CN116917947A - Pixel-aligned volumetric avatar - Google Patents

Abstract

A method of forming a pixel-aligned volumetric avatar includes receiving a plurality of two-dimensional images having at least two or more fields of view of a subject. The method further includes extracting a plurality of image features from the plurality of two-dimensional images using the set of learnable weights; projecting the plurality of image features along a direction between the three-dimensional model of the subject and the selected viewpoint of the viewer; and providing an image of the three-dimensional model of the subject to the viewer. A system and a non-transitory computer readable medium storing instructions to perform the above-described method are also provided.

Description

Pixel-aligned volumetric avatar

Cross Reference to Related Applications

The present disclosure relates TO and is in accordance with 35 U.S. c. ≡119 (e) claims priority from U.S. provisional application No. 63/129,989, entitled l armgardi et al, entitled l earning TO PREDICT IMPLICIT VOLUMETRIC AVATARS, filed 12/23, 2020, the contents of which are hereby incorporated by reference in their entirety for all purposes.

Technical Field

The present disclosure relates to generating faithful facial expressions in Virtual Reality (VR) and augmented reality (augmented reality, AR) applications for generating real-time volumetric avatars. More specifically, the present disclosure provides real-time volumetric avatars in multi-identity settings for VR/AR applications.

Background

In the VR/AR application field, acquiring and rendering a realistic human head is a challenging problem to implement virtual telepresence. Currently, the highest quality is achieved by volumetric methods that train in a person-specific manner on multi-view data. These models represent fine structures, such as hair, better than simpler mesh-based models. The volumetric model typically employs global codes to represent facial expressions so that the volumetric model can be driven by a small set of animation parameters. While such architectures achieve impressive rendering quality, they cannot be easily extended to multi-identity settings, and they are computationally expensive and difficult to practice in "real-time" applications.

Disclosure of Invention

In a first aspect, there is provided a computer-implemented method comprising: receiving a plurality of two-dimensional images, the plurality of two-dimensional images having at least two or more fields of view of a subject; extracting a plurality of image features from the plurality of two-dimensional images using a set of learnable weights; projecting the plurality of image features along a direction between the three-dimensional model of the subject and the selected viewpoint of the viewer; and providing an image of the three-dimensional model of the subject to the viewer.

Extracting the plurality of image features may include: the intrinsic characteristics of the camera used to collect the plurality of two-dimensional images are extracted.

Projecting the plurality of image features along a direction between the three-dimensional model of the subject and the selected viewpoint of the viewer may include: the feature map associated with the first direction and the feature map associated with the second direction are interpolated.

Projecting the plurality of image features along a direction between the three-dimensional model of the subject and the selected viewpoint may include: a plurality of image features are aggregated for a plurality of pixels along a direction between the three-dimensional model of the subject and the selected viewpoint.

Projecting the plurality of image features along a direction between the three-dimensional model of the subject and the selected viewpoint may include: the plurality of feature maps generated by each of the plurality of cameras, each of which has inherent characteristics, are concatenated in a permutation-invariant combination.

The method may further comprise: the method further includes evaluating a loss function based on a difference between an image of the three-dimensional model of the subject and a ground truth image of the subject, and updating at least one of the set of learnable weights based on the loss function.

The principal may be a user of a client device having a webcam directed to the user. The method may further include identifying the selected viewpoint as a location of a webcam pointing from the client device to the user.

The viewer may be using a network-coupled client device, and providing an image of the three-dimensional model of the subject may include: streaming the following video to the network-coupled client device: the video has a plurality of images of a three-dimensional model of the subject.

The subject may be a user of a client device having an immersive reality application running therein, and the method may further comprise: the selected viewpoint is identified as the location within the immersive reality application where the viewer is located.

In a second aspect, there is provided a system comprising: a memory storing a plurality of instructions; and one or more processors configured to execute the plurality of instructions to cause the system to perform the method of the first aspect.

A computer program product is also described, comprising instructions which, when the program is executed by a computer, cause the computer to perform the method of the first aspect.

A computer-readable storage medium is also described, comprising instructions which, when executed by a computer, cause the computer to perform the method of the first aspect. The medium may be non-transitory.

In a third aspect, there is provided a computer-implemented method for training a model to provide a view of a subject to an autostereoscopic display in a virtual reality head mounted viewer, the method comprising: collecting a plurality of ground truth images from the faces of a plurality of users; correcting the plurality of ground truth images with a plurality of stored calibrated stereo image pairs; generating a plurality of composite views of a plurality of subjects with a three-dimensional facial model, wherein the plurality of composite views of the plurality of subjects include interpolation of a plurality of feature maps projected along different directions corresponding to the plurality of views of the plurality of subjects; and training the three-dimensional face model based on differences between the plurality of ground truth images and the plurality of synthesized views of the plurality of subjects.

Generating the plurality of composite views may include: image features from each of a plurality of ground truth images are projected along a selected viewing direction, and a plurality of feature maps generated from each of the plurality of ground truth images, each of the plurality of ground truth images having inherent characteristics, are concatenated in a permutation-invariant combination.

Training the three-dimensional face model may include: at least one of the learner weights for each of the plurality of features in the plurality of feature maps is updated based on a value of a loss function indicative of a difference between the plurality of ground truth images and the plurality of composite views of the plurality of subjects.

Training the three-dimensional face model may include: the background value for each of the plurality of pixels in the plurality of ground truth images is trained based on the pixel background values projected from the plurality of ground truth images.

The method may further include interpolating the plurality of feature maps by: the plurality of feature vectors from the plurality of cameras are averaged to form a camera aggregate feature vector of different directions at the desired point.

Training the three-dimensional face model may include: a background model is generated using specific features of each of a plurality of cameras used to collect the plurality of ground truth images.

A system is also described, comprising: a memory storing a plurality of instructions; and one or more processors configured to execute the plurality of instructions to cause the system to perform the method of the third aspect.

A computer program product is also described, comprising instructions which, when the program is executed by a computer, cause the computer to perform the method of the third aspect.

A computer-readable storage medium is also described, comprising instructions which, when executed by a computer, cause the computer to perform the method of the third aspect. The medium may be non-transitory.

Drawings

FIG. 1 illustrates an example architecture suitable for providing real-time, clothing-worn subject animation in a virtual reality environment.

Fig. 2 is a block diagram illustrating an example server and client from the architecture of fig. 1.

Fig. 3 shows a block diagram of the following model architecture: the model architecture is used for 3D rendering of a portion of the face of a VR/AR head-mounted viewer (head set) user.

Fig. 4A to 4C show a volumetric avatar calculated with only two views given as inputs.

Fig. 5 shows a different technique compared to ground truth identity: techniques for reality capture, neural volume, globally regulated neural radiation field (Neural Radiance Field, neRF), and pixel alignment.

FIG. 6 shows alpha/normal/avatar generated at a typical viewpoint using eNerf and pixel alignment compared to ground truth.

Fig. 7 shows the predicted texture with respect to the number of views.

Fig. 8 shows the background ablation results.

FIG. 9 shows the sensitivity of the pixel aligned features to the choice of feature extractor used, and the shallow convolutional network.

Fig. 10 illustrates a camera perception feature summarization strategy.

Fig. 11 shows a flow chart of the method as follows: the method is for rendering a three-dimensional (3D) view of a portion of a user's face from a plurality of 2D images of the portion of the user's face.

