Skip to main content
Version: 3.15.0

Face Estimation

Age & Gender​

Note: If you need to estimate age and gender on a video stream, see Estimation of age, gender, and emotions in the section Video Stream Processing.

For age and gender estimation create the AgeGenderEstimator class by calling the FacerecService.createAgeGenderEstimator method and providing the configuration file.

Currently, two configuration files are available:

  • age_gender_estimator.xml - First implementation of the AgeGenderEstimator interface.
  • age_gender_estimator_v2.xml - Improved version of the AgeGenderEstimator interface, which provides higher accuracy of age and gender estimation given that you follow Guidelines for Cameras.
  • age_gender_estimator_v3.xml - Improved Gender estimation algorithm.

With AgeGenderEstimator you can estimate age and gender of a captured face using AgeGenderEstimator.estimateAgeGender. The result is the AgeGenderEstimator.AgeGender structure that contains the number of ages (in years), age group (AgeGenderEstimator.Age) and gender (AgeGenderEstimator.Gender).

See the example of using the AgeGenderEstimator in demo.cpp.

You can learn how to estimate Age & Gender in an image in our tutorial Estimating Age, Gender, and Emotions.

Note: Gender Estimation also can be performed through Processing Block API - see Gender Estimation.

Quality​

At the moment there are two quality estimation interfaces: QualityEstimator and FaceQualityEstimator.

  • QualityEstimator provides discrete grade of quality for flare, lighting, noise and sharpness.
  • FaceQualityEstimator provides quality as a single real value that aggregates sample usability for face recognition (i.e. pose, occlusion, noise, blur and lighting), which is very useful for comparing samples of one person from video tracking.

QualityEstimator​

To create the QualityEstimator object, call the FacerecService.createQualityEstimator method by passing the configuration file. Currently, two configuration files are available:

  • quality_estimator.xml – First implementation of the QualityEstimator quality estimation interface.
  • quality_estimator_iso.xml (recommended) – Improved version of the QualityEstimator quality estimation interface, provides higher accuracy of quality estimation.

With QualityEstimator you can estimate the quality of a captured face using QualityEstimator.estimateQuality. The result is the QualityEstimator.Quality structure that contains estimated flare, lighting, noise, and sharpness level.

See the example of using the QualityEstimator in demo.cpp.

FaceQualityEstimator​

To create the FaceQualityEstimator object, call the FacerecService.createFaceQualityEstimator method by passing the configuration file. Currently, there is only one configuration file available, which is face_quality_estimator.xml. With FaceQualityEstimator you can estimate the quality of a captured face using FaceQualityEstimator.estimateQuality. This results in a real number (the greater the number, the higher the quality), which aggregates sample usability for face recognition.

See the example of using the FaceQualityEstimator in demo.cpp.

Liveness​

The main purpose of liveness estimation is to prevent spoofing attacks (using a photo of a person instead of a real face). Currently, you can estimate liveness in three ways - by processing a depth map, by processing an IR image or by processing an RGB image from your camera. You can also estimate liveness using the Active Liveness, which presupposes that a user has to perform a sequence of certain actions.

You can learn how to estimate liveness of a face in our tutorial Liveness Detection.

DepthLivenessEstimator​

To estimate liveness with a depth map, create the DepthLivenessEstimator object using FacerecService.createDepthLivenessEstimator.

The following configuration files are available:

  • depth_liveness_estimator.xml – The first implementation (not recommended; used only for backward compatibility);
  • depth_liveness_estimator_cnn.xml – Implementation based on neural networks (recommended, used in VideoWorker by default).

To use this algorithm, it is necessary to obtain synchronized and registered frames (color image + depth map), use a color image for face tracking / detection, and pass the corresponding depth map to the DepthLivenessEstimator.estimateLiveness method.

To get an estimated result, call the pbio.DepthLivenessEstimator.estimateLiveness method. You'll get one of the following results:

  • DepthLivenessEstimator.NOT_ENOUGH_DATA – too many missing depth values on the depth map.
  • DepthLivenessEstimator.REAL – the observed face belongs to a living person.
  • DepthLivenessEstimator.FAKE – the observed face is taken from a photo.

IRLivenessEstimator​

To estimate liveness using an infrared image from a camera, create the IRLivenessEstimator object using the FacerecService.createIRLivenessEstimator method. Currently, only one configuration file is available – ir_liveness_estimator_cnn.xml (implementation based on neural networks). To use this algorithm, get color frames from the camera in addition to the IR frames.

To get an estimated result, you can call the IRLivenessEstimator.estimateLiveness method. You'll get one of the following results:

  • IRLivenessEstimator.Liveness.NOT_ENOUGH_DATA – Too many missing values in the IR image.
  • IRLivenessEstimator.Liveness.REAL – The observed face belongs to a living person.
  • IRLivenessEstimator.Liveness.FAKE – The observed face is taken from a photo.

Liveness2DEstimator​

To estimate liveness with an RGB map, create the Liveness2DEstimator object using the FacerecService.createLiveness2DEstimator method. Currently, three configuration files are available:

  • liveness_2d_estimator.xml – The first implementation (not recommended; used only for backward compatibility).
  • liveness_2d_estimator_v2.xml – An accelerated and improved version of the current module.
  • liveness_2d_estimator_v3.xml – Liveness Estimation with several additional checks such as face presence, face frontality and image quality.

Two methods can be used to obtain the evaluation result: Liveness2DEstimator.estimateLiveness and Liveness2DEstimator.estimate.

