Methods and systems for animating facial expressions

Description

CROSS-REFERENCE TO RELATED APPLICATIONS

The present invention claims priority to U.S. Provisional Patent Application Ser. No. 62/677,586, filed May 29, 2018. The disclosure of U.S. Provisional Patent Application Ser. No. 62/677,586 is herein incorporated by reference in its entirety.

FIELD OF THE INVENTION

This invention relates to animating a 3D avatar based upon captured images. More particularly, this invention relates to animating facial expressions for a 3D avatar based on captured images of a user's face and their facial expressions.

BACKGROUND OF THE INVENTION

The animation of computer generated 3D content is becoming increasingly popular. 3D models and avatars are being introduced in many different fields and applications. However, the animation of facial expressions for 3D models can be technically challenging and can require extensive manual animation processes.

SUMMARY OF THE INVENTION

Systems and methods for generating animations for a 3D model in accordance with embodiments of the invention are illustrated. One embodiment includes a set of one or more processors, a memory readable by the set of processors, and instructions stored in the memory that when read by the set of processors directs the set of processors to identify a first set of landmarks from a first set of one or more images of a user's face, identify a neutral frame based on the identified set of landmarks, identify a second set of landmarks from a second set of one or more images of the user's face, classify a facial expression of the user's face in the second set of images based on the identified neutral frame and the second set of landmarks, identify a set of one or more predefined expression weights based on the facial expression, and calculate a set of final morph target weights from the predefined expression weights and the second set of landmarks based on the second set of images, wherein the 3D model is animated based on the calculated set of final morph target weights for morph targets of the 3D model.

In another embodiment, identifying the neutral frame comprises classifying a facial expression of the user's face in the first set of images, adjusting the first set of landmarks for the classified expression to an estimated neutral expression, and setting the adjusted set of landmarks as a neutral expression.

In a further embodiment, classifying a facial expression comprises identifying an expression state based on the identified set of landmarks.

In still another embodiment, adjusting the first set of landmarks comprises identifying an expected ratio between two sets of facial landmarks for the classified facial expression, and adjusting the first set of landmarks based on a difference between the expected ratio for the classified facial expression and an expected ratio for a neutral facial expression.

In a still further embodiment, the first set of landmarks are identified using at least one of a mnemonic descent method (MDM) deep learning approach and an ensemble of regression trees (ERT) approach.

In yet another embodiment, calculating the set of final morph target weights comprises calculating a set of initial morph target weights for the second set of images based on the second set of landmarks, blending the set of initial morph target weights and the predefined expression weights to compute the set of final morph target weights.

In a yet further embodiment, the morph targets of the 3D model comprise a set of one or more base shapes and a set of one or more corrective shapes.

In another additional embodiment, the morph targets of the 3D model are stored in a standardized model file format.

In a further additional embodiment, the instructions further direct the set of processors to adjust the identified neutral frame based on second set of images.

In another embodiment again, the instructions further direct the set of processors to calculate a second set of final morph target weights based on a third set of one or more images, wherein the transition between the first set of final morph target weights and the second set of final morph target weights is based on a linear function to control the rate of morphing between different expressions.

Additional embodiments and features are set forth in part in the description that follows, and in part will become apparent to those skilled in the art upon examination of the specification or may be learned by the practice of the invention. A further understanding of the nature and advantages of the present invention may be realized by reference to the remaining portions of the specification and the drawings, which forms a part of this disclosure.

BRIEF DESCRIPTION OF THE DRAWINGS

FIG. 1 illustrates a system for performing one or more processes to animate facial expressions for a 3D model of a head in accordance with various embodiments of the invention.

FIG. 2 illustrates components of a data processing element that executes processes to animate facial expressions for a 3D model of a head in accordance with various embodiments of the invention.

FIG. 3 illustrates components of a mapping application that executes to map facial expressions captured in images to a 3D model of a head in accordance with an embodiment of the invention.

FIG. 4 illustrates examples of different facial ratios used for estimating neutral expressions.

FIG. 5 illustrates a weighting engine that executes to solve for weights of morph targets of a 3D model of a head in accordance with an embodiment of the invention.

