A gaze estimation method for simultaneously estimating eye gaze and facial gaze

Abstract

The present invention relates to a sight estimation method for simultaneously estimating eye sight and facial sight, comprising the following steps: obtaining a left eye image, a right eye image and a facial image; performing standardized coordinate system processing on the left eye image, the right eye image and the facial image respectively; inputting the standardized left eye image, the right eye image and the facial image into a sight estimation model, and outputting predicted eye sight and facial sight; the sight estimation model comprises: a feature extraction layer, a feature information conversion layer, a first feature communication layer, a second feature communication layer, a first fully connected layer and a second fully connected layer. The present invention can simultaneously predict eye sight and facial sight.

Description

Technical Field

[0001] The present invention relates to the technical field of sight line estimation, and in particular to a sight line estimation method for simultaneously estimating eye sight lines and facial sight lines. Background Art

[0002] Gaze estimation refers to the study of estimating the gaze direction or gaze point of a person using eye images or face images as processing objects. According to the different processing objects, it can be divided into eye gaze estimation and facial gaze estimation. Existing methods regard eye gaze estimation and facial gaze estimation as two independent tasks. Most works only use eye images to estimate eye gaze or only use face images to estimate face gaze. Some works use both eye images and face images, but they still only estimate eye gaze or face gaze.

[0003] For example, Fisher et al. used both eye images and facial images to estimate line of sight, but the facial image was only used to extract head posture information, and the eye image was mainly processed, ignoring the rich line of sight related information of the facial image. The model's prediction result was the eye line of sight; Chen et al. proposed a multi-region dilated convolution network structure, and for the eye image, an additional dilated convolution layer was added to extract features. However, this method only used eye features as a supplement to facial features, and the model's prediction result was facial line of sight.

[0004] Cai et al. introduced the multi-head self-attention (MHSA) module based on iTracker and designed the iTracker-MHSA network. Figure 1 As shown in the figure: first use three different backbone networks to extract the left eye features, right eye features and facial features respectively, then regard the three features as a sequence of length 3, calculate the multi-head self-attention of this sequence, fuse the processed binocular features first, and then splice them to the facial features, and use multiple fully connected layers and ReLU activation functions to predict the facial line of sight.

[0005] However, the multi-head self-attention module is sufficient to fully exchange eye feature information and facial feature information. Further fusion after this module will cause the eye sight features to interfere with the representation of facial sight features, making the facial sight features contain irrelevant noise. In addition, this method simply superimposes the eye information, does not make further constraints based on the spatial relationship between the eye sight and the facial sight, and does not use the true value of the eye sight to directly constrain the expression of the eye features. Summary of the invention

[0006] The technical problem to be solved by the present invention is to provide a sight line estimation method for simultaneously estimating eye sight line and facial sight line, which can simultaneously predict eye sight line and facial sight line.

[0007] The technical solution adopted by the present invention to solve the technical problem is: to provide a sight line estimation method for simultaneously estimating eye sight lines and facial sight lines, comprising the following steps:

[0008] Acquire a left-eye image, a right-eye image, and a facial image;

[0009] Performing standardized coordinate system processing on the left eye image, the right eye image and the face image respectively;

[0010] The standardized left eye image, right eye image and facial image are input into a sight line assessment model, and the predicted eye sight line and facial sight line are output; wherein the sight line assessment model includes:

[0011] A feature extraction layer, used for performing feature extraction on the left eye image, right eye image and face image after the standardization process, to obtain left eye feature information in the left eye standardization space, right eye feature information in the right eye standardization space and facial feature information in the face standardization space; wherein the facial feature information includes sight line features and facial appearance features;

[0012] A feature information conversion layer, used for converting the left eye feature information in the left eye standardized space into the left eye feature information in the facial standardized coordinate system, and converting the right eye feature information in the right eye standardized space into the right eye feature information in the facial standardized coordinate system;

[0013] A first feature exchange layer is used to input the left eye feature information of the left eye standardized space, the right eye feature information of the right eye standardized space, and the facial feature information of the facial standardized space to obtain a first eye output feature and a first facial output feature; wherein the first eye output feature includes a first left eye output feature and a first right eye output feature;

[0014] The second feature exchange layer is used to input the left eye feature information in the facial standardized coordinate system, the right eye feature information in the facial standardized coordinate system and the facial feature information in the facial standardized space to obtain a second eye output feature and a second facial output feature; wherein the second eye output feature includes a second left eye output feature and a second right eye output feature;

[0015] a first fully connected layer, configured to input the first eye output feature and the first facial output feature, and output a predicted eye sight line in an eye normalized space and a first facial sight line in a facial normalized space, respectively;

[0016] The second fully connected layer is used to input the second eye output feature and the second facial output feature, and respectively output the predicted eye sight line in the facial standardized space and the second facial sight line in the facial standardized space.

