the question of face_recognition_sface_2021dec_int8 for java

[12341234-yangboran]

[12341234-yangboran]

        FaceRecognizerSF faceRecognizerSF = FaceRecognizerSF.create("C:/Users/y1793/Downloads/face_recognition_sface_2021dec.onnx", "");
        Mat alignFaceA = new Mat();
        faceRecognizerSF.alignCrop(imageA, faceA, alignFaceA);
        Mat alignFaceB = new Mat();
        faceRecognizerSF.alignCrop(imageB, faceB, alignFaceB);

        
        Mat featureA = new Mat();
        faceRecognizerSF.feature(alignFaceA, featureA);
        Mat featureB = new Mat();
        faceRecognizerSF.feature(alignFaceB, featureB);

[fengyuentau]

Expression: vector subscript out of range

Check whether faceA or faceB is empty.

[12341234-yangboran]

faceRecognizerSF.alignCrop(imageA, faceA.row(0), alignFaceA);
like this?

[fengyuentau]

See Python demo:

opencv_zoo/models/face_recognition_sface/demo.py

Line 80 in f872270

result = recognizer.match(img1, face1[0][:-1], img2, face2[0][:-1])

opencv_zoo/models/face_recognition_sface/sface.py

Lines 54 to 63 in f872270

	def match(self, image1, face1, image2, face2):
	feature1 = self.infer(image1, face1)
	feature2 = self.infer(image2, face2)
	if self._disType == 0: # COSINE
	cosine_score = self._model.match(feature1, feature2, self._disType)
	return 1 if cosine_score >= self._threshold_cosine else 0
	else: # NORM_L2
	norml2_distance = self._model.match(feature1, feature2, self._disType)
	return 1 if norml2_distance <= self._threshold_norml2 else 0

A face box should be sent to alignCrop. The face box should have the format as follows:

[x1, y1, w, h, lm1_x, lm1_y, ..., lm5_x, lm5_y]

, where x1, y1 are the top-left corner point and lm1, lm2, lm3, lm4, lm5 are right eye, left eye, nose tip, right corner of mouth, left corner of mouth respectively. See https://github.com/opencv/opencv_zoo/tree/main/models/face_detection_yunet for more details.

[12341234-yangboran]

face_recognition_sface_2021dec_int8.onnx may be have something wrong.


there are my image
faceRecognizerSF.match return 0.99999

[12341234-yangboran]

FaceDetectorYN faceDetectorYN = FaceDetectorYN.create("C:/Users/y1793/Downloads/face_detection_yunet_2022mar_int8.onnx", "", new Size(), score_thresh, nms_thresh, top_k);
FaceRecognizerSF faceRecognizerSF = FaceRecognizerSF.create("C:/Users/y1793/Downloads/face_recognition_sface_2021dec_int8.onnx", "");

[rh-id]

hi @fengyuentau

+1 for this, i think something is wrong with the model scoring. i tested 2 faces that are very different, yet the score indicates it was identical.

my code:
https://github.com/rh-id/a-jarwis/blob/master/ml-engine/src/androidTest/java/m/co/rh/id/a_jarwis/ml_engine/provider/component/FaceEngineTest.java#L77-L92

https://github.com/rh-id/a-jarwis/blob/master/ml-engine/src/main/java/m/co/rh/id/a_jarwis/ml_engine/provider/component/FaceEngine.java#L61-L79

https://github.com/rh-id/a-jarwis/blob/master/ml-engine/src/main/java/m/co/rh/id/a_jarwis/ml_engine/provider/component/FaceEngine.java#L139-L164

[fengyuentau]

hi @fengyuentau

+1 for this, i think something is wrong with the model scoring. i tested 2 faces that are very different, yet the score indicates it was identical.

my code: https://github.com/rh-id/a-jarwis/blob/master/ml-engine/src/androidTest/java/m/co/rh/id/a_jarwis/ml_engine/provider/component/FaceEngineTest.java#L77-L92

https://github.com/rh-id/a-jarwis/blob/master/ml-engine/src/main/java/m/co/rh/id/a_jarwis/ml_engine/provider/component/FaceEngine.java#L61-L79

https://github.com/rh-id/a-jarwis/blob/master/ml-engine/src/main/java/m/co/rh/id/a_jarwis/ml_engine/provider/component/FaceEngine.java#L139-L164

1. Draw the detected face bounding boxes on the image to see whether faces are detected correctly.
2. Make sure you send the correct face bounding box along with its matching image to the API.
3. Could you test your image with the Python demo here? If the results are still wrong, then this model does not satisfy your need.

[rh-id]

hi @fengyuentau , I think the demo.py result is correct only my code is incorrect, you could ignore my questions. thanks and sorry for the trouble

[fengyuentau]

@12341234-yangboran Could you doublecheck with the Python demo as well with your images as input?

[12341234-yangboran]

hi @fengyuentau , I think the demo.py result is correct only my code is incorrect, you could ignore my questions. thanks and sorry for the trouble

You need clone the result of face recogniz feature to a new entity。

[12341234-yangboran]

hi @fengyuentau , I think the demo.py result is correct only my code is incorrect, you could ignore my questions. thanks and sorry for the trouble

like this
Mat featureA = new Mat();
faceRecognizer.feature(alignFaceA, featureA);
featureA = featureA.clone();
Mat featureB = new Mat();
faceRecognizer.feature(alignFaceB, featureB);
featureB = featureB.clone();

[rh-id]

hi @12341234-yangboran

Yes that is right, it works. I think you could close this issue since it seemed the issue is solved.