Fig. 12 shows a flow chart of the method as follows: the method is for training a model to render a three-dimensional (3D) view of a portion of a user's face from a plurality of two-dimensional (2D) images of the portion of the user's face.

Fig. 13 illustrates a computer system configured to perform at least some of the plurality of methods for using an AR or VR device.

In the drawings, like elements are similarly numbered according to their description unless explicitly stated otherwise.

Detailed Description

In the following detailed description, numerous specific details are set forth in order to provide a thorough understanding of the present disclosure. It will be apparent, however, to one of ordinary skill in the art that the embodiments of the present disclosure may be practiced without some of these specific details. In other instances, well-known structures and techniques have not been shown in detail in order not to obscure the disclosure.

General overview

Virtual telepresence applications attempt to represent human heads with high precision and fidelity. Due to the complex geometric and cosmetic characteristics of the human head: subsurface scattering of skin, fine surface detail, thin structured hair (thin-structured hair), and human eyes and teeth are all specular and translucent, and thus modeling and rendering human heads is challenging. Existing methods include complex and expensive multi-view capture devices (with up to hundreds of cameras) to reconstruct even individual-specific models of human heads. Currently, high quality methods employ volumetric models rather than texture meshes, because volumetric models can better learn to represent fine structures (e.g., hair) on the face, which is critical to achieving a realistic appearance. Volumetric models typically employ global codes to represent facial expressions or are applicable only to static scenes. While such an architecture achieves an impressive rendering quality, it is difficult for such an architecture to accommodate multi-identity settings. Such as global codes for controlling expressions, are insufficient to model identity changes across subjects. Some attempts to solve this problem include implicit models for representing scenes and objects. These models have the following advantages: the scene is represented as a parametric function in continuous space, which allows fine-grained inference of geometry and texture. However, these methods fail to model view-dependent effects presented, for example, with hair having a textured surface. This approach may generalize across objects, but only at low resolution, and may only handle pure lambertian surfaces (Lambertian surface), which is insufficient for human heads. One disadvantage of the above methods is that they are trained to model only a single scene or object. Methods that can generate multiple objects are often limited in terms of quality and resolution of the predicted texture and geometry. For example, methods that generate a set of weights (e.g., scene representation network (Scene Representation Network, SRN)) from global image coding (e.g., single potential code vector per image) are difficult to generalize to local variations (e.g., facial expressions), and these methods fail to recover high frequency detail even when these local variations are visible in the input image. This is because the global potential code aggregates the information in the image and some of the information must be discarded to generate a compact encoding of the data.

To address the above-described problems in the field of immersive reality applications of computer networks, embodiments disclosed herein implement a predicted volumetric avatar of a human head from a limited number of inputs. To achieve this, embodiments disclosed herein implement model generalization across multiple identities by parameterizing as follows: the parameterization combines the neural radiation field with local, pixel-aligned features extracted directly from the model input. This approach results in a shallow and simple network that can be implemented in real-time immersive applications. In some embodiments, models trained on photometric re-rendering loss functions may render a subject-based avatar in real-time without explicit 3D supervision. The model disclosed herein generates faithful facial expressions in multi-identity settings and is therefore suitable for real-time group immersive applications. Embodiments disclosed herein generalize in real time to multiple unseen identities and expressions and provide a good representation of a temporal image sequence.

Some embodiments include a pixel-aligned volumetric avatar (PVA) model for estimating a volumetric 3D avatar using only a few input images of the human head. The PVA model can be generalized in real time to an unseen identity. To enhance generalization across identities, the PVA model parameterizes the volumetric model via local, pixel-aligned features extracted from the input image. Thus, the PVA model may synthesize new views for unseen identities and expressions while preserving high frequency details in the rendered avatar. Furthermore, some embodiments include a pixel-aligned radiation field that predicts an implicit shape and appearance from a sparse set of pose images in any view direction for any point in space.

Example System architecture

FIG. 1 illustrates an example architecture 100 suitable for accessing a volumetric avatar engine. Architecture 100 includes a plurality of servers 130 communicatively coupled to a plurality of client devices 110 and at least one database 152 via a network 150. One of the plurality of servers 130 is configured to host memory that includes instructions that, when executed by a processor, cause the server 130 to perform at least some of the plurality of steps of the methods disclosed herein. In some examples, the processor is configured to control a graphical user interface (graphical user interface, GUI) to cause a user of one of the plurality of client devices 110 to access the volumetric avatar engine with the immersive reality application. Accordingly, the processor may include a dashboard tool configured to display components and graphical results to a user via the GUI. For load balancing purposes, the plurality of servers 130 may host a plurality of memories including instructions to the one or more processors, and the plurality of servers 130 may host a history log and a database 152, the database 152 including a plurality of training profiles for the volumetric avatar engine. Further, in some examples, multiple users of multiple client devices 110 may access the same volumetric avatar engine to run one or more immersive reality applications. In some examples, a single user with a single client device 110 may provide images and data to train one or more machine learning models running in parallel in one or more servers 130. Accordingly, client device 110 and server 130 may communicate with each other via network 150 and resources located therein (e.g., data in database 152).

Server 130 may include any of the following: the any device has a suitable processor, memory, and communication capability for hosting the volumetric avatar engine (including a plurality of tools associated therewith). The volumetric avatar engine may be accessed by various clients 110 over the network 150. The client 110 may be, for example, a desktop computer, a mobile computer, a tablet computer (e.g., including an electronic book reader), a mobile device (e.g., a smart phone or a Personal Digital Assistant (PDA)), or any other device as follows: the any other device has a suitable processor, memory, and communication capability for accessing a volumetric avatar engine located on one or more of the plurality of servers 130. In some examples, the client device 110 may include a VR/AR headset configured to run an immersive reality application using a volumetric avatar supported by one or more of the plurality of servers 130. The network 150 may include, for example, any one or more of a Local Area Network (LAN), a Wide Area Network (WAN), the internet, and the like. Further, network 150 may include, but is not limited to, any one or more of the following tool topologies: including bus networks, star networks, ring networks, mesh networks, star bus networks, tree networks, hierarchical networks, etc.

Fig. 2 is a block diagram 200 illustrating an example server 130 and client device 110 from architecture 100. Client device 110 and server 130 are communicatively coupled via respective communication modules 218-1 and 218-2 (hereinafter collectively referred to as "communication modules 218") through network 150. The communication module 218 is configured to connect with the network 150 to send information (e.g., data, requests, responses, and commands) and receive information (e.g., data, requests, responses, and commands) to other devices via the network 150. The communication module 218 may be, for example, a modem or an ethernet card, and may include radio hardware and software for wireless communication, for example, via electromagnetic radiation such as Radio Frequency (RF), near field communication (near field communication, NFC), wiFi, and bluetooth radio technologies. A user may interact with client device 110 via input device 214 and output device 216. Input devices 214 may include a mouse, keyboard, pointer, touch screen, microphone, joystick, virtual joystick, or the like. In some examples, the input device 214 may include a camera, microphone and sensors (e.g., touch sensors, acoustic sensors, inertial motion units (inertial motion unit, IMU), and other sensors configured to provide input data to the VR/AR headset. For example, in some examples, the input device 214 may include an eye tracking device to detect a position of a pupil of the user in the VR/AR headset. The output device 216 may be a screen display, touch screen, speaker, etc. Client device 110 may include memory 220-1 and processor 212-1. Memory 220-1 may include an application 222 and a GUI 225, the application 222 and the GUI 225 configured to run in client device 110 and to couple with input device 214 and output device 216. The application 222 may be downloaded by a user from the server 130 and may be hosted by the server 130. In some examples, the client device 110 is a VR/AR headset and the application 222 is an immersive reality application.

