I have managed to get a face recognition program working using LBPH. I am using ChiSquare comparison for the histograms. With K=1 and threshold of 100. Other values gridx=y=8 and radius = 1 and samples =8.
When I print the returned value of the recognise method, the list
List<independentpair\<detectedface, list\<scoredannotation\<string="">>>> contains the image name and a score, this appears as</independentpair\<detectedface,>
The values are the probability of a particular class (or person), based on the number of nearest neighbour instances requested (or the total number of selected neighbours (with distance below the threshold) if it's smaller), and number of neighbour instances belonging to each class. If K was 3, and two of the neighbours were 'Jon' and one was 'Sina', then the values would be 0.6666 and 0.3333 respectively. Normally, you'd just pick the class with the highest probability as the best guess.
If you would like to refer to this comment somewhere else in this project, copy and paste the following link:
Many thanks, I did guess that it was the probability . I wanted to compare the distance value computed by your code against that of opencv when using the same set of images - is there a method that will expose this value?, I looked in the KNNAnnotator class - the distance value is discarded after comparison.
If you would like to refer to this comment somewhere else in this project, copy and paste the following link:
There's no way to do it using the KNNAnnotator, but you could compute it directly as follows:
final FaceDetector<DetectedFace, FImage> faceDetector = HaarCascadeDetector.BuiltInCascade.frontalface_alt2
.load();
final LocalLBPHistogram.Extractor<DetectedFace> extr = new LocalLBPHistogram.Extractor<DetectedFace>(
new IdentityAligner<DetectedFace>());
final FImage img1 = ImageUtilities.readF(new File("image1.jpg"));
final FImage img2 = ImageUtilities.readF(new File("image2.jpg"));
//assuming a single face in each image
FloatFV feature1 = extr.extractFeature(faceDetector.detectFaces(img1).get(0)).getFeatureVector();
FloatFV feature2 = extr.extractFeature(faceDetector.detectFaces(img2).get(0)).getFeatureVector();
System.out.println(FloatFVComparison.CHI_SQUARE.compare(feature1, feature2));
If you would like to refer to this comment somewhere else in this project, copy and paste the following link:
I have managed to get a face recognition program working using LBPH. I am using ChiSquare comparison for the histograms. With K=1 and threshold of 100. Other values gridx=y=8 and radius = 1 and samples =8.
When I print the returned value of the recognise method, the list
List<independentpair\<detectedface, list\<scoredannotation\<string="">>>> contains the image name and a score, this appears as</independentpair\<detectedface,>
[[org.openimaj.image.processing.face.detection.DetectedFace@15ebe925,[(Current7, 1.0)]]]
What does the float value signify ?
Because if I set K=2, I get two images returned with equal values
[[org.openimaj.image.processing.face.detection.DetectedFace@41f99142,[(man, 0.5), (Current9, 0.5)]]]
In OpenCV I am used to getting a distance value which is normally below the threshold i.e in this case a value less than 100
Can you please explain how this value returned by openIMAJ recogniser is used.
The values are the probability of a particular class (or person), based on the number of nearest neighbour instances requested (or the total number of selected neighbours (with distance below the threshold) if it's smaller), and number of neighbour instances belonging to each class. If K was 3, and two of the neighbours were 'Jon' and one was 'Sina', then the values would be 0.6666 and 0.3333 respectively. Normally, you'd just pick the class with the highest probability as the best guess.
Many thanks, I did guess that it was the probability . I wanted to compare the distance value computed by your code against that of opencv when using the same set of images - is there a method that will expose this value?, I looked in the KNNAnnotator class - the distance value is discarded after comparison.
There's no way to do it using the
KNNAnnotator, but you could compute it directly as follows:final FaceDetector<DetectedFace, FImage> faceDetector = HaarCascadeDetector.BuiltInCascade.frontalface_alt2 .load(); final LocalLBPHistogram.Extractor<DetectedFace> extr = new LocalLBPHistogram.Extractor<DetectedFace>( new IdentityAligner<DetectedFace>()); final FImage img1 = ImageUtilities.readF(new File("image1.jpg")); final FImage img2 = ImageUtilities.readF(new File("image2.jpg")); //assuming a single face in each image FloatFV feature1 = extr.extractFeature(faceDetector.detectFaces(img1).get(0)).getFeatureVector(); FloatFV feature2 = extr.extractFeature(faceDetector.detectFaces(img2).get(0)).getFeatureVector(); System.out.println(FloatFVComparison.CHI_SQUARE.compare(feature1, feature2));Many thanks for your quick reply.