FIG. 6 is a conceptual diagram of an expression state machine for classifying expressions of a user from captured images of the user in accordance with an embodiment of the invention.

FIG. 7 illustrates a flow diagram of a process for calculating weights for morph targets in accordance with an embodiment of the invention.

FIG. 8 illustrates an example function curve for ramping into an expression.

FIG. 9 illustrates an example function curve for blending out of a detected expression.

FIG. 10 illustrates a flow diagram of a process for identifying an initial neutral frame for a user in accordance with an embodiment of the invention.

FIG. 11 illustrates a flow diagram of a process for updating a neutral frame in accordance with an embodiment of the invention.

FIG. 12 illustrates action units of the facial action coding system (FACS).

FIGS. 13-18 illustrate the activation of morph targets for basic shapes of a 3D model.

FIG. 19 illustrates the activation of a morph target for a corrective shape of a 3D model.

DETAILED DESCRIPTION

Turning now to the drawings, systems and methods for animating facial expressions of 3D models from captured images of a user's face in accordance with various embodiments of the invention are illustrated. Systems and processes in accordance with many embodiments of the invention provide a process for identifying landmarks from captured images of a user's face to calculate weights for morph targets of a 3D model in order to animate the expressions of the 3D model. In a number of embodiments, calculating the weights for the morph targets is based on one or more of an identified expressive state of the user's face, a calculated neutral state, and/or identified landmark positions from the user's face. In accordance with some other embodiments, the processes are performed by a “cloud” server system, a user device, and/or combination of devices local and/or remote from a user. Systems and processes in accordance with many embodiments of the invention employ a file format that includes basic shapes that animate individual action units of a face, along with corrective shapes that are used to correct the animations of action unit combinations.

Computing Morph Target Weights for a 3D Model

A system that provides animation of a 3D model of a head from received images in accordance with some embodiments of the invention is shown in FIG. 1. Network 100 includes a communications network 160. The communications network 160 is a network such as the Internet that allows devices connected to the network 160 to communicate with other connected devices. Server systems 110, 140, and 170 are connected to the network 160. Each of the server systems 110, 140, and 170 is a group of one or more servers communicatively connected to one another via networks that execute processes that provide cloud services to users over the network 160. For purposes of this discussion, cloud services are one or more applications that are executed by one or more server systems to provide data and/or executable applications to devices over a network. The server systems 110, 140, and 170 are shown each having three servers in the internal network. However, the server systems 110, 140 and 170 may include any number of servers and any additional number of server systems may be connected to the network 160 to provide cloud services. In accordance with various embodiments of this invention, processes for animating expressions for a 3D model based upon facial expressions captured in images are provided by executing one or more processes on a single server system and/or a group of server systems communicating over network 160.

Users may use personal devices 180 and 120 that connect to the network 160 to perform processes for capturing images (or video) of a user, identifying landmarks, and animating expressions for a 3D model in accordance with various embodiments of the invention. In the illustrated embodiment, the personal devices 180 are shown as desktop computers that are connected via a conventional “wired” connection to the network 160. However, the personal device 180 may be a desktop computer, a laptop computer, a smart television, an entertainment gaming console, or any other device that connects to the network 160 via a “wired” and/or “wireless” connection. The mobile device 120 connects to network 160 using a wireless connection. A wireless connection is a connection that uses Radio Frequency (RF) signals, Infrared signals, or any other form of wireless signaling to connect to the network 160. In FIG. 1, the mobile device 120 is a mobile telephone. However, mobile device 120 may be a mobile phone, Personal Digital Assistant (PDA), a tablet, a smartphone, or any other type of device that connects to network 160 via a wireless connection without departing from this invention. In accordance with some embodiments of the invention, processes for animating expressions of a 3D model based upon facial expressions captured in images are performed by the user device. In many embodiments, an application being executed by the user device may capture or obtain images including a face image and transmit the captured images to a server system that performs additional processing based upon the received images. Although references are made to images throughout this application, one skilled in the art will recognize that processes described in this application can clearly be applied to video (or video frames) without departing from this invention.