The server 130 includes a memory 220-2, a processor 212-2, and a communication module 218-2. Hereinafter, the processors 212-1 and 212-2 and the memories 220-1 and 220-2 will be collectively referred to as "processor 212" and "memory 220", respectively. The processor 212 is configured to execute instructions stored in the memory 220. In some examples, memory 220-2 includes a volumetric avatar engine 232. The volumetric avatar model engine 232 may share or provide features and resources to the GUI 225, including multiple tools associated with training and using a three-dimensional avatar rendering model for an immersive reality application (e.g., application 222). The user may access the volumetric avatar engine 232 through the application 222 installed in the memory 220-1 of the client device 110. Accordingly, the application 222 (including the GUI 225) may be installed by the server 130 and execute scripts and other routines provided by the server 130 through any of a number of tools. Execution of the application 222 may be controlled by the processor 212-1.

In this regard, as disclosed herein, the volumetric avatar engine 232 may be configured to create, store, update, and maintain PVA models 240.PVA phantom 240 may include encoder-decoder tool 242, ray traveling tool 244, and radiation field tool 246. The encoder-decoder tool 242 collects multiple input images with multiple simultaneous views of the subject and extracts pixel-aligned features to adjust the radiation field tool 246 via a ray-traveling program in the ray-traveling tool 244. PVA model 240 may generate a new view of the unseen subject from one or more sample images processed by encoder-decoder tool 242. In some examples, encoder-decoder tool 242 is a shallow (e.g., comprising several single-node layers or two-node layers) convolutional network. In some examples, the radiation field tool 246 converts three-dimensional position and pixel-aligned features into color and opacity fields that can be projected in any desired view direction.

In some examples, volumetric avatar engine 232 may access one or more machine learning models stored in training database 252. Training database 252 includes training files and other data files that may be used by volumetric avatar model engine 232 in training of a machine learning model based on user input through application 222. Further, in some examples, at least one or more training files or machine learning models may be stored in any one of the plurality of memories 220 and accessed by a user through the application 222.

The volumetric avatar engine 232 may include the following algorithm: the algorithm is trained for the specific purpose of the engines and tools included in the volumetric avatar engine 232. The algorithm may include a machine learning algorithm or an artificial intelligence algorithm using any linear or nonlinear algorithm, such as a neural network algorithm or a multiple regression algorithm. In some examples, the machine learning model may include a Neural Network (NN), a convolutional neural network (convolutional neural network, CNN), a generative antagonism neural network (generative adversarial neural network, GAN), a deep reinforcement learning (deep reinforcement learning, DRL) algorithm, a deep-loop neural network (deep recurrent neural network, DRNN), a classical machine learning algorithm (e.g., a random forest, a k-nearest neighbor (KNN) algorithm, a k-means clustering algorithm), or any combination thereof. More generally, the machine learning model may include any machine learning model that involves a training step and an optimization step. In some examples, training database 252 may include a training archive to modify coefficients according to desired results of the machine learning model. Thus, in some examples, volumetric avatar model engine 232 is configured to access training database 252 to retrieve documents and archives as input to the machine learning model. In some examples, volumetric avatar engine 232, tools contained in volumetric avatar engine 232, and at least a portion of training database 252 may be hosted in different servers accessible by server 130 or client device 110.

Fig. 3 shows a block diagram of a model architecture 300 for 3D rendering of face portions of VR/AR headset users. Model architecture 300 is a pixel-aligned volumetric avatar (PVA) model. PVA model 300 is learned from a set of multi-view images that produce a plurality of 2D input images 301-1, 301-2, and 301-n (hereinafter, collectively referred to as "input images 301"). Each of the plurality of input images 301 is associated with a camera view vector v _i (e.g., v ₁ 、v ₂ And v _n ) In association, the camera view vector indicates a view direction of the user's face for the particular image. In some examples, multiple input images 301 are collected simultaneously or quasi-simultaneously such that different view vectors v _i Pointing to the same volumetric representation of the body. A plurality of vectors v _i Each vector in (a) is a known viewpoint 311, the known viewpoint 311 and the camera intrinsic parameter K _i And rotating R _i (e.g., { K _i ,[R|t] _i }) are associated. Intrinsic parameters K of camera _i Brightness, color mapping, sensor efficiency, and other camera related parameters may be included. Rotation R _i Indicating the orientation (and distance) of the head of the subject relative to the camera. Different camera sensors have slightly different responses to the same incident radiation, despite the fact that these different camera sensors are the same camera model. If no measures are taken to solve this problem, the intensity difference will eventually be incorporated into the scene representation N, which will lead to an unnatural brightening or darkening of the image from some point of view. To solve this problem, we learn per-camera bias (per-camera bias) and gain values. This allows the system to interpret this change in data in a 'easier' way.

The value of 'n' is purely exemplary, as any number n of input images 301 may be used, as will be appreciated by any of ordinary skill. PVA model 300 produces volumetric rendering 321 of the head-mounted viewer user. The volume rendering 321 is a 3D model (e.g., "avatar") as follows: the 3D model may be used to generate a 2D image of the subject from the target viewpoint. The 2D image changes as the target viewpoint changes (e.g., as the viewer moves around the head-mounted viewer user).

PVA phantom 300 includes a convolutional encoder-decoder 310A, a ray-traveling stage 310B, and a radiation field stage 310C (hereinafter collectively referred to as "PVA stage 310"). PVA model 300 is trained using gradient descent with a plurality of input images 301 selected from a multi-identity training corpus. Thus, PVA model 300 includes a loss function defined between a plurality of predicted images from a plurality of subjects and corresponding ground truth values. This enables PVA model 300 to render accurate volume rendering 321 independent of the subject.

Convolutional encoder-decoder network 310A takes a plurality of input images 301 and generates pixel-aligned feature maps 303-1, 303-2, and 303-n (hereinafter collectively referred to as "feature map 303" f ⁽ⁱ⁾ ). Ray-progression stage 310B follows the path defined by { K } _j ,[R|t] _j The ray in the defined object view j follows each of the plurality of pixels, thereby atThe color C and optical density ("opacity") produced by the radiation field stage 310C are accumulated at each point. The radiation field stage 310C (N) converts the 3D position and pixel aligned features to color and opacity to render the radiation field 315 (C, σ).

The input image 301 is a 3D object as follows: the 3D object has a direction v with the camera _i The height (h) and width (w) corresponding to the collected 2D image, and the depth of 3 layers for each color pixel R, G, B. The feature map 303 is a 3D object having dimensions h×w×d. The encoder-decoder network 310A encodes the input image 301 using the learnable weights 320-1, 320-2, … …, 320-n (hereinafter collectively referred to as "learnable weights 320"). The ray-traveling stage 310B performs world-to-camera projection 323, bilinear interpolation 325, position encoding 327, and feature aggregation 329.

