemented about gocv HOT 4 CLOSED

Keeping the same patterns as used by the other wrappers is the best bet. But I see from the PR that you already also came to that same conclusion 😸

Thanks for working on this @Cartermel

from gocv.

I noticed the keypoints are being copied back over to c++ in a couple of functions:

Should be trivial to implement for BRIEF, I was originally worried about the performance impact of copying the keypoints since, for my use case at least, I'll be running this against thousands of images. But I ran some benchmarks and copying the struct between c++ / go is extremely lightweight, which makes sense since the keypoint struct is very small.

Here's some data for anyone curious:
(I only used time.Now() and time.Since() for measuring elapsed, there's probably a better way but I didn't search)

elapsed times are the average of 3 runs rounded to nearest 100ms

5000 iterations:
No copy: 26200ms
Copy:    27800ms
No copy 1600ms faster 
~5.9% difference

10000 iterations:
No copy: 52000ms
Copy:    55400ms
No copy 3400ms faster
~6.3% difference

This was using this function for computing the descriptors using BRIEF and copying from go (my vscode syntax highlighting is broken...) :

Where VectorOfKeyPoint is a custom struct that's just a ptr to a vectorcv::KeyPoint for testing. And the c++ implementation was:

And the native c++ implementation without copying was this:

Where I passed in the same FastFeatureDector from golang. Not shown is the code for creating the BriefDescriptorExtractor, but its the same as all the other feature2d alg initialization.

This is running on a Ryzen 7 5800X.

TLDR

Copying the keypoints from golang for use in BRIEF results in about a 6% time loss (on my minimal testing), but leaves all the existing keypoint implementation intact. Personally I think this is a fine trade off, I'm interested in knowing what everyone thinks!

In the meantime I'll create a PR for BRIEF and copy the golang keypoints, unless someone's currently building this!

from gocv.

Forgot to add, this was the simple benchmark function:

and the native c++ call was basically exactly the same but with the c call.
And it just uses the lenna photo from opencv.

from gocv.

Released as part of 0.36 so now closing. Thank you!

from gocv.