FaceRecognizer

FaceRecognizer

class FaceRecognizer : public Algorithm

All face recognition models in OpenCV are derived from the abstract base class FaceRecognizer, which provides a unified access to all face recongition algorithms in OpenCV.

class FaceRecognizer : public Algorithm
{
public:
    //! virtual destructor
    virtual ~FaceRecognizer() {}

    // Trains a FaceRecognizer.
    virtual void train(InputArray src, InputArray labels) = 0;

    // Updates a FaceRecognizer.
    virtual void update(InputArrayOfArrays src, InputArray labels);

    // Gets a prediction from a FaceRecognizer.
    virtual int predict(InputArray src) const = 0;

    // Predicts the label and confidence for a given sample.
    virtual void predict(InputArray src, int &label, double &confidence) const = 0;

    // Serializes this object to a given filename.
    virtual void save(const String& filename) const;

    // Deserializes this object from a given filename.
    virtual void load(const String& filename);

    // Serializes this object to a given cv::FileStorage.
    virtual void save(FileStorage& fs) const = 0;

    // Deserializes this object from a given cv::FileStorage.
    virtual void load(const FileStorage& fs) = 0;
};

Description

I’ll go a bit more into detail explaining FaceRecognizer, because it doesn’t look like a powerful interface at first sight. But: Every FaceRecognizer is an Algorithm, so you can easily get/set all model internals (if allowed by the implementation). Algorithm is a relatively new OpenCV concept, which is available since the 2.4 release. I suggest you take a look at its description.

Algorithm provides the following features for all derived classes:

  • So called “virtual constructor”. That is, each Algorithm derivative is registered at program start and you can get the list of registered algorithms and create instance of a particular algorithm by its name (see Algorithm::create()). If you plan to add your own algorithms, it is good practice to add a unique prefix to your algorithms to distinguish them from other algorithms.
  • Setting/Retrieving algorithm parameters by name. If you used video capturing functionality from OpenCV highgui module, you are probably familar with cvSetCaptureProperty(), cvGetCaptureProperty(), VideoCapture::set() and VideoCapture::get(). Algorithm provides similar method where instead of integer id’s you specify the parameter names as text Strings. See Algorithm::set() and Algorithm::get() for details.
  • Reading and writing parameters from/to XML or YAML files. Every Algorithm derivative can store all its parameters and then read them back. There is no need to re-implement it each time.

Moreover every FaceRecognizer supports the:

  • Training of a FaceRecognizer with FaceRecognizer::train() on a given set of images (your face database!).
  • Prediction of a given sample image, that means a face. The image is given as a Mat.
  • Loading/Saving the model state from/to a given XML or YAML.

Note

When using the FaceRecognizer interface in combination with Python, please stick to Python 2. Some underlying scripts like create_csv will not work in other versions, like Python 3.

Setting the Thresholds

Sometimes you run into the situation, when you want to apply a threshold on the prediction. A common scenario in face recognition is to tell, whether a face belongs to the training dataset or if it is unknown. You might wonder, why there’s no public API in FaceRecognizer to set the threshold for the prediction, but rest assured: It’s supported. It just means there’s no generic way in an abstract class to provide an interface for setting/getting the thresholds of every possible FaceRecognizer algorithm. The appropriate place to set the thresholds is in the constructor of the specific FaceRecognizer and since every FaceRecognizer is a Algorithm (see above), you can get/set the thresholds at runtime!