In some examples, for the adjustment view (conditioning view) v _i ∈R ^h×w×3 The feature map 303 may be defined as a function of:

wherein phi (X): r is R ³ →R ^6×l Is to use 2X l different basis functions to point 330 (X ε R ³ ) And performing position coding. The point 330 (X) is directed along the 2D image from the subject to a specific viewpoint 331r ₀ Is a ray of the ray of (c). Feature map 303 (f ⁽ⁱ⁾ ∈R ^h×w×d ) And camera position vector v _i Associated, where d is the number of feature channels, h and w are the image height and width, and f _X ∈R ^d ' is the aggregate image feature associated with point X. For each feature map f ⁽ⁱ⁾ The ray-progression stage 310B obtains f by projecting the 3D point X along the ray using the camera-intrinsic parameter (K) and extrinsic parameter (R, t) of the particular viewpoint _X ∈R ^d ，

x _i ＝∏(X；K _i [R|t] _i ) (3)

Where n is the perspective projection function to the camera pixel coordinates and F (F, x) is the bilinear interpolation 325 of F at pixel location x. The ray-traveling stage 310B aligns the pixel-aligned features f from multiple images ⁽ⁱ⁾ _X And combined for the radiation field level 310C.

For the camera with internal reference K _j And rotating R _j And translation t _j Each given training image v of (2) _j For a given viewpoint, at the focal plane and center 331 (r ₀ )∈R ³ Pixel p e R of (2) ² Is obtained by advancing rays into the scene using the following matrix: camera-to-world projection matrix P ^-1 ＝[R _i |t _i ] ^-1 K ^-1 _i Wherein the direction of the rays is given by the following formula.

Ray-progression stage 310B follows a path defined by r (t) =r ₀ +td (where t ε [ t ] _near ,t _far ]) The defined rays 335 accumulate the radiation values and opacity values as follows:

wherein, the liquid crystal display device comprises a liquid crystal display device,

in some examples, ray-progression stage 310B pairs a set of n _s Points t to t _Near-to-near ,t _{Far distance} ]The sampling is performed uniformly. Setting x=r (t), the integration rule can be used to approximate the integrals 6 and 7. Function I _α (p) can be defined as:

wherein alpha is _i ＝1-exp(-δ _i ·σ _i ) Wherein delta _i Is the distance between the (i+1) th sample point and the (i) th sample point along ray 335.

At a view point v with a plurality of known cameras _i And a fixed number of adjusted views, the ray-progression stage 310B aggregates multiple features by simple concatenation. Specifically, for the case with the composition represented by { R } _i } ⁿ _i＝1 And { ti } ni=1, and n adjusted images { v) of the corresponding rotation matrix and translation matrix _i } ⁿ _i＝1 The feature { f } is used for each point X as in equation (3) ⁽ⁱ⁾ _X } ⁿ _i＝1 The ray-progression stage 310B generates the following final features:

wherein the method comprises the steps ofRepresenting a series connection along the depth dimension. This preserves { v } from multiple views _i } ⁿ _i＝1 To assist PVA model 300 in determining the best combination and using tuning information.

In some examples, PVA model 300 is insensitive to the viewpoint and the number of adjustment views. In this case, a simple concatenation as described above is not sufficient, since the number of adjustment views may not be known a priori, resulting in different feature sizes (d) during the inference time. To summarize the characteristics of the multiview settings, some examples include a permutation invariant function (permutation invariant function) G: r is R ^n×d →R ^d So that for any permutation ψ:

G(f ⁽¹⁾ ,…,f ⁽ⁿ⁾ )＝G([f ^ψ(1) ,f ^ψ(2) …,f ^ψ(n) ])

a simple permutation invariant function for feature aggregation is to average the sampled feature map 303. This aggregation process may be desirable when depth information during training is available. However, in the presence of depth blur (e.g., for points projected onto the feature map 303 prior to sampling), the above aggregation may lead to artifacts. To avoid this, some examples consider camera information to include effective adjustments in the radiated field level 310C. Thus, some examples include a regulating function network (conditioning function network) N _cf ：R ^d+7 →R ^d ' the adjustment function network obtains a feature vector f ⁽ⁱ⁾ _X And camera information (ci) and generates a camera summary feature vector f' ⁽ⁱ⁾ _X . These modified vectors are then averaged over multiple or all adjustment views as follows:

an advantage of this approach is that the camera summary feature may take into account possible occlusions before feature averaging is performed. Camera information is encoded into a 4-dimensional (4D) rotation quaternion and a 3D camera position.

Some examples may also include a context estimation network N _bg To avoid learning the portions of the background that are in the scene representation. Background estimation network N _bg Can be defined as: n (N) _bg ：R ^nc :→R ^h×w×3 To learn the fixed background per camera. In some examples, the radiation field level 310C may use N _bg To predict the final image pixel as:

I _p ＝I _rgb +(1-I _α )·I _bg (11)

wherein for camera c _i ，Wherein->Is an initial estimate of the background using repair extraction, and I _α As defined in equation (8). These prosthetic backgrounds are often noisy, resulting in a 'halo' effect around the human head. To avoid this, N _bg The model learns the residuals to the restored background. This has the following advantages: high capacity networks are not required to account for background.

For ground truth target image v _j PVA model 300 trains both radiated field level 310C and feature extraction network using simple photometric reconstruction loss as follows:

fig. 4A to 4C show volume avatars 421A-1, 421A-2, 421A-3, 421A-4, and 421A-5 (hereinafter, collectively referred to as "volume avatars 421A"), 421B-1, 421B-2, 421B-3, 421B-4, and 421B-5 (hereinafter, collectively referred to as "volume avatars 421B"), 421C-1, 421C-2, 421C-3, 421C-4, and 421C-5 (hereinafter, collectively referred to as "volume avatars 421C"). Volumetric avatars 421A, 421B, and 421C (hereinafter collectively referred to as "volumetric avatar 421") are high fidelity renderings of different objects obtained using only two body views as inputs.

Volumetric avatar 421 shows that the PVA model disclosed herein may generate multiple views of different subject avatars from a large number of new viewpoints given only two views as input.

FIG. 5 shows avatars 521A-1, 521A-2, 521A-3, 521A-4 (hereinafter, collectively referred to as "reality capturing avatars 521A"), avatars 521B-1, 521B-2, 521B-3, 521B-4 (hereinafter, collectively referred to as "nerve volume avatars 521B"), avatars 521C-1, 521C-2, 521C-3, 521C-4 (hereinafter, collectively referred to as "globally regulated, cNeRF avatars 521C"), and avatars 521D-1, 521D-2, 521D-3, 521D-4 (hereinafter, collectively referred to as "volume avatars 521D"). Avatars 521A, 521B, 521C, and 521D will be collectively referred to hereinafter as "avatars 521". An avatar 521 is obtained from the two input images of the new identity and this avatar 521 is compared with ground truth images 501-1, 501-2, 501-3 and 501-4 (hereinafter collectively referred to as "ground truth images 501") as input to the computational reconstruction.

The reality capturing avatar 521A is obtained using a structure-from-motion (SFM) and multi-view stereoo (MVS) algorithm that reconstructs a 3D model from a set of captured images. The neural volume avatar 521B is obtained using a voxel-based extrapolation method that globally encodes a dynamic image of the scene and decodes the voxel grid and warped fields representing the scene. The cNeRF avatar 521C is a variant of the NeRF algorithm implemented with global identity adjustment (cNeRF). In some examples, the cneref avatar 521C uses a Visual Geometry Group (VGG) network to extract a single 64-dimensional (64D) feature vector for each training identity and additionally adjusts the NeRF model on this input. As disclosed herein, volume avatar 521D is obtained with a PVA model.