Various components of a data processing element that executes one or more processes to provide an animated 3D model of a head in accordance with various embodiments of the invention are illustrated in FIG. 2. Data processing element 200 includes processor 205, image capture device 210, network interface 215, and memory 220. One skilled in the art will recognize that a particular data processing element may include other components that are omitted for brevity without departing from this invention. The processor 205 can include (but is not limited to) a processor, microprocessor, controller, or a combination of processors, microprocessor, and/or controllers that performs instructions stored in the memory 220 to manipulate data stored in the memory. Processor instructions can configure the processor 205 to perform processes in accordance with certain embodiments of the invention. Image capture device 215 can capture and/or retrieve images for the data processing element. Image capture devices can include (but are not limited to) cameras and other sensors that can capture image data of a scene. Network interface 215 allows data processing element 200 to transmit and receive data over a network based upon the instructions performed by processor 205.

Memory 220 includes a mapping application 225, morph targets 230, expression weights 235, and model parameters 240. Mapping applications in accordance with several embodiments of the invention are used to animate expressions for 3D models using model parameters and/or weights that are calculated for morph targets of a 3D model. In some embodiments, the morph targets of a 3D model include a set of base shapes and a set of corrective shapes. Base shapes in accordance with several embodiments of the invention include linear base shapes that represent various action units for a facial model, such as those defined by the facial action coding system (FACS). Corrective shapes in accordance with certain embodiments of the invention are a non-linear function of two or more base shapes used to represent a combined state of multiple action units. In some embodiments, the non-linear function is specified by a corrective interpolator function. Corrective shapes in accordance with several embodiments of the invention are used to correct for rendering errors that can occur when a combination of different base shapes are animated together. In some embodiments, corrective shapes are used to smooth the animation of certain combinations of action units that may result in unnatural looking animations when the animations of the individual base shapes are interpolated.

Although a specific example of a processing system 200 is illustrated in FIG. 2, any of a variety of processing systems can be utilized to perform processes similar to those described herein as appropriate to the requirements of specific applications in accordance with embodiments of the invention.

Components of a mapping application that executes to map images of a user to 3D head movements of a 3D model in accordance with an embodiment of the invention are illustrated in FIG. 3. Mapping application 300 includes image processing engine 305, neutral frame engine 310, weighting engine 315, and mapping engine 320.

Image processing engines in accordance with many embodiments of the invention process images captured by an image capture device to perform a variety of functions, including (but not limited to) landmark identification, camera parameter identification, and image preprocessing. In this example, image processing engine 305 includes landmark engine 307, which can be used to detect and track landmarks from captured images. The motion of one or more landmarks and/or 3D shapes in visible video can be tracked, and the expressions of the 3D model video can be recomputed based on the tracked landmarks.

Facial landmarks in accordance with a number of embodiments of the invention can include various facial features, including (but not limited to) eye corners, mouth corners, noses, chins, etc. Methods that can be used for detecting facial landmarks in accordance with several embodiments of the invention include (but are not limited to) a Mnemonic Descent Method (MDM) deep learning approach and a standard ensemble of regression trees (ERT) approach. The MDM deep learning approach is described in U.S. Pat. No. 9,786,084, entitled “Systems and Methods for Generating Computer Ready Animation Models of a Human Head from Captured Data Images”, issued Oct. 10, 2017, the disclosure of which is incorporated by reference herein in its entirety.

In certain embodiments, image processing engines analyze images to identify parameters of the camera used to capture the images. Image processing engines in accordance with some embodiments of the invention include 3D image processing engines that infer 3D camera parameters from images of faces, such as (but not limited to) video. Camera parameters in accordance with several embodiments of the invention include (but are not limited to) rotation, tilt, and focal length. In some embodiments, camera parameters can be used for several processes including (but not limited to) adjusting identified landmark positions, classifying facial expressions, as well as animating head and/or neck motions. For example, in some embodiments, the inverse of a camera's motion is applied to animate an avatar's neck and head joints. Animating head and/or neck motions can allow for head motions of an avatar to match with movements of a user. Focal lengths can be used to stabilize an avatar's movement, relative to a camera, as the user moves the camera closer and farther away from the user's face.

