The match function of SFace in C++ returns incorrect results about opencv_zoo

[fengyuentau]

Python's API (match) is actually calling the C++ implementation. So that must be something wrong in your code.

---

printf("ret=%.10lf\n",ret);

You dont need l to print a double when using printf. See a duplicate here: #140 (comment)

[zhaodianbo]

printf("ret=%.10f\n",ret); is not working！ result is same as %.10lf
tried std::cout <<std::fixed << std::setprecision(10) << ret << std::endl; not working!

I also tested use Go, result is also wrong

[zhaodianbo]

Actually, I first tested with Go code and found that the comparison results were incorrect. That's why I used C++ code to test it. However, I found that the results were inconsistent with Python's results. The code is very simple, which confuses me.

[fengyuentau]

-    cv::Mat feature1=getFaceFeature(faceDetector,faceRecognizer,image1);
-    cv::Mat feature2=getFaceFeature(faceDetector,faceRecognizer,image2);
+    cv::Mat feature1=getFaceFeature(faceDetector,faceRecognizer,image1).clone();
+    cv::Mat feature1=getFaceFeature(faceDetector,faceRecognizer,image1).clone();

Classic C++ misuse.

1. Your feature is extracted in a function, so it is a local variable.
2. When you call getFaceFeature the first time, feature is extracted and then returned, resulting in a shallow copy to feature1. In my changes above, .clone() makes it a deep copy.
3. When you call getFaceFeature the second time, local variable feature is reused to save memory, and it is overwritten by the features of image2. This changes the values in feature1 as well due to shallow copy.
4. So you get feature1 and feature2, both have exact same values, resulting in very high cosine score (because they are the same one!).

[zhaodianbo]

Thank you very much, the correct value was returned after using the clone() method. I will continue testing in the future, thx！
I check the source code of get feature, outputarray of mat type just give a shallow copy ,so mat need to clone if use it later.