The volumetric avatar 521D is a more complete reconstruction than the real-world captured avatar 521A, which typically uses more ground truth images 501 to obtain a good reconstruction. The volumetric avatar 521D also enables a more detailed reconstruction than cnorf avatar 521C due to the pixel alignment features in the PVA model disclosed herein, which provide more complete information to the model at test time.

Table 1 below is a comparison of the performance of the volumetric avatar 521D with the Neural Volume (NV) avatar 521B and cNeRF avatar 521C using different metrics. The structural similarity index (structural similarity index, SSIM) preferably has a maximum value of one (1), and the learning-aware image block similarity (learned perceptual image patch similarity, lpas) metric and the mean square error (mean squared error, MSE) metric preferably have lower values.

TABLE 1

	SSIM	NSE	LPIPS
				cNeRF	0.7663	1611.01 12	4.3775
NV	0.8027	1208.36	3.1112
				PVA	0.8889	383.71	1.7392

The PVA model disclosed herein addresses certain shortcomings of global identity coding methods (e.g., nerve volume and cnenf) that train in a scenario-specific manner, which do not generalize well to unseen identities. For example, cNeRF avatar 521C smoothes the facial features and loses some local detail of the unseen identity (e.g., facial hair in 521C-3 and 521C-4, and hair length in 521C-2) because the model relies heavily on a global prior of learning. Since there is no a priori model built into the sfm+mvs framework, the reality capturing avatar 521A fails to capture the head structure, resulting in incomplete reconstruction. For a real-world capture (RC) 521A model, a large number of images would be required in order to faithfully reconstruct a new identity. The neural volume avatar 521B generates a better texture because the generated warp field takes into account some degree of local information. However, the neural volume avatar 521B uses an encoder configured with a global encoding and projecting the test time identity into the most recent training time identity, resulting in inaccurate avatar predictions. The volumetric avatar 521D reconstructs the structure of the volumetric head along with the hair from only two example viewpoints.

FIG. 6 shows alpha views 631A-1, 631A-2, and 631A-3 (hereinafter collectively referred to as "alpha views 631A"), normal views 633A-1, 633A-2, and 633A-3 (hereinafter collectively referred to as "normal views 633A") and avatar views 621A-1, 621A-2, and 621A-3 (hereinafter collectively referred to as "avatar 621A") generated using eNerf, and associated ground truth images 601-1, 601-2, and 601-3 for three different subjects. As disclosed herein, alpha views 631B-1, 631B-2, and 631B-3 (hereinafter, collectively referred to as "alpha views 631B"), normal views 633B-1, 633B-2, and 633B-3 (hereinafter, collectively referred to as "normal views 633B"), and avatar views 621B-1, 621B-2, and 621B-3 (hereinafter, collectively referred to as "avatar 621B") of an avatar are aligned with pixels obtained with a PVA model.

Compared to avatar 621A, avatar 621B is well suited to capturing expression information, and avatar 621A has more difficulty generalizing facial expressions to new identities. To obtain 621A avatar, a NeRF model (ennerf) is adjusted according to one-hot expression codes and one-hot identity information about the test time identity. However, although all identities have been seen during training, avatar 621B generalizes better to dynamic expressions of multiple identities than avatar 621A. Because the PVA model is tuned with local features, avatar 621B captures dynamic effects (both geometry and texture) on a particular identity better than avatar 621A.

FIG. 7 shows predicted avatars 721-1, 721-2, 721-3, 721-4, and 725-5 (hereinafter, collectively referred to as "avatars 721") with respect to the number of views (rows). Normal views 733-1, 733-2, 733-3, 733-4, and 733-5 (hereinafter, collectively referred to as "normal views 733") are associated with each of the plurality of avatars 721.

The avatar 721 and the normal view 733 show different perspectives of the subject from the perspectives captured in the images 701-1, 701-2, 701-3, 701-4, and 701-5 (hereinafter, collectively referred to as "ground truth images 701"). Because the PVA model learns shape priors from training identities, the normal view 733 is consistent with the identity of the ground truth image 701. However, when extrapolated to extreme perspectives (733-1 and 721-1), artifacts appear in the portion of the face that was not seen in the "adjusted" ground truth image 701. This is due to the inherent depth blur caused by the projection of the sample points onto the ground truth image 701-1. Adding the second view (e.g., ground truth image 701-2) has significantly reduced these artifacts in the normal view 733-2 and avatar 721-2 because the PVA model now has more information about features from different perspectives and thus depth information. In general, the PVA models disclosed herein can achieve a large degree of view extrapolation with only two adjustment views.

FIG. 8 shows background ablation results of avatars 821-1 and 821-2 (hereinafter, collectively referred to as "avatars 821") derived from normal views 833-1 and 833-2 (hereinafter, collectively referred to as "normal views 833") based on input images 801-1 and 801-2 (hereinafter, collectively referred to as "input images 801").

Fig. 9 shows the sensitivity of the PVA model to the feature extractors 921A ("hourglass network"), 921B ("UNet") and 921C (shallow convolutional network) used for selection based on the adjustment view 901. 921A and 921B are reliable feature extractors. In some examples, a shallow encoder-decoder architecture 921C may be required to perform because it retains more local information without having to encode all pixel level information into the bottleneck layer.

Fig. 10 illustrates a camera perception feature summarization strategy. For the first subject, input images 1001A-1, 1001B-1, and 1001C-1 (hereinafter, collectively referred to as "input images 1001-1") corresponding to different views and collected with different cameras are averaged without camera-specific information (avatar 1021A-1), or averaged with camera-specific information (avatar 1021B-1). Similarly, for the second subject, input images 1001A-2, 1001B-2, and 1001C-2 (hereinafter, collectively referred to as "input images 1001-2") corresponding to different views and cameras are averaged without camera-specific information (avatar 1021A-2), or averaged with camera-specific information (avatar 1021B-2). And for the third subject, input images 1001A-3, 1001B-3, and 1001C-3 (hereinafter, collectively referred to as "input images 1001-3") corresponding to different views and cameras are averaged without camera-specific information (see camera-intrinsic parameter (K) and extrinsic parameter (R, t) in equation (3)), or averaged with camera-specific information (avatar 1021B-3).

Specifically, without camera information, avatars 1021A-1, 1021A-2, and 1021A-3 present fringes in the generated images (particularly in avatars 1021A-1 and 1021A-2) due to inconsistent averages of information from different viewpoints.

Fig. 11 shows a flow chart in a method 1100 for rendering a three-dimensional (3D) view of a portion of a user's face from a plurality of two-dimensional (2D) images of the portion of the user's face. The steps in method 1100 may be performed at least in part by a processor executing instructions stored in a memory, where the processor and memory are part of a client device or VR/AR headset disclosed herein (e.g., memory 220, processor 212, client device 110). In other examples, at least one or more of the steps of a method consistent with method 1100 may be performed by a processor executing instructions stored in a memory, wherein at least one of the processor and the memory is remotely located in a cloud server, and the head-mounted viewer device is communicatively coupled to the cloud server via a communication module coupled to a network (see server 130, communication module 218). In some examples, the method 1100 may be performed using a volumetric avatar engine configured to train a PVA model including an encoder-decoder tool, a ray travel tool, and a radiation field tool, including neural network architecture in a machine learning algorithm or an artificial intelligence algorithm, as disclosed herein (e.g., volumetric avatar engine 232, PVA model 240, encoder-decoder tool 242, ray travel tool 244, and radiation field tool 246). In some examples, a method consistent with the present disclosure may include at least one or more steps from method 1100 as follows: at least one or more steps from method 1100 are performed in a different order, simultaneously, quasi-simultaneously, or overlapping in time.