The neutral frame engine 310 in accordance with many embodiments of the invention internally manages a neutral frame estimate (in 3D) that approximates the shape of the user's face in a neutral expression. In some embodiments, neutral models of a user's face are used to define a neutral state, where other facial expressions can be measured against the neutral state to classify the other facial expressions. Neutral frame engines in accordance with some embodiments of the invention can use any of a number of approaches to classify a neutral expression, including (but not limited to) using an identity solver from 2D and/or 3D landmarks, using statistics based on ratios of 3D face distances, and maintaining an adaptive neutral geometry in 3D by temporally accumulating data across multiple frames.

Examples of different facial ratios used for estimating neutral expressions are illustrated in FIG. 4. In this example, 3D distances are measured between the centers of each person's eyes (illustrated with an x) and the distance between the outer corners of the mouth (illustrated with a y). For faces in a neutral pose, the ratio of y/x has been measured with a Gaussian distribution with a mean of 0.8 and standard deviation of 0.15 across different genders and ethnicities. That ratio can be much higher or lower when a user is not in a neutral expression. In this example, the smiling face illustrated in the first face of each row has a y/x ration that is much higher than 1.0. In many embodiments, several such ratios of distances between facial features can be used to compute an overall neutral face distribution. Neutral face distributions can be used in accordance with several embodiments of the invention for a variety of uses including, but not limited to, estimating a neutral frame, updating an existing neutral frame, and/or classifying a user's current expression. For example, in some embodiments, a user's current expression is adjusted with ratios from a neutral face distribution to estimate the user's neutral frame.

In some embodiments, weighting engines can be used to determine weights for different morph targets of a 3D model. Weighting engines in accordance with a number of embodiments calculate the weights based on results from image processing engines and/or neutral frame engines. Weights for the different morph targets are used to animate different facial expressions on a 3D model. A weighting engine in accordance with an embodiment of the invention is described in greater detail below with reference to FIG. 5.

Mapping engines in accordance with a number of embodiments of the invention can animate expressions of a 3D model by morphing between the morph targets of a 3D model based on the weights calculated by weighting engines. In some embodiments, the morphing is a linear combination of the morph targets with their corresponding weights.

A weighting engine that executes to solve for weights of morph targets of a 3D model of a head in accordance with an embodiment of the invention is illustrated in FIG. 5. Weighting engines in accordance with several embodiments of the invention are used to determine weights for different morph targets of a 3D model. Weighting engine 500 includes a numerical solver 510, expression classifier 520, and morph target weighting generator 530.

Numerical solvers, or optimizers, in accordance with many embodiments of the invention can receive an array of 2D facial landmarks to set up a numerical optimization problem that can solve for the facial geometry and animation weights on facial morph targets. In accordance with many embodiments, the Facial Action Coding System (FACS) morph targets may be constructed by an artist using 3D modelling tools. In several embodiments, the FACS blend shapes may be synthesized using muscle-based physical simulation systems.

In the example of FIG. 5, numerical solver 510 includes facial geometry solver 512 and corrective interpolator 514. Facial geometry solvers can use identified landmarks to determine the facial geometry of a user and to determine their relative positions in order to identify weights for morph targets. In many embodiments, numerical solvers operate in a multi-stage operation that uses landmarks from images to first identify a neutral frame (or an identity-basis) for a user and to then calculate a differential for subsequent images based on the neutral frame to classify expressions from the user's face. In several embodiments, numerical solvers use a neutral state shape in the optimization and update the neural state shape as the system sees more images of the users face in a video sequence. Neutral shapes can be used in conjunction with subsequent images of a user's face to more accurately classify a user's expression.

Correctives are often useful for correcting certain rendering errors that can arise when multiple action units are activated together. Corrective interpolators in accordance with several embodiments of the invention are used to calculate weights to interpolate between various combinations of base shapes.

In accordance with many of these embodiments, numerical solvers can solve for camera parameters. The camera parameters may include, but are not limited to, camera rotation, camera translation, Field of View (FOV), and focal length. In some embodiments, camera parameters are stored in metadata associated with images or in a memory of the device.