[0017] The specific steps of performing standardized coordinate system processing on the left eye image, the right eye image and the face image are as follows:

[0018] Rotate the real camera coordinate system so that the z-axis of the camera points to the center of the left eye, and ensure that the x-axis of the camera and the x-axis of the left eye coordinate system are coplanar; then move the camera so that the distance between the camera and the left eye is a preset value. The camera coordinate system at this time is called the left eye standardized coordinate system;

[0019] Rotate the real camera coordinate system so that the z-axis of the camera points to the center of the right eye, and ensure that the x-axis of the camera and the x-axis of the right eye coordinate system are coplanar; then move the camera so that the distance between the camera and the right eye is a preset value. The camera coordinate system at this time is called the right eye standardized coordinate system;

[0020] The real camera coordinate system is rotated so that the z-axis of the camera points to the center of the face, and the x-axis of the camera is ensured to be coplanar with the x-axis of the face coordinate system; then the camera is moved so that the distance between the camera and the face is a preset value. The camera coordinate system at this time is called the normalized face coordinate system.

[0021] The sight loss function of the sight evaluation model is: in, represents the predicted left eye gaze in the normalized left eye space, Represents the true left eye line of sight in the left eye normalized space, represents the predicted first facial gaze in the normalized face space, represents the predicted second facial gaze in the face normalized space, represents the facial gaze in the true face normalized space, represents the predicted right eye gaze in the right eye normalized space, Represents the true right eye line of sight in the right eye normalized space, represents the predicted left eye gaze in the normalized face space, represents the left eye gaze in the real face normalized space, represents the predicted right eye gaze in the normalized face space, Represents the right eye gaze in the true face normalized space.

[0022] The line of sight evaluation model also includes: a personal identity fully connected layer, which is used to input the facial appearance features and output a predicted personal identity number.

[0023] The face recognition loss function of the sight line evaluation model is: Among them, N represents the number of samples, M represents the number of categories, and y icis a sign function, if the true category of sample i is equal to c, it takes 1, otherwise it takes 0; p ic is the predicted probability that the observed sample i belongs to category c.

[0024] The line of sight assessment model also includes: a yaw angle fully connected layer, which is used to input the difference between the second left eye output feature and the second right eye output feature, and output the difference in yaw angle between the left eye line of sight and the right eye line of sight in the predicted facial standardized coordinate system.

[0025] The yaw angle loss function of the line of sight evaluation model is: in, represents the difference in the yaw angles of the left and right eye sights in the predicted facial normalization space, represents the second left eye output feature output through the second feature communication layer, represents the second right eye output feature output by the second feature communication layer, Represents the difference in yaw angles between the left and right eye gazes in the real face normalization space.

[0026] The line of sight evaluation model also includes: a fully connected layer for pitch angle, which is used to input the second left eye output feature plus the second right eye output feature minus the second facial output feature, and output the predicted value of the left eye line of sight pitch angle plus the right eye line of sight pitch angle minus the facial line of sight pitch angle in the facial standardized space.

[0027] The pitch angle loss function of the sight line evaluation model is: in, It represents the difference between the left and right eye sights and the face sight pitch angle in the predicted face standardization space. represents the second left eye output feature output by the second feature communication layer, represents the second right eye output feature output by the second feature communication layer, represents the second facial output feature output by the second feature exchange layer, Represents the difference between the actual left and right eye sights and the facial sight pitch angle in the actual facial normalized space.