Step 1102 includes: a plurality of two-dimensional images is received, the plurality of two-dimensional images having at least two or more fields of view of a subject.

Step 1104 includes: a plurality of image features are extracted from a plurality of two-dimensional images using a set of learnable weights. In some examples, step 1104 includes: the intrinsic characteristics of a camera for collecting a plurality of two-dimensional images are extracted.

Step 1106 includes: a plurality of image features are projected along a direction between the three-dimensional model of the subject and the selected viewpoint of the viewer. In some examples, step 1106 includes: the feature map associated with the first direction and the feature map associated with the second direction are interpolated. In some examples, step 1106 includes: a plurality of image features for a plurality of pixels are aggregated along a direction between the three-dimensional model of the subject and the selected viewpoint. In some examples, step 1106 includes: the plurality of feature maps generated by each of the plurality of cameras, each of which has inherent characteristics, are concatenated in a permutation-invariant combination.

In some examples, the principal is a user of a client device having a webcam pointing to the user, and step 1106 includes: the selected viewpoint is identified as the location of the webcam pointing from the client device to the user. In some examples, the subject is a user of a client device having an immersive reality application running therein, and step 1106 further comprises: the selected viewpoint is identified as the location within the immersive reality application where the viewer is located.

Step 1108 includes: an image of the three-dimensional model of the subject is provided to a viewer. In some examples, step 1108 includes: the method further includes evaluating a loss function based on a difference between an image of the three-dimensional model of the subject and a ground truth image of the subject, and updating at least one of the set of learnable weights based on the loss function. In some examples, the viewer is using a network-coupled client device, and step 1108 includes: video of a plurality of images having a three-dimensional model of a subject is streamed to a network-coupled client device.

Fig. 12 shows a flow chart in a method 1200, the method 1200 for training a model to render a three-dimensional (3D) view of a portion of a user's face from a plurality of two-dimensional (2D) images of the portion of the user's face. The steps in method 1200 may be performed, at least in part, by a processor executing instructions stored in a memory, where the processor and memory are part of a client device or VR/AR headset disclosed herein (e.g., memory 220, processor 212, client device 110). In other examples, at least one or more of the steps of a method consistent with method 1200 may be performed by a processor executing instructions stored in a memory, wherein at least one of the processor and the memory is remotely located in a cloud server, and the head-mounted viewer device is communicatively coupled to the cloud server via a communication module coupled to a network (see server 130, communication module 218). In some examples, the method 1200 may be performed using a volumetric avatar engine configured to train a PVA model including an encoder-decoder tool, a ray travel tool, and a radiation field tool, including neural network architecture in a machine learning algorithm or an artificial intelligence algorithm, as disclosed herein (e.g., volumetric avatar engine 232, PVA model 240, encoder-decoder tool 242, ray travel tool 244, and radiation field tool 246). In some examples, a method consistent with the present disclosure may include at least one or more steps from method 1200 as follows: at least one or more steps from method 1200 are performed in a different order, simultaneously, quasi-simultaneously, or overlapping in time.

Step 1202 includes: a plurality of ground truth images are collected from the faces of a plurality of users.

Step 1204 includes: the plurality of ground truth images are modified with a plurality of stored calibrated stereo image pairs.

Step 1206 includes: a plurality of composite views of the plurality of subjects are generated with the three-dimensional facial model, wherein the plurality of composite views of the plurality of subjects include an interpolation of a plurality of feature maps projected along different directions corresponding to the plurality of views of the plurality of subjects. In some examples, step 1206 includes: image features from each of a plurality of ground truth images are projected along a selected viewing direction and a plurality of feature maps generated from each of the plurality of ground truth images, each of the plurality of ground truth images having inherent characteristics, are concatenated in a permutation-invariant combination. In some examples, step 1206 further comprises: the plurality of feature maps are interpolated by: the plurality of feature vectors from the plurality of cameras are averaged to form a camera aggregate feature vector of different directions at the desired point.

Step 1208 includes: a three-dimensional facial model is trained based on differences between the plurality of ground truth images and the plurality of synthesized views of the plurality of subjects. In some examples, step 1208 includes: at least one of the learner weights for each of the plurality of features in the plurality of feature maps is updated based on a value of a loss function indicative of a difference between the plurality of ground truth images and the plurality of composite views of the plurality of subjects. In some examples, step 1208 includes: the background value for each of the plurality of pixels in the plurality of ground truth images is trained based on the pixel background values projected from the plurality of ground truth images. In some examples, step 1208 includes: a background model is generated using specific features of each of a plurality of cameras used to collect the plurality of ground truth images.

Hardware overview

Fig. 13 is a block diagram illustrating an exemplary computer system 1300 with which the head-mounted and other client devices 110, as well as methods 1100 and 1200, may be implemented by the computer system 1300. In some aspects, computer system 1300 may be implemented using hardware, or a combination of software and hardware, in a dedicated server, or integrated into another entity, or distributed across multiple entities. Computer system 1300 may include a desktop computer, a laptop computer, a tablet phone, a smart phone, a feature phone, a server computer, or others. The server computer may be remotely located in a data center or stored locally.

Computer system 1300 includes a bus 1308 or other communication mechanism for communicating information, and a processor 1302 (e.g., processor 212) coupled with bus 1308 for processing information. For example, computer system 1300 may be implemented with one or more processors 1302. The processor 1302 may be a general purpose microprocessor, microcontroller, digital signal processor (Digital Signal Processor, DSP), application specific integrated circuit (Application Specific Integrated Circuit, ASIC), field programmable gate array (Field Programmable Gate Array, FPGA), programmable logic device (Programmable Logic Device, PLD), controller, state machine, gating logic, discrete hardware components, or any other suitable entity that can perform the computation or other operation of information.

In addition to hardware, computer system 1300 may include code that creates an execution environment for the relevant computer program, e.g., code that constitutes the following stored in an included memory 1304 (e.g., memory 220): processor firmware, protocol stacks, a database management system, an operating system, or a combination of one or more thereof, which comprises a Memory 1304, such as random access Memory (Random Access Memory, RAM), flash Memory, read-Only Memory (ROM), programmable Read-Only Memory (Programmable Read-Only Memory, PROM), erasable programmable Read-Only Memory (EPROM), registers, hard disk, a removable disk, a CD-ROM, a DVD, or any other suitable storage device, for example, a Memory 1304 coupled with bus 1308 for storing information and instructions to be executed by processor 1302. The processor 1302 and the memory 1304 can be supplemented by, or incorporated in, special purpose logic circuitry.