Here is an example of setting a threshold for the Eigenfaces method, when creating the model:

// Let's say we want to keep 10 Eigenfaces and have a threshold value of 10.0
int num_components = 10;
double threshold = 10.0;
// Then if you want to have a cv::FaceRecognizer with a confidence threshold,
// create the concrete implementation with the appropiate parameters:
Ptr<FaceRecognizer> model = createEigenFaceRecognizer(num_components, threshold);

Sometimes it’s impossible to train the model, just to experiment with threshold values. Thanks to Algorithm it’s possible to set internal model thresholds during runtime. Let’s see how we would set/get the prediction for the Eigenface model, we’ve created above:

// The following line reads the threshold from the Eigenfaces model:
double current_threshold = model->getDouble("threshold");
// And this line sets the threshold to 0.0:
model->set("threshold", 0.0);

If you’ve set the threshold to 0.0 as we did above, then:

//
Mat img = imread("person1/3.jpg", CV_LOAD_IMAGE_GRAYSCALE);
// Get a prediction from the model. Note: We've set a threshold of 0.0 above,
// since the distance is almost always larger than 0.0, you'll get -1 as
// label, which indicates, this face is unknown
int predicted_label = model->predict(img);
// ...

is going to yield -1 as predicted label, which states this face is unknown.

Getting the name of a FaceRecognizer

Since every FaceRecognizer is a Algorithm, you can use Algorithm::name() to get the name of a FaceRecognizer:

// Create a FaceRecognizer:
Ptr<FaceRecognizer> model = createEigenFaceRecognizer();
// And here's how to get its name:
String name = model->name();

FaceRecognizer::train

Trains a FaceRecognizer with given data and associated labels.

C++: void FaceRecognizer::train(InputArrayOfArrays src, InputArray labels) = 0
Parameters:
  • src – The training images, that means the faces you want to learn. The data has to be given as a vector<Mat>.
  • labels – The labels corresponding to the images have to be given either as a vector<int> or a

The following source code snippet shows you how to learn a Fisherfaces model on a given set of images. The images are read with imread() and pushed into a std::vector<Mat>. The labels of each image are stored within a std::vector<int> (you could also use a Mat of type CV_32SC1). Think of the label as the subject (the person) this image belongs to, so same subjects (persons) should have the same label. For the available FaceRecognizer you don’t have to pay any attention to the order of the labels, just make sure same persons have the same label:

// holds images and labels
vector<Mat> images;
vector<int> labels;
// images for first person
images.push_back(imread("person0/0.jpg", CV_LOAD_IMAGE_GRAYSCALE)); labels.push_back(0);
images.push_back(imread("person0/1.jpg", CV_LOAD_IMAGE_GRAYSCALE)); labels.push_back(0);
images.push_back(imread("person0/2.jpg", CV_LOAD_IMAGE_GRAYSCALE)); labels.push_back(0);
// images for second person
images.push_back(imread("person1/0.jpg", CV_LOAD_IMAGE_GRAYSCALE)); labels.push_back(1);
images.push_back(imread("person1/1.jpg", CV_LOAD_IMAGE_GRAYSCALE)); labels.push_back(1);
images.push_back(imread("person1/2.jpg", CV_LOAD_IMAGE_GRAYSCALE)); labels.push_back(1);

Now that you have read some images, we can create a new FaceRecognizer. In this example I’ll create a Fisherfaces model and decide to keep all of the possible Fisherfaces:

// Create a new Fisherfaces model and retain all available Fisherfaces,
// this is the most common usage of this specific FaceRecognizer:
//
Ptr<FaceRecognizer> model =  createFisherFaceRecognizer();

And finally train it on the given dataset (the face images and labels):

// This is the common interface to train all of the available cv::FaceRecognizer
// implementations:
//
model->train(images, labels);

FaceRecognizer::update

Updates a FaceRecognizer with given data and associated labels.

C++: void FaceRecognizer::update(InputArrayOfArrays src, InputArray labels)
Parameters:
  • src – The training images, that means the faces you want to learn. The data has to be given as a vector<Mat>.
  • labels – The labels corresponding to the images have to be given either as a vector<int> or a

This method updates a (probably trained) FaceRecognizer, but only if the algorithm supports it. The Local Binary Patterns Histograms (LBPH) recognizer (see createLBPHFaceRecognizer()) can be updated. For the Eigenfaces and Fisherfaces method, this is algorithmically not possible and you have to re-estimate the model with FaceRecognizer::train(). In any case, a call to train empties the existing model and learns a new model, while update does not delete any model data.