[0028] Beneficial Effects

[0029] Due to the adoption of the above technical scheme, the present invention has the following advantages and positive effects compared with the prior art: the present invention increases the true value of eye sight, so that the network can simultaneously predict the eye sight and facial sight, and at the same time, considering the changes in sight under different standardized coordinate systems, a dual-channel MHSA structure is designed to make the true value of sight correspond to the image; the present invention also separates some appearance features from facial features for face recognition, and utilizes the spatial geometric relationship between the sight of both eyes and the sight of the face during gaze to constrain the expression of eye features and facial features, so that the extracted features are more accurate and effective. BRIEF DESCRIPTION OF THE DRAWINGS

[0030] Figure 1 It is a schematic diagram of the iTracker-MHSA network structure in the prior art;

[0031] Figure 2 is a flow chart of an embodiment of the present invention;

[0032] Figure 3 is a schematic diagram of the network structure of the sight line assessment model in an embodiment of the present invention;

[0033] Figure 4 is a diagram of the processing process of the left eye center standardized coordinate system in an embodiment of the present invention;

[0034] Figure 5 It is a graph of the relationship between the left and right eye gazes and the facial gaze during fixation;

[0035] Figure 6 It is a schematic diagram of pitch angle, yaw angle and roll angle. DETAILED DESCRIPTION

[0036] The present invention will be further described below in conjunction with specific embodiments. It should be understood that these embodiments are only used to illustrate the present invention and are not intended to limit the scope of the present invention. In addition, it should be understood that after reading the content taught by the present invention, those skilled in the art can make various changes or modifications to the present invention, and these equivalent forms fall within the scope limited by the appended claims of the application equally.

[0037] An embodiment of the present invention relates to a gaze estimation method for simultaneously estimating eye gaze and facial gaze, such as Figure 2 As shown, the following steps are included: obtaining a left eye image, a right eye image and a facial image; performing standardized coordinate system processing on the left eye image, the right eye image and the facial image respectively; inputting the standardized left eye image, the right eye image and the facial image into a sight line assessment model, and outputting predicted eye sight lines and facial sight lines.

[0038] Among them, the sight line assessment model, such as Figure 3 As shown, including:

[0039] A feature extraction layer, used for performing feature extraction on the left eye image, the right eye image and the face image after the standardization process, to obtain left eye feature information of the left eye standardized space (recorded as the first left eye feature information), right eye feature information of the right eye standardized space (recorded as the first right eye feature information) and facial feature information of the facial standardized space (including sight line features and appearance features);

[0040] a feature information conversion layer, used for converting the left eye feature information in the left eye standardized space into left eye feature information in a facial standardized coordinate system (recorded as second left eye feature information), and converting the right eye feature information in the right eye standardized space into right eye feature information in a facial standardized coordinate system (recorded as second right eye feature information);

[0041] A first feature exchange layer is used to input the left eye feature information of the left eye standardized space, the right eye feature information of the right eye standardized space, and the facial feature information of the facial standardized space to obtain a first eye output feature and a first facial output feature; wherein the first eye output feature includes a first left eye output feature and a first right eye output feature;

[0042] The second feature exchange layer is used to input the left eye feature information in the facial standardized coordinate system, the right eye feature information in the facial standardized coordinate system and the facial feature information in the facial standardized space to obtain a second eye output feature and a second facial output feature; wherein the second eye output feature includes a second left eye output feature and a second right eye output feature;

[0043] a first fully connected layer, configured to input the first eye output feature and the first facial output feature, and output a predicted eye sight line in an eye normalized space and a first facial sight line in a facial normalized space, respectively;

[0044] A second fully connected layer is used to input the second eye output feature and the second facial output feature, and output the predicted eye sight line in the facial standardized space and the second facial sight line in the facial standardized space respectively;

[0045] A personal identity fully connected layer, used to input the facial appearance features and output a predicted personal identity number;

[0046] A first feature processing module, configured to obtain a first processed feature by subtracting a right eye feature from a left eye feature in a second eye output feature output through the facial normalization space MHSA layer;

[0047] A yaw angle fully connected layer, used for inputting a difference between the second left eye output feature and the second right eye output feature, and outputting a difference between the left eye sight and the right eye sight in the yaw angle in the predicted facial standardized coordinate system;

[0048] The pitch angle fully connected layer is used to input the second left eye output feature plus the second right eye output feature minus the second facial output feature, and output the predicted value of the left eye sight pitch angle plus the right eye sight pitch angle minus the facial sight pitch angle in the facial standardized space.