  1. Liveness2DEstimator.estimateLiveness. This method returns a Liveness2DEstimator.Liveness object.

    1.1. Using the liveness_2d_estimator.xml and liveness_2d_estimator_v2.xml configurations allow to obtain one of the following results:
    Liveness2DEstimator.Liveness.NOT_ENOUGH_DATA – Not enough data to make a decision.
    Liveness2DEstimator.Liveness.REAL – The observed person belongs to a living person.
    Liveness2DEstimator.Liveness.FAKE – The observed face is taken from a photo.

    1.2. Using the liveness_2d_estimator_v3.xml configuration allows to obtain one of the following results:
    Liveness2DEstimator.Liveness.REAL - The observed person belongs to a living person.
    Liveness2DEstimator.Liveness.FAKE - The observed face is taken from a photo.
    Liveness2DEstimator.Liveness.IN_PROCESS - Liveness estimation can not be done.
    Liveness2DEstimator.Liveness.NO_FACES - There is no faces on the input image.
    Liveness2DEstimator.Liveness.MANY_FACES - There are more than one face on the input image.
    Liveness2DEstimator.Liveness.FACE_OUT - Observed face is out of the input image boundaries.
    Liveness2DEstimator.Liveness.FACE_TURNED_RIGHT - Observed face is not frontal and turned right.
    Liveness2DEstimator.Liveness.FACE_TURNED_LEFT - Observed face is not frontal and turned left.
    Liveness2DEstimator.Liveness.FACE_TURNED_UP - Observed face is not frontal and turned up.
    Liveness2DEstimator.Liveness.FACE_TURNED_DOWN - Observed face is not frontal and turned down.
    Liveness2DEstimator.Liveness.BAD_IMAGE_LIGHTING - Input image have bad lighting conditions.
    Liveness2DEstimator.Liveness.BAD_IMAGE_NOISE - Input image is too noisy.
    Liveness2DEstimator.Liveness.BAD_IMAGE_BLUR - Input image is too blurry.
    Liveness2DEstimator.Liveness.BAD_IMAGE_FLARE - Input image is too flared.

2.Liveness2DEstimator.estimate. This method returns a Liveness2DEstimator.LivenessAndScore object that contains the following fields:

  • liveness - Object of the Liveness2DEstimator.Liveness class/structure (see above).
  • score – a numeric value in the range from 0 to 1 indicating the probability that the face belongs to a real person (for liveness_2d_estimator.xml only 0 or 1).

Both methods take a RawSample object as output. Examples are available in the demo sample (C++/C#/Android).

Note: Liveness Estimation also can be performed through Processing Block API - see 2D RGB Liveness Estimation.

Timing Characteristics (ms)​

VersionCore i7 4.5 ГГц (Single-Core)Google Pixel 3
liveness_2d_estimator.xml250126 (GPU) / 550 (CPU)
liveness_2d_estimator_v2.xml1020

Quality metrics​

DatasetTAR@FAR=1e-2
CASIA Face Anti-spoofing0.99

Active Liveness​

This type of liveness estimation presupposes that a user needs to perform certain actions. For example, "turn the head", "blink", etc. Estimation is performed through the VideoWorker object based on the video stream. See more detailed description in Video Stream Processing.

Emotions​

Note: If you need to estimate emotions on a video stream, see Estimation of age, gender, and emotions in the section Video Stream Processing.

To estimate emotions, create the EmotionsEstimator object using FacerecService.createEmotionsEstimator and pass the configuration file. Currently, there are two available configuration files:

  • emotions_estimator.xml - allows estimate four emotions: happy, surprised, neutral, angry.
  • emotions_estimator_v2.xml - allows estimate seven emotions: happy, surprised, neutral, angry, disgusted, sad, scared.

With the EmotionsEstimator object you can estimate the emotion of a captured face using the EmotionsEstimator.estimateEmotions method. The result is a vector with the EmotionsEstimator.EmotionConfidence elements which contain emotions with a confidence value.

See the example of using the EmotionsEstimator in demo.cpp.

Note: Emotion Estimation also can be performed through Processing Block API - see Emotion Estimation.

FaceAttributesEstimator​

This class is a universal module for evaluation of face attributes. To get the score, call the FaceAttributesEstimator.estimate(RawSample) method. The evaluation result is an Attribute object that contains the following attributes:

  • score – The probability that a person has the required attribute, a value from 0 to 1 (if the value is set to -1, then this field is not available for the specified type of assessment).

  • verdict – The probability that a person has the required attribute, boolean value (true/false).

  • mask_attribute – An object of the class/structure FaceAttributesEstimator.FaceAttributes.Attribute, which contains the following values:

    • NOT_COMPUTED – No estimation made.
    • NO_MASK – Face without a mask.
    • HAS_MASK – Face with a mask.

  • left_eye_state, right_eye_state – objects of the class/structure FaceAttributesEstimator.FaceAttributes.EyeStateScore, which contains the score attribute and the EyeState structure with the following fields:

    • NOT_COMPUTED – No estimation is made.
    • CLOSED – The eye is closed.
    • OPENED – The eye is open.

Presence of a Mask on the Face​

To check the presence of a mask on the face, use the FaceAttributesEstimator in conjunction with the face_mask_estimator.xml configuration file. This returns the score, verdict, mask attributes in the Attribute object.

State of the Eyes (Open/Closed)​

To check the state of the eyes (open/closed), use the FaceAttributesEstimator in conjunction with the eyes_openness_estimator_v2.xml configuration file. This returns the left_eye_state, right_eye_state attributes in the Attribute object.