// Create a new LBPH model (it can be updated) and use the default parameters,
// this is the most common usage of this specific FaceRecognizer:
//
Ptr<FaceRecognizer> model =  createLBPHFaceRecognizer();
// This is the common interface to train all of the available cv::FaceRecognizer
// implementations:
//
model->train(images, labels);
// Some containers to hold new image:
vector<Mat> newImages;
vector<int> newLabels;
// You should add some images to the containers:
//
// ...
//
// Now updating the model is as easy as calling:
model->update(newImages,newLabels);
// This will preserve the old model data and extend the existing model
// with the new features extracted from newImages!

Calling update on an Eigenfaces model (see createEigenFaceRecognizer()), which doesn’t support updating, will throw an error similar to:

OpenCV Error: The function/feature is not implemented (This FaceRecognizer (FaceRecognizer.Eigenfaces) does not support updating, you have to use FaceRecognizer::train to update it.) in update, file /home/philipp/git/opencv/modules/contrib/src/facerec.cpp, line 305
terminate called after throwing an instance of 'cv::Exception'

Please note: The FaceRecognizer does not store your training images, because this would be very memory intense and it’s not the responsibility of te FaceRecognizer to do so. The caller is responsible for maintaining the dataset, he want to work with.

FaceRecognizer::predict

C++: int FaceRecognizer::predict(InputArray src) const = 0
C++: void FaceRecognizer::predict(InputArray src, int& label, double& confidence) const = 0

Predicts a label and associated confidence (e.g. distance) for a given input image.

Parameters:
  • src – Sample image to get a prediction from.
  • label – The predicted label for the given image.
  • confidence – Associated confidence (e.g. distance) for the predicted label.

The suffix const means that prediction does not affect the internal model state, so the method can be safely called from within different threads.

The following example shows how to get a prediction from a trained model:

using namespace cv;
// Do your initialization here (create the cv::FaceRecognizer model) ...
// ...
// Read in a sample image:
Mat img = imread("person1/3.jpg", CV_LOAD_IMAGE_GRAYSCALE);
// And get a prediction from the cv::FaceRecognizer:
int predicted = model->predict(img);

Or to get a prediction and the associated confidence (e.g. distance):

using namespace cv;
// Do your initialization here (create the cv::FaceRecognizer model) ...
// ...
Mat img = imread("person1/3.jpg", CV_LOAD_IMAGE_GRAYSCALE);
// Some variables for the predicted label and associated confidence (e.g. distance):
int predicted_label = -1;
double predicted_confidence = 0.0;
// Get the prediction and associated confidence from the model
model->predict(img, predicted_label, predicted_confidence);

FaceRecognizer::save

Saves a FaceRecognizer and its model state.

C++: void FaceRecognizer::save(const String& filename) const

Saves this model to a given filename, either as XML or YAML.

Parameters:
  • filename – The filename to store this FaceRecognizer to (either XML/YAML).
C++: void FaceRecognizer::save(FileStorage& fs) const

Saves this model to a given FileStorage.

Parameters:

Every FaceRecognizer overwrites FaceRecognizer::save(FileStorage& fs) to save the internal model state. FaceRecognizer::save(const String& filename) saves the state of a model to the given filename.

The suffix const means that prediction does not affect the internal model state, so the method can be safely called from within different threads.

FaceRecognizer::load

Loads a FaceRecognizer and its model state.

C++: void FaceRecognizer::load(const String& filename)
C++: void FaceRecognizer::load(const FileStorage& fs) = 0

Loads a persisted model and state from a given XML or YAML file . Every FaceRecognizer has to overwrite FaceRecognizer::load(FileStorage& fs) to enable loading the model state. FaceRecognizer::load(FileStorage& fs) in turn gets called by FaceRecognizer::load(const String& filename), to ease saving a model.

createEigenFaceRecognizer

C++: Ptr<FaceRecognizer> createEigenFaceRecognizer(int num_components=0, double threshold=DBL_MAX)
Parameters:
  • num_components – The number of components (read: Eigenfaces) kept for this Prinicpal Component Analysis. As a hint: There’s no rule how many components (read: Eigenfaces) should be kept for good reconstruction capabilities. It is based on your input data, so experiment with the number. Keeping 80 components should almost always be sufficient.
  • threshold – The threshold applied in the prediciton.