[0049] This implementation has designed a network that can simultaneously predict eye gaze (Eye Gaze) and facial gaze (Face Gaze) - EFGaze (i.e., the gaze evaluation model of this implementation). Its main improvements include: adding the true value of eye gaze so that the network can simultaneously predict eye gaze and facial gaze; considering the changes in gaze under different standardized coordinate systems, a dual-channel MHSA structure is designed so that the true value of gaze corresponds to the image; in addition, some surface features are separated from facial features for face recognition; the spatial geometric relationship between binocular gaze and facial gaze during gaze is used to constrain the expression of eye features and facial features, so that the extracted features are more accurate and effective. According to the above improvements, different loss functions can be constructed, and the specific designs are as follows:

[0050] 1. Loss gaze

[0051] In this embodiment, the midpoint between the inner and outer corners of the left eye is taken as the left eye center, the midpoint between the inner and outer corners of the right eye is taken as the right eye center, and the center of the six points including the four corners of the left and right eyes and the two corners of the mouth is taken as the facial center.

[0052] Eye images and facial images need to be processed in a standardized coordinate system before being input into the EFGaze network, such as Figure 4 As shown, specifically, first rotate the real camera coordinate system so that the camera's z-axis points to the center of the left eye, and ensure that the camera's x-axis is coplanar with the x-axis of the left eye coordinate system; then move the camera so that the distance between the camera and the left eye is a preset value. The camera coordinate system at this time is called the left eye standardized coordinate system. In the same way, make the camera's z-axis point to the center of the right eye and the center of the face respectively, and you can get the right eye standardized coordinate system and the face standardized coordinate system, specifically:

[0053] Rotate the real camera coordinate system so that the z-axis of the camera points to the center of the right eye, and ensure that the x-axis of the camera and the x-axis of the right eye coordinate system are coplanar; then move the camera so that the distance between the camera and the right eye is a preset value. The camera coordinate system at this time is called the right eye standardized coordinate system;

[0054] The real camera coordinate system is rotated so that the z-axis of the camera points to the center of the face, and the x-axis of the camera is ensured to be coplanar with the x-axis of the face coordinate system; then the camera is moved so that the distance between the camera and the face is a preset value. The camera coordinate system at this time is called the normalized face coordinate system.

[0055] The left eye image in the left eye standardized coordinate system, the right eye image in the right eye standardized coordinate system, and the facial image in the facial standardized coordinate system are input into EFGaze, and the corresponding features are extracted through the backbone network (i.e., the feature extraction layer). The rotation matrix R from the left eye normalized coordinate system to the face normalized coordinate system faceleye The left eye features in the normalized coordinate system of the left eye Transformed into the left eye feature in the facial standardized coordinate system The formula is The rotation matrix R from the right eye normalized coordinate system to the face normalized coordinate system facereye The right eye features in the right eye standardized coordinate system Transformed into the right eye feature in the facial standardized coordinate system The formula is The facial features in the facial standardized coordinate system remain unchanged. (like Figure 3 As shown, let’s assume that the dimensions of these features are 3×64), where As Figure 3 The top three rows are the inputs of the standardized spatial MHSA pathway. and As Figure 3 The middle and lower three rows show the inputs to the MHSA pathway in face normalization space.

[0056] Then, the eye features and facial features processed by the two MHSA pathways are passed through multiple fully connected layers and ReLU activation functions to output the predicted eye sight and facial sight, respectively. The L1 loss function of these predicted sights and the corresponding true values ​​is calculated to obtain the sight loss function:

[0057]

[0058] in, represents the predicted left eye gaze in the normalized left eye space, Represents the true left eye line of sight in the left eye normalized space, represents the predicted first facial gaze in the normalized face space, represents the predicted second facial gaze in the face normalized space, represents the facial gaze in the true face normalized space, represents the predicted right eye gaze in the right eye normalized space, Represents the true right eye line of sight in the right eye normalized space, represents the predicted left eye gaze in the normalized face space, represents the left eye gaze in the real face normalized space, represents the predicted right eye gaze in the normalized face space, Represents the right eye gaze in the true face normalized space.

[0059] 2. Face recognition loss function (Loss person id → Loss pid)

[0060] For the facial features separated from the face image by the backbone network, except for the line of sight estimation Some are also used for face recognition, such as Figure 3 As shown in the small and medium squares, let's assume that the dimension of this part of the features is 1×320. After passing this part of the features through the fully connected layer and the ReLU activation function, the id of the current face is output, and the cross entropy loss function is used as the face recognition loss function, which is:

[0061]

[0062] Among them, N represents the number of samples, M represents the number of categories, and y ic is a sign function, if the true category of sample i is equal to c, it takes 1, otherwise it takes 0; p ic is the predicted probability that the observed sample i belongs to category c.