In several embodiments, solved facial geometries and/or morph target weights from a numerical solver are used by expression classifiers to classify the facial geometry into an expression state. In some embodiments, this is implemented with an expression state machine. A conceptual diagram of an expression state machine for classifying expressions of a user from captured images of the user in accordance with an embodiment of the invention is illustrated in FIG. 6. Expression states in accordance with some embodiments of the invention include the universal expressions (happy, sad, angry, etc). In certain embodiments, expression states include two special expression states: “neutral” and “unknown”, which can be very helpful in capturing the facial state when a person is talking without a strong facial expression, or when the user is making a face that cannot be classified into a traditional expression. The expression classification can be achieved by analyzing the morph target weights directly, or in combination with image and landmark inputs using machine learning techniques.

Once an expression has been classified, morph target weighting generators in accordance with many embodiments of the invention identify predefined weights for the classified expression. In some embodiments, morph target weighting generators use a combination of the predefined weights and the weights calculated by the numerical solver to determine a final set of weights for the morph targets. By classifying a user's expression and using predefined weights for various emotions, expressions of different 3D models (or avatars) can be standardized, allowing a designer to create morph targets of a model based on predictable weightings for various expressions, such as (but not limited to) a happy face, a surprised face, a sad face, etc. Each expression may include multiple weights for various different morph targets in a 3D model. In certain embodiments, when the expression is unknown, the weights calculated by the numerical solver are used as the final set of weights for the morph targets.

A process for calculating a final set of weights for morph targets in accordance with an embodiment of the invention is illustrated in FIG. 7. Process 700 receives (705) images of a user's face. In certain embodiments, images of a user are captured with an image capture device (e.g., a camera) of the user's device. In accordance with several of these embodiments, an application executed by the user device controls the image capture device to capture the image data. In accordance with some embodiments, images are read from memory. In accordance with some other embodiments, images can be received from another device via a communication network. Optionally, image capture information such as depth information, focal length and/or other camera information may be received along with a captured image. In several embodiments, the make and/or model of the image capture device is received and used to look-up camera parameters stored in a table or other data structure in memory.

In many embodiments, the image is a 3D image capturing an image of a human face. Images in accordance with some embodiments of the invention include depth measurements obtained from a depth sensor including (but not limited to) a time of flight camera and/or a structured light camera. In a number of embodiments, the images include images captured by multiple imagers captured using a plenoptic camera and/or images captured using a coded aperture from which depth information concerning distances to objects within a scene may be determined.

Process 700 identifies (710) landmarks from the received images. The identification of certain features or landmarks of a face can be useful in a variety of applications including (but not limited to) face detection, face recognition, and computer animation. In many embodiments, the identification of landmarks includes the identification of 3D points of a user's face, which may aid in animation and/or modification of a face in a 3D model. In accordance with some embodiments of an invention, a Mnemonic Descent Method (MDM) is used for facial landmark tracking. The goal of tracking the facial landmarks is to predict a set of points on an image of a face that locate salient features (such as eyes, lip corners, jawline, etc.).

Process 700 identifies (715) a neutral frame from the received images. Neutral frames in accordance with several embodiments of the invention represent the positions of facial landmarks when the user's facial expression is in a neutral state. In many embodiments, identifying a neutral frame is an iterative process that refines the identified neutral state as more information about the user's face is gathered. In certain embodiments, the neutral frame is solved (and updated) in 3D from multiple image and/or video measurements. An example of a process for identifying a neutral frame in accordance with an embodiment of the invention is described below with reference to FIG. 10.

Process 700 solves (720) for camera parameters. Camera parameters in accordance with a number of embodiments of the invention include (but are not limited to) focal length, rotation, and tilt. In some embodiments, camera parameters are used for a variety of reasons, including (but not limited to) identifying landmarks, building/maintaining a neutral frame, and adjusting the calculated weights for morph targets.

