Why the function match of FaceRecognizerSF alterates the feature?

C++
1

As I known, sorry if I misunderstood, the method match from the class FaceRecognizerSF pass the features as references by using the &. My first question is why?

digss
digss1046×131 9.39 KB

And then came another doubt crossed my mind the usage of match in this example is correct? Because when we call match by the second time, feature 1 and feature 2 are not the same as in the beggining.
Reference code

feature1, feature2;

faceRecognizer->feature(aligned_face1, feature1);

feature1 = feature1.clone()

faceRecognizer->feature(aligned_face2, feature2);

feature2 = feature2.clone()

double cos_score = faceRecognizer->match(feature1, feature2, FaceRecognizerSF::DisType::FR_COSINE);

double L2_score = faceRecognizer->match(feature1, feature2, FaceRecognizerSF::DisType::FR_NORM_L2);
1 Like
2

the answer is here:

and i think, you’ve found a bug ! :]
(const things should not be altered)

1 Like
3

i’d even move the normalize() into the feature() function
as of now, if you match 1 feature against 5 others,
there are 10 normalize() where only 6 are needed

4

Hey, Berak thanks for answering! I totally agree about the normalize() suggestion. I just reproduced the possible inconsistency, using the example from the OpenCV documentation and some print’s. Take a look about the changing of the features after passing by the match method:

tstin
tstin1886×1068 309 KB

5

could you do this somehow as TEXT instead of a screenshot ? ty.

6

First I took this example to my C++ environment and pass an image as input1 and another image as input2, creating 2 templates. After passing by:

// Run feature extraction with given aligned_face

Mat feature1, feature2;

faceRecognizer->feature(aligned_face1, feature1);

feature1 = feature1.clone()

faceRecognizer->feature(aligned_face2, feature2);

feature2 = feature2.clone()

Despite that, my templates are changed after passing by the method match, even with the type const in the declaration of match method in face_recognize.cpp.

double match(InputArray _face_feature1, InputArray _face_feature2, int dis_type) const override
    {
        Mat face_feature1 = _face_feature1.getMat(), face_feature2 = _face_feature2.getMat();
        normalize(face_feature1, face_feature1);
        normalize(face_feature2, face_feature2);

        if(dis_type == DisType::FR_COSINE){
            return sum(face_feature1.mul(face_feature2))[0];
        }else if(dis_type == DisType::FR_NORM_L2){
            return norm(face_feature1, face_feature2);
        }else{
            throw std::invalid_argument("invalid parameter " + std::to_string(dis_type));
        }

    };

The main idea is the features generated along the code aren’t supposed to be imuttable? This doubt came because I’m testing deploying a facial server with Yunet and SFace, which I have to compare each template from different user for enroll. But after passing to many comparsions in match, my templates are changing their original value. The path I followed to work around this problem is creating clones, even between cos_score and l2_score but this sounds weird and maybe there’s a real bug as you said.