[0063] 3. Loss yaw

[0064] When a person looks at an object, the left and right eye sights and the facial sights satisfy certain spatial geometric relationships, such as Figure 5 As shown, g L is the left eye sight, g R is the right eye sight, g T is the facial sight line, T is the object point being looked at, at this time, the left eye center, the right eye center and the object point form a triangle, the left eye sight line and the right eye sight line are the two sides of the triangle respectively, and the projection of the facial sight line on this triangle is the midline of the triangle passing through the object point.

[0065] Figure 6 This is a diagram of pitch, yaw and roll angles. Figure 5 and Figure 6 It can be found that the pitch angles of the left eye and the right eye are the same, only the yaw angles are different. Based on this characteristic, the present embodiment designs a yaw angle loss function. Figure 3 As shown in the figure, the left eye feature output by the MHSA pathway in the facial standardization space is subtracted from the right eye feature, and then passed through multiple fully connected layers and ReLU activation functions to predict the difference in the yaw angles of the left and right eyes in the facial standardization space. The yaw angle loss function is:

[0066]

[0067] in, represents the difference in the yaw angles of the left and right eye sights in the predicted facial normalization space, represents the second left eye output feature output through the second feature communication layer, represents the second right eye output feature output by the second feature communication layer, Represents the difference in yaw angles between the left and right eye gazes in the real face normalization space.

[0068] 4. Loss pitch

[0069] Since the facial sight line is at the center of the left eye, the projection of the triangle formed by the center of the right eye and the object point is the midline of the triangle passing through the object point, the sum of the left and right eye sight lines is the same as the yaw angle of the facial sight line, and only the pitch angle is different. Based on this feature, the pitch angle loss function is designed. Figure 3 As shown in the figure, the left eye feature output by the MHSA pathway in the facial standardization space is added to the right eye feature and then subtracted from the facial feature. After passing through multiple fully connected layers and ReLU activation functions, the difference between the left and right eye sights and the facial sight pitch angle in the facial standardization space is predicted. The pitch angle loss function is:

[0070]

[0071] in, It represents the difference between the left and right eye sights and the face sight pitch angle in the predicted face standardization space. represents the second left eye output feature output by the second feature communication layer, represents the second right eye output feature output by the second feature communication layer, represents the second facial output feature output by the second feature exchange layer, Represents the difference between the actual left and right eye sights and the facial sight pitch angle in the actual facial normalized space.

[0072] 5. Total loss function

[0073] The total loss function of the gaze evaluation model in this embodiment is the sum of the above four loss functions, that is, Losstotal=Loss gaze+Loss pid+Loss yaw+Loss pitch.

Claims

1. A method for simultaneously estimating eye sight and facial sight, characterized in that: The following steps are involved: Acquire a left-eye image, a right-eye image, and a facial image; Performing standardized coordinate system processing on the left eye image, the right eye image and the face image respectively; The standardized left eye image, right eye image and facial image are input into a sight line assessment model, and the predicted eye sight line and facial sight line are output; wherein the sight line assessment model includes: A feature extraction layer, used for performing feature extraction on the left eye image, right eye image and face image after the standardization process, to obtain left eye feature information in the left eye standardization space, right eye feature information in the right eye standardization space and facial feature information in the face standardization space; wherein the facial feature information includes sight line features and facial appearance features; A feature information conversion layer, used for converting the left eye feature information in the left eye standardized space into the left eye feature information in the facial standardized coordinate system, and converting the right eye feature information in the right eye standardized space into the right eye feature information in the facial standardized coordinate system; A first feature exchange layer is used to input the left eye feature information of the left eye standardized space, the right eye feature information of the right eye standardized space, and the facial feature information of the facial standardized space to obtain a first eye output feature and a first facial output feature; wherein the first eye output feature includes a first left eye output feature and a first right eye output feature; The second feature exchange layer is used to input the left eye feature information in the facial standardized coordinate system, the right eye feature information in the facial standardized coordinate system and the facial feature information in the facial standardized space to obtain a second eye output feature and a second facial output feature; wherein the second eye output feature includes a second left eye output feature and a second right eye output feature; a first fully connected layer, configured to input the first eye output feature and the first facial output feature, and output a predicted eye sight line in an eye normalized space and a first facial sight line in a facial normalized space, respectively; The second fully connected layer is used to input the second eye output feature and the second facial output feature, and respectively output the predicted eye sight line in the facial standardized space and the second facial sight line in the facial standardized space.