Process 700 classifies (725) the landmarks as a facial expression. Classifying the landmarks in accordance with many embodiments of the invention include classifying the landmarks as an indication of a particular emotion, such as (but not limited to) happiness, sadness, surprise, and anger. In several embodiments, classification is performed based on differentials calculated based on differentials between a user's neutral frame and the landmarks from a current image of the user's face. Classification as emotions and/or other facial expressions can be performed using a variety of machine learning techniques, including, but not limited to, convolutional neural networks, support vector machines, and decision trees. In several embodiments, expressions can be obtained by classifying 3D morph target weights produced by a numerical solver.

Process 700 identifies (730) predefined expression weights for the classified expression. In some embodiments, predefined expression weights include weights for multiple morph targets that are weighted to be animated together. In the case that the expression is Unknown, processes in a variety of embodiments do not identify any additional predefined weights, but allow the weights to be calculated based on the identified landmarks.

Process 700 calculates (735) a final set of weights for the morph targets to animate a 3D model. Final sets of weights in accordance with a number of embodiments of the invention are calculated based on one or more of the predefined expression weights and the calculated landmark positions. The final weight in accordance with several embodiments of the invention can be computed via a non-linear blend function that combines the predefined weights (on a subset of the morph target shapes) with the solved weights from the numerical solver. In some embodiments, a user's expression is matched closely with the captured images, but is adjusted based on predefined weights for classified expressions. The strength of the user's expression and the predefined weights can be adjusted in accordance with certain embodiments of the invention.

In some embodiments, function curves are used to control the rate of blending weights to morph between different expressions. Function curves in accordance with a number of embodiments of the invention can be used to ensure smooth blending of the morph target weights between detected expressions. In certain embodiments, it can be desirable to provide a fast ramp in of a current detected expression, and a fast blend out of the previous expression (or neutral). In many embodiments, a function curve used for blending into an expression can be different from a function curve for blending out of an expression, providing different rates of change for smoother and more natural transitions.

Example function curves for expression blending are illustrated in FIGS. 8-9. In FIG. 8, an example function curve for ramping into an expression is illustrated. The function curve of this figure is used to modify a morph target corresponding to a “detected” expression, allowing the morph target weight to ramp up very quickly to a full expression (at 1.0). The morph target weight then remains at 1.0, as the solved weight increases past the threshold (set to 0.4 in this example).

Any curves that are not part of a detected expression can be blended out according to a different function curve. FIG. 9 illustrates a function curve for blending out of a detected expression. In this example, the expression weight of a single morph target (i.e., the expression that is being blending out) accelerates down as the solved weight reduces from 1 to 0.

Although processes for calculating morph weights in accordance with the illustrated embodiment is described above, other processes for calculating morph weights in accordance with other embodiments of the invention may be used depending on the requirements of the particular system implemented.

Computing Neutral Frames

The framework in accordance with a number of embodiments of the invention builds an internal 3D representation of the users face in neutral expression (or neutral frame) by combining facial landmarks across multiple frames of video. In several embodiments, neutral frames are computed by a numerical solver and are continually updated over time with additional frames of video. Neutral frames in accordance with many embodiments of the invention are estimated in multiple stages: at the very first frame of video, and subsequently by accumulating multiple frames that are classified as “neutral” by the expression classifier.

An example of a process for identifying an initial neutral frame for a user in accordance with an embodiment of the invention is described below with reference to FIG. 10. In order to identify a neutral frame, process 1000 detects (1005) a user and identifies (1010) landmarks of an initial set of images of the user. In several embodiments, the identified facial landmarks from images of a user's face are fitted to a 3D template based on a distribution of the 3D points. 3D templates in accordance with a number of embodiments of the invention can be generated based on landmarks identified from neutral expressions of several different people, allowing the template to accommodate a broader range of landmark distributions.

Process 1000 classifies (1015) the constellation of landmark positions to identify an expression state (e.g., smiling, sad, neutral, etc.) based on the identified landmarks. In some embodiments, classification of the constellation of landmark positions is performed using one or more machine learning techniques, including (but not limited to) convolutional neural networks, support vector machines, and decision trees. Classifying the landmark positions in accordance with some embodiments of the invention are performed using a heuristics based approach that classifies an expression as neutral based on the distribution of the landmark positions. For example, a variety of factors including (but not limited to) mouth widths, eye shapes, and mouth shapes can be used to predict whether a person is performing a variety of different non-neutral expressions including, but not limited to, smiling, winking, frowning, etc.