The instructions may be stored in the memory 1304 and may be implemented in one or more computer program products, such as one or more modules of computer program instructions encoded on a computer readable medium for execution by the computer system 1300 or to control the operation of the computer system 1300, and according to any method known to those skilled in the art, the instructions include, but are not limited to, the following computer languages: such as data oriented languages (e.g., SQL, dBase), system languages (e.g., C, objective-C, C ++, assembly), architecture languages (e.g., java, ·net), and application languages (e.g., PHP, ruby, perl, python). The instructions may also be implemented in a computer language as follows: such as array language, aspect oriented language, assembly language, authoring language (authoring language), command line interface language, compiled language, concurrency language, curly language (curly-curly language), data streaming language, data structured language, declarative language, deep language (esoteric language), extension language (extension language), fourth generation language, functional language, interactive mode language, interpreted language, iterative language (iterative language), list-based language (list-based language), small language (littlelanguage), logic-based language, machine language, macro language, meta-programming language (metaprogramming language), multi-paradigm language (multiparadigm language), numerical analysis, non-English-based language (non-englist-based language), object-oriented class-based language, object-oriented prototype-based language, offside rule language (off-side rule language), procedural language, reflection language (reflective language), rule-based language, script language, stack-based language, synchronization language, grammar processing language (syntax handling language), visual language, and xml-based language. Memory 1304 may also be used for storing temporary variables or other intermediate information during execution of instructions to be executed by processor 1302.

The computer programs discussed herein do not necessarily correspond to files in a file system. A program may be stored in a portion of a file that holds other programs or data (e.g., one or more scripts stored in a markup language document), in a single file dedicated to the relevant program, or in multiple coordinated files (e.g., files that store one or more modules, one or more sub-programs, or portions of code). A computer program can be deployed to be executed on one computer or on multiple computers at one site or distributed across multiple sites and interconnected by a communication network. The processes and logic flows described in this specification can be performed by one or more programmable processors executing one or more computer programs to perform functions by operating on input data and generating output.

Computer system 1300 also includes a data storage device 1306 (e.g., magnetic or optical disk) coupled to bus 1308 for storing information and instructions. Computer system 1300 may be coupled to a variety of devices via input/output module 1310. The input/output module 1310 may be any input/output module. The exemplary input/output module 1310 includes a data port, such as a USB port. The input/output module 1310 is configured to be connected to a communication module 1312. Exemplary communication module 1312 includes a network interface card, such as an ethernet card and a modem. In certain aspects, the input/output module 1310 is configured to connect to a plurality of devices, such as the input device 1314 and/or the output device 1316. Exemplary input devices 1314 include a keyboard and a pointing device (e.g., a mouse or a trackball) by which a user can provide input to computer system 1300. Other kinds of input devices 1314 may also be used to provide for interaction with a user, such as a tactile input device, a visual input device, an audio input device, or a brain-computer interface device. For example, feedback provided to the user may be any form of sensory feedback, e.g., visual feedback, auditory feedback, or tactile feedback; and may receive any form of input from the user, including acoustic input, speech input, tactile input, or brain wave input. The exemplary output device 1316 includes a display device such as a liquid crystal display (liquid crystal display, LCD) monitor for displaying information to a user.

In accordance with an aspect of the disclosure, the head-mounted viewer and client device 110 may be implemented at least in part using the computer system 1300 in response to the processor 1302 executing one or more sequences of one or more instructions contained in the memory 1304. These instructions may be read into memory 1304 from another machine-readable medium, such as data storage device 1306. Execution of the sequences of instructions contained in main memory 1304 causes processor 1302 to perform the process steps described herein. One or more processors in a multi-processing arrangement may also be employed to execute the sequences of instructions contained in memory 1304. In alternative aspects, hard-wired circuitry may be used in place of or in combination with software instructions to implement the various aspects of the disclosure. Thus, aspects of the present disclosure are not limited to any specific combination of hardware circuitry and software.

Aspects of the subject matter described in this specification can be implemented in a computing system that includes a back-end component (e.g., a data server), or that includes a middleware component (e.g., an application server), or that includes a front-end component (e.g., a client computer having a graphical user interface or a web browser through which a user can interact with an implementation of the subject matter described in this specification); or aspects of the subject matter described in this specification can be implemented in any combination of one or more such back-end components, one or more such middleware components, or one or more such front-end components. The various components of the system may be interconnected by any form or medium of digital data communication (e.g., a communication network). The communication network may include, for example, any one or more of a LAN, a WAN, the Internet, and the like. Further, the communication network may include, but is not limited to, for example, any one or more of the following network topologies: including bus networks, star networks, ring networks, mesh networks, star bus networks, tree networks, hierarchical networks, etc. The communication module may be, for example, a modem or an ethernet card.

Computer system 1300 may include clients and servers. The client and server are typically remote from each other and typically interact through a communication network. The relationship of client and server arises by virtue of computer programs running on the respective computers and having a client-server relationship to each other. Computer system 1300 may be, for example, but is not limited to, a desktop computer, a laptop computer, or a tablet computer. Computer system 1300 may also be embedded in another device such as, but not limited to, a mobile phone, a PDA, a mobile audio player, a global positioning system (Global Positioning System, GPS) receiver, a video game console, and/or a television set top box.

The term "machine-readable storage medium" or "computer-readable medium" as used herein refers to any medium or media that participates in providing instructions to processor 1302 for execution. Such a medium may take many forms, including but not limited to, non-volatile media, and transmission media. Non-volatile media includes, for example, optical or magnetic disks, such as data storage device 1306. Volatile media includes dynamic memory, such as memory 1304. Transmission media includes coaxial cables, copper wire and fiber optics, including the wires that form bus 1308. Common forms of machine-readable media include, for example, a floppy disk, a flexible disk, hard disk, magnetic tape, any other magnetic medium, a CD-ROM, DVD, any other optical medium, punch cards, paper tape, any other physical medium with patterns of holes, RAM, PROM, EPROM, FLASH EPROM, any other memory chip or cartridge, or any other medium from which a computer can read. The machine-readable storage medium may be a machine-readable storage device, a machine-readable storage substrate, a memory device, a combination of substances affecting a machine-readable propagated signal, or a combination of one or more of them.

To illustrate the interchangeability of hardware and software, various illustrative blocks, modules, components, methods, operations, instructions, and algorithms have been described in general terms in terms of their functionality. Whether such functionality is implemented as hardware, software, or a combination of hardware and software depends upon the particular application and design constraints imposed on the overall system. Skilled artisans may implement the described functionality in varying ways for each particular application.

As used herein, the phrase "at least one of" after a series of items (any of which are separated by the term "and" or ") modifies the list as a whole, rather than modifying each member (e.g., each item) of the list. The phrase "at least one of" does not require that at least one item be selected; rather, the meaning of the phrase includes at least one of any of these items, and/or at least one of any combination of these items, and/or at least one of each of these items. For example, the phrase "at least one of A, B and C" or "at least one of A, B or C" each refer to: only a, only B or only C; A. any combination of B and C; and/or at least one of each of A, B and C.

The word "exemplary" is used herein to mean "serving as an example, instance, or illustration. Any embodiment described herein as "exemplary" is not necessarily to be construed as preferred or advantageous over other embodiments. Phrases such as one aspect, this aspect, another aspect, some aspects, one or more aspects, an implementation, the implementation, another implementation, some implementations, one or more implementations, an embodiment, the embodiment, another embodiment, some embodiments, one or more embodiments, a configuration, the configuration, another configuration, some configurations, one or more configurations, subject technology, the disclosure, the present disclosure, other variations thereof, and the like are for convenience and do not imply that the disclosure associated with one or more such phrases is essential to the subject technology, or that the disclosure applies to all configurations of the subject technology. The disclosure relating to one or more such phrases may apply to all configurations, or one or more configurations. A disclosure relating to one or more such phrases may provide one or more examples. A phrase such as one or more aspects may refer to one or more aspects and vice versa, and the same applies similarly to other preceding phrases.

References to elements in the singular are not intended to mean "one and only one" unless specifically stated, but rather "one or more. The term "some" refers to one or more. The underlined and/or italicized headings and subheadings are used for convenience only, do not limit the subject technology, and are not referred to in connection with the interpretation of the description of the subject technology. Relational terms such as first and second, and the like may be used to distinguish one entity or action from another entity or action without necessarily requiring or implying any actual such relationship or order between such entities or actions. All structural and functional equivalents to the elements of the various configurations described throughout this disclosure that are known or later come to be known to those of ordinary skill in the art are expressly incorporated herein by reference and are intended to be encompassed by the subject technology. Moreover, nothing disclosed herein is intended to be dedicated to the public regardless of whether such disclosure is explicitly recited in the above description. No claim element should be construed as in accordance with the specification of 35u.s.c. ≡112 paragraph 6 unless the element is explicitly recited using the phrase "means for … …" or, in the case of method claims, the element is recited using the phrase "step for … …".

While this specification contains many specifics, these should not be construed as limitations on the scope of what may be described, but rather as descriptions of specific embodiments of the subject matter. Certain features that are described in this specification in the context of separate embodiments can also be implemented in combination in a single embodiment. Conversely, various features that are described in the context of a single embodiment can also be implemented in multiple embodiments separately or in any suitable subcombination. Furthermore, although features may be described above as acting in certain combinations and even initially claimed as such, one or more features from a described combination can in some cases be excised from the combination, and the described combination may be directed to a subcombination or variation of a subcombination.

The subject matter of the present specification has been described with respect to particular aspects, but other aspects can be practiced and other aspects are within the scope of the following claims. For example, although operations are depicted in the drawings in a particular order, this should not be understood as requiring that such operations be performed in the particular order shown or in sequential order, or that all illustrated operations be performed, to achieve desirable results. The actions recited in the claims can be performed in a different order and still achieve desirable results. As one example, the processes depicted in the accompanying figures do not necessarily require the particular order shown, or sequential order, to achieve desirable results. In some cases, multitasking and parallel processing may be advantageous. Moreover, the separation of various system components in the various aspects described above should not be understood as requiring such separation in all aspects, and it should be understood that the described program components and systems can generally be integrated together in a single software product or packaged into multiple software products.

The title, background, brief description of the drawings, abstract and drawings are incorporated herein by reference into the present disclosure and are provided as illustrative examples of the present disclosure and not as limiting descriptions. It should be construed that they are not intended to limit the scope or meaning of the claims. Furthermore, in the detailed description, it can be seen that the present description provides illustrative examples, and various features are grouped together in various embodiments for the purpose of streamlining the disclosure. The methods of the present disclosure should not be construed as reflecting the following intent: the described subject matter requires more features than are expressly recited in each claim. Rather, as the following claims reflect, inventive subject matter lies in less than all features of a single disclosed configuration or operation. The claims are hereby incorporated into the detailed description, with each claim standing on its own as a separately delineated subject matter.

The claims are not intended to be limited to the aspects described herein but are to be accorded the full scope consistent with the language claims and encompassing all legal equivalents. However, none of the claims are intended to contain subject matter that fails to meet applicable patent statutes, nor should they be construed in this way.

Claims (15)

1. A computer-implemented method, comprising:

receiving a plurality of two-dimensional images, the plurality of two-dimensional images having at least two or more fields of view of a subject;

extracting a plurality of image features from the plurality of two-dimensional images using a set of learnable weights;

projecting the plurality of image features along a direction between the three-dimensional model of the subject and the selected viewpoint of the viewer; and

an image of the three-dimensional model of the subject is provided to the viewer.

2. The computer-implemented method of claim 1, wherein extracting a plurality of image features comprises: extracting inherent characteristics of a camera, wherein the camera is used for collecting the plurality of two-dimensional images.

3. The computer-implemented method of claim 1 or 2, wherein projecting a plurality of image features along a direction between the three-dimensional model of the subject and the selected viewpoint of the viewer comprises: the feature map associated with the first direction and the feature map associated with the second direction are interpolated.

4. The computer-implemented method of any preceding claim, wherein projecting a plurality of image features along a direction between the three-dimensional model of the subject and the selected viewpoint comprises: the plurality of image features are aggregated for a plurality of pixels along the direction between the three-dimensional model of the subject and the selected viewpoint.

5. The computer-implemented method of any preceding claim, wherein projecting a plurality of image features along a direction between the three-dimensional model of the subject and the selected viewpoint comprises: the plurality of feature maps generated by each of the plurality of cameras, each of the plurality of cameras having inherent characteristics, are concatenated in a permutation-invariant combination.

6. The computer-implemented method of any preceding claim, further comprising: a loss function is evaluated based on a difference between an image of the three-dimensional model of the subject and a ground truth image of the subject, and at least one of the set of learnable weights is updated based on the loss function.

7. The computer-implemented method of any preceding claim, wherein the principal is a user of a client device having a webcam directed to the user, the method further comprising: the selected viewpoint is identified as the location of the webcam pointing from the client device to the user.

8. The computer-implemented method of any preceding claim, wherein the viewer is using a network-coupled client device, and providing an image of the three-dimensional model of the subject comprises: streaming the following video to the network-coupled client device: the video has a plurality of images of a three-dimensional model of the subject.

9. The computer-implemented method of any preceding claim, wherein the subject is a user of a client device having an immersive reality application running therein, the method further comprising: the selected viewpoint is identified as the location within the immersive reality application where the viewer is located.

10. A system, comprising:

a memory storing a plurality of instructions; and

one or more processors configured to execute the plurality of instructions to cause the system to perform the method of any preceding claim.

11. A computer-implemented method for training a model to provide a view of a subject to an autostereoscopic display in a virtual reality head mounted viewer, the method comprising:

collecting a plurality of ground truth images from the faces of a plurality of users;

correcting the plurality of ground truth images with a plurality of stored calibrated stereo image pairs;

generating a plurality of composite views of a plurality of subjects with a three-dimensional facial model, wherein the plurality of composite views of the plurality of subjects include interpolation of a plurality of feature maps projected along different directions corresponding to the plurality of views of the plurality of subjects; and

The three-dimensional facial model is trained based on differences between the plurality of ground truth images and a plurality of synthesized views of the plurality of subjects.

12. The computer-implemented method of claim 11, wherein generating a plurality of composite views comprises: the image features from each of the plurality of ground truth images are projected along the selected viewing direction, and the plurality of feature maps generated by each of the plurality of ground truth images, each of the plurality of ground truth images having inherent characteristics, are concatenated in a permutation-invariant combination.

13. The computer-implemented method of claim 11 or 12, wherein training the three-dimensional facial model comprises: at least one of the learnable weights for each of the plurality of features in the plurality of feature maps is updated based on a value of a loss function indicative of the difference between the plurality of ground truth images and a plurality of composite views of the plurality of subjects.

14. The computer-implemented method of any of claims 11 to 13, wherein training the three-dimensional face model comprises:

Training a background value for each of a plurality of pixels in the plurality of ground truth images based on pixel background values projected from the plurality of ground truth images; and/or

A background model is generated using specific features of each of a plurality of cameras used to collect the plurality of ground truth images.

15. The computer-implemented method of any of claims 11 to 14, further comprising: the plurality of feature maps are interpolated by: the plurality of feature vectors from the plurality of cameras are averaged to form a camera aggregate feature vector of different directions at the desired point.