2. The method for simultaneously estimating eye sight lines and facial sight lines according to claim 1, characterized in that: The specific steps of performing standardized coordinate system processing on the left eye image, the right eye image and the face image are as follows: Rotate the real camera coordinate system so that the z-axis of the camera points to the center of the left eye, and ensure that the x-axis of the camera and the x-axis of the left eye coordinate system are coplanar; then move the camera so that the distance between the camera and the left eye is a preset value. The camera coordinate system at this time is called the left eye standardized coordinate system; Rotate the real camera coordinate system so that the z-axis of the camera points to the center of the right eye, and ensure that the x-axis of the camera and the x-axis of the right eye coordinate system are coplanar; then move the camera so that the distance between the camera and the right eye is a preset value. The camera coordinate system at this time is called the right eye standardized coordinate system; The real camera coordinate system is rotated so that the z-axis of the camera points to the center of the face, and the x-axis of the camera is ensured to be coplanar with the x-axis of the face coordinate system; then the camera is moved so that the distance between the camera and the face is a preset value. The camera coordinate system at this time is called the normalized face coordinate system.

3. The method for simultaneously estimating eye sight lines and facial sight lines according to claim 1, characterized in that: The sight loss function of the sight evaluation model is: in, represents the predicted left eye gaze in the normalized left eye space, Represents the true left eye line of sight in the normalized left eye space, represents the predicted first facial gaze in the normalized face space, represents the predicted second facial gaze in the face normalized space, represents the facial gaze in the true face normalized space, represents the predicted right eye gaze in the right eye normalized space, Represents the true right eye line of sight in the right eye normalized space, represents the predicted left eye gaze in the normalized face space, represents the left eye gaze in the real face normalized space, represents the predicted right eye gaze in the normalized face space, Represents the right eye gaze in the true face normalized space.

4. The method for simultaneously estimating eye sight lines and facial sight lines according to claim 1, characterized in that: The line of sight evaluation model also includes: a personal identity fully connected layer, which is used to input the facial appearance features and output a predicted personal identity number.

5. The method for simultaneously estimating eye sight lines and facial sight lines according to claim 4, characterized in that: The face recognition loss function of the sight line evaluation model is: Among them, N represents the number of samples, M represents the number of categories, and y ic is a sign function, if the true category of sample i is equal to c, it takes 1, otherwise it takes 0; p ic is the predicted probability that the observed sample i belongs to category c.

6. The method for simultaneously estimating eye sight lines and facial sight lines according to claim 1, characterized in that: The sight line assessment model also includes: The yaw angle fully connected layer is used to input the difference between the second left eye output feature and the second right eye output feature, and output the difference between the left eye sight and the right eye sight in the yaw angle in the predicted facial standardized coordinate system.

7. The method for simultaneously estimating eye sight lines and facial sight lines according to claim 6, characterized in that: The yaw angle loss function of the line of sight evaluation model is: in, represents the difference in the yaw angles of the left and right eye sights in the predicted facial normalization space, represents the second left eye output feature output by the second feature communication layer, represents the second right eye output feature output by the second feature exchange layer, Represents the difference in yaw angles between the left and right eye gazes in the real face normalization space.

8. The method for simultaneously estimating eye sight lines and facial sight lines according to claim 1, characterized in that: The sight line assessment model also includes: The pitch angle fully connected layer is used to input the second left eye output feature plus the second right eye output feature minus the second facial output feature, and output the predicted value of the left eye sight pitch angle plus the right eye sight pitch angle minus the facial sight pitch angle in the facial standardized space.

9. The method for simultaneously estimating eye sight lines and facial sight lines according to claim 8, characterized in that: The pitch angle loss function of the sight line evaluation model is: in, It represents the difference between the left and right eye sights and the face sight pitch angle in the predicted face standardization space. represents the second left eye output feature output by the second feature communication layer, represents the second right eye output feature output by the second feature exchange layer, represents the second facial output feature output by the second feature exchange layer, Represents the difference between the actual left and right eye sights and the facial sight pitch angle in the actual facial normalized space.

Images