When process 1000 determines (1020) that the expression state is neutral, the process sets (1025) the identified landmarks of the initial image set as the initial neutral frame. When process 1000 determines (1020) that the expression state is not neutral, the process attempts to estimate (1030) an initial frame. Estimating an initial frame can include (but is not limited) various methods including factorizing out identified expressions and initializing the neutral frame to a base template that is not customized to any specific user. Factorizing out an identified expression in accordance with a number of embodiments of the invention can include modifying the positions of the identified landmarks based on the determined expression state. For example, in certain embodiments, if a user is determined to be smiling from an initial set of images, neutral frames can be calculated to adjust landmark positions of the neutral frame by reversing the expected changes of a smile from a neutral position based on neutral frames from other models. In a number of embodiments, when no expression state can be identified, a base template can be used as an initial neutral frame until enough images are gathered to generate a better neutral frame for the user. Once the initial neutral frame is set, process 1000 ends. In many embodiments, neutral frames are continuously updated as more images of a user are captured in order to reach a better neutral frame.

Dynamically updated neutral frames can make facial expression tracking significantly robust to several real-world scenarios, where a user's face may go out of frame and returns back into the field of view of the camera, or where the view may switch to a different user's face. User changes can occur because a different user comes into the view of the camera, or because a current user turns their head away from the camera and subsequently comes back into the view. In such cases, it can be important to detect if the user has changed, or whether it is the same user that was previously in view, in order to determine whether to reset the neutral frame for a new user, or to update the existing neutral frame for the same user.

A process for updating a neutral frame in accordance with an embodiment of the invention is illustrated in FIG. 11. Process 1100 detects (1105) a potential user change. In several embodiments, potential user changes can be detected when a user's face can no longer be detected in an image. In certain embodiments, potential user changes can be detected based on sudden or large movements of the image capture device. Such movements can be detected based on sensors, such as (but not limited to) accelerometers, or through analysis of the captured images.

Process 1100 predicts (1110) a new neutral frame based on images captured after the detected potential user change. Process 1100 then compares (1115) the new neutral frame to a prior neutral frame that was calculated prior to the detected potential user change. Process 1100 then determines (1120) whether the difference between the new neutral frame and the prior neutral frame exceeds a particular threshold. When the difference between the new neutral frame and the prior neutral frame exceeds the threshold, the process resets (1125) the neutral frame based on the newly predicted neutral frame. In some embodiments, prior neutral frames are discarded or stored for the case that the original user returns into view. When the difference does not exceed the threshold, or after the neutral frame is reset (1125), process 1100 ends.

Although a process for updating a neutral frame in accordance with the illustrated embodiment is described above, other processes in accordance with other embodiments of the invention may be used depending on the requirements of the particular system implemented.

Model File Format

Mapper frameworks in accordance with several embodiments of the invention are designed to provide maximum flexibility in the characters that are driven by a user's face. In some embodiments, the facial movements of a person's face can be coded based on the facial action coding system (FACS). Action units of FACS are illustrated in FIG. 12. Action units identify a set of basic facial movements that a face can make. The action units can be combined to build expressions. Morph targets in accordance with many embodiments of the invention include base shapes that represent the movements of the individual action units as well as corrective shapes that represent the movements of two or more action units. In the sample of FIG. 12, arrows indicate the direction and magnitude of various action units.

In some embodiments, morph targets (including base shapes and corrective shapes) are stored in a standardized model file format. In certain embodiments, the facial rig is embedded using a modified version of the GL transmission format (gITF), a specification for 3D models. The table below shows an example of a modified gITF specification with 75 morph targets, including 43 base shapes and 32 corrective shapes.

In some embodiments, morph targets are assigned identifiers based on the associated base shape(s) and/or shape type. Morph target identifiers in accordance with many embodiments of the invention are built according to a set of naming codes. In a number of embodiments, naming codes include affixes (i.e., prefixes and suffixes) to identify various characteristics of each morph target. Affixes for the side of a head, whether or not it is a corrective, a number of base shapes are involved in the corrective, higher level shapes (e.g., head, neck), joints. In the example above, a prefix of “c_” indicates a non-split shape, in which both the left and right sides of the head are combined, while prefixes of “l_” and “r_” indicate a split shape on the left and right sides of the head respectively.

FIGS. 13 and 14 illustrate examples of morph targets, l_EC and r_EC respectively. The l_EC morph target shows the activation of an action unit to close the left (for the character) eye. The r_EC morph target shows the activation of the corresponding action unit for the right eye. Many morph targets have separate, but corresponding actions for the left and right side of the face.

FIGS. 15 and 16 illustrate more examples of morph targets. Some morph targets have a primary effect, but may also have a secondary effect on other parts of the face as well. For example, morph targets l_LCP and r_LCP represent the left and right lip corner pullers respectively, which raise the corners of the left/right side of the mouth upwards in a smile, and squints the eyes.

FIGS. 17 and 18 illustrate examples of morph targets c_JD and c_PK. The c_JD morph target shows the activation of an action unit to drop the jaw open. The c_PK morph target shows the activation of an action unit to pucker. Unlike the previous examples, these morph targets are not isolated to a single side of the face or direction, but rather affect both sides of the face. A prefix of “x2_” indicates a two-way corrective (or “x3” for a three-way corrective), which combines two base shapes. In certain embodiments, correctives are standardized within a file format, such that each file within the format contains the same base shapes and corrective combinations. In many embodiments, affix-based corrective schemes can be generalized to be fully user-defined across multiple pre-defined base shapes (e.g., shapes following the FACS convention). For example, if the user creates x3_LCP_JD_FN, the system can automatically detect that it is a 3-way corrective between these three base shapes (LCP, JD and FN). Standardized file formats in accordance with many embodiments of the invention can be extended (e.g., by third parties) to include additional correctives for which weights can be generated based on images of a user's face.

Correctives are often useful for correcting certain rendering errors that can arise when multiple action units are activated together. In some embodiments, the two base shapes are dialed to 100%, then any correction delta is saved as an “x2_”. To control the value of an “x2” is multiplicative math. For example, x2_c_JD×PK is Jaw Drop combined with Pucker. If you set JD to 1 (1 being 100%) and PK to 1, the value of x2_c_JD×PK=1. If JD=0.5 and PK=0.5, x2_c_JD_PK=0.25. In some embodiments, include suffixes for geometry and joints. FIG. 19 illustrates an example of a corrective x2_c_JD×PK. Corrective x2_c_JD×PK corrects the pucker when the mouth is open. Other examples of morph targets are included in an appendix filed with this application.

The modified file formats allow users of this mapper framework to design a wide range of characters—human and non-human—with the set of morph targets in the modified file format. The key flexibility is for the designers of the characters to sculpt both base shapes and corrective shapes which gives them complete control over the design and movement of the character. Mappers in accordance with a number of embodiments of the invention automatically generate weights for the morph targets that match a user's facial movements and expressions. The use of such a format can allow designers to not only have increased control over the expressions of their characters, but to be able to use their characters with any mapper in accordance with such embodiments of the invention. In such a manner, the mappers can function across a wide range of characters (i.e., any character that is built with the modified specification).

Although, a particular model file format in accordance with the illustrated embodiment is described above, other model file formats in accordance with other embodiments of the invention may be used depending on the requirements of the particular system implemented. For example, although particular affixes are described, different affixes can be used in accordance with various embodiments of the invention.

Although the present invention has been described in certain specific aspects, many additional modifications and variations would be apparent to those skilled in the art. It is therefore to be understood that the present invention may be practiced otherwise than specifically described, including any variety of models of and machine learning techniques to animate the 3D shape of human faces to mimic facial movements, without departing from the scope and spirit of the present invention. Thus, embodiments of the present invention should be considered in all respects as illustrative and not restrictive.