Notes:

  • Training and prediction must be done on grayscale images, use cvtColor() to convert between the color spaces.
  • THE EIGENFACES METHOD MAKES THE ASSUMPTION, THAT THE TRAINING AND TEST IMAGES ARE OF EQUAL SIZE. (caps-lock, because I got so many mails asking for this). You have to make sure your input data has the correct shape, else a meaningful exception is thrown. Use resize() to resize the images.
  • This model does not support updating.

Model internal data:

  • num_components see createEigenFaceRecognizer().
  • threshold see createEigenFaceRecognizer().
  • eigenvalues The eigenvalues for this Principal Component Analysis (ordered descending).
  • eigenvectors The eigenvectors for this Principal Component Analysis (ordered by their eigenvalue).
  • mean The sample mean calculated from the training data.
  • projections The projections of the training data.
  • labels The threshold applied in the prediction. If the distance to the nearest neighbor is larger than the threshold, this method returns -1.

createFisherFaceRecognizer

C++: Ptr<FaceRecognizer> createFisherFaceRecognizer(int num_components=0, double threshold=DBL_MAX)
Parameters:
  • num_components – The number of components (read: Fisherfaces) kept for this Linear Discriminant Analysis with the Fisherfaces criterion. It’s useful to keep all components, that means the number of your classes c (read: subjects, persons you want to recognize). If you leave this at the default (0) or set it to a value less-equal 0 or greater (c-1), it will be set to the correct number (c-1) automatically.
  • threshold – The threshold applied in the prediction. If the distance to the nearest neighbor is larger than the threshold, this method returns -1.

Notes:

  • Training and prediction must be done on grayscale images, use cvtColor() to convert between the color spaces.
  • THE FISHERFACES METHOD MAKES THE ASSUMPTION, THAT THE TRAINING AND TEST IMAGES ARE OF EQUAL SIZE. (caps-lock, because I got so many mails asking for this). You have to make sure your input data has the correct shape, else a meaningful exception is thrown. Use resize() to resize the images.
  • This model does not support updating.

Model internal data:

  • num_components see createFisherFaceRecognizer().
  • threshold see createFisherFaceRecognizer().
  • eigenvalues The eigenvalues for this Linear Discriminant Analysis (ordered descending).
  • eigenvectors The eigenvectors for this Linear Discriminant Analysis (ordered by their eigenvalue).
  • mean The sample mean calculated from the training data.
  • projections The projections of the training data.
  • labels The labels corresponding to the projections.

createLBPHFaceRecognizer

C++: Ptr<FaceRecognizer> createLBPHFaceRecognizer(int radius=1, int neighbors=8, int grid_x=8, int grid_y=8, double threshold=DBL_MAX)
Parameters:
  • radius – The radius used for building the Circular Local Binary Pattern. The greater the radius, the
  • neighbors – The number of sample points to build a Circular Local Binary Pattern from. An appropriate value is to use `` 8`` sample points. Keep in mind: the more sample points you include, the higher the computational cost.
  • grid_x – The number of cells in the horizontal direction, 8 is a common value used in publications. The more cells, the finer the grid, the higher the dimensionality of the resulting feature vector.
  • grid_y – The number of cells in the vertical direction, 8 is a common value used in publications. The more cells, the finer the grid, the higher the dimensionality of the resulting feature vector.
  • threshold – The threshold applied in the prediction. If the distance to the nearest neighbor is larger than the threshold, this method returns -1.

Notes:

  • The Circular Local Binary Patterns (used in training and prediction) expect the data given as grayscale images, use cvtColor() to convert between the color spaces.
  • This model supports updating.

Model internal data: