EP4264557A1 - Pixel-aligned volumetric avatars - Google Patents

Abstract

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

Description

PIXEL-ALIGNED VOLUMETRIC AVATARS

CROSS-REFERENCE TO RELATED APPLICATIONS

[0001] The present disclosure is related and claims priority under 35 U.S.C. §119(e) to US Prov. Appln. No. 63/129,989, entitled LEARNING TO PREDICT IMPLICIT VOLUMETRIC AVATARS to Lombardi, et-al., filed on Dec. 23, 2020, the contents of which are hereby incorporated by reference in their entirety, for all purposes.

TECHNICAL FIELD

[0002] The present disclosure is related to generating faithful facial expressions for generating real-time volumetric avatars in virtual reality (VR) and augmented reality (AR) applications. More specifically, the present disclosure provides real-time volumetric avatars in a multi-identity setting for VR/AR applications.

BACKGROUND

[0003] In the field of VR/AR applications, acquisition and rendering of photo-realistic human heads is a challenging problem for achieving virtual telepresence. Currently, the highest quality is achieved by volumetric approaches trained in a person-specific manner on multiview data. These models better represent fine structure, such as hair, compared to simpler mesh-based models. Volumetric models typically employ a global code to represent facial expressions, such that they can be driven by a small set of animation parameters. While such architectures achieve impressive rendering quality, they cannot easily be extended to the multiidentity setting, and are computationally expensive and hard to practice in “real-time” applications.

SUMMARY

[0004] In a first aspect, there is provided a computer-implemented method, comprising: receiving multiple two-dimensional images having at least two or more fields of view of a subject; extracting multiple image features from the two-dimensional images using a set of learnable weights; projecting the image features along a direction between a three-dimensional model of the subject and a selected observation point for a viewer; and providing, to the viewer, an image of the three-dimensional model of the subject.

[0005] Extracting image features may comprise extracting intrinsic properties of a camera used to collect the two-dimensional images.

[0006] Proj ecting image features along a direction between a three-dimensional model of the subject and a selected observation point for a viewer may comprise interpolating a feature map associated with a first direction with a feature map associated with a second direction. [0007] Proj ecting image features along a direction between a three-dimensional model of the subject and a selected observation point may comprise aggregating the image features for multiple pixels along the direction between the three-dimensional model of the subject and the selected observation point.

[0008] Proj ecting image features along a direction between a three-dimensional model of the subject and a selected observation point may comprise concatenating multiple feature maps produced by each of multiple cameras in a permutation invariant combination, each of the multiple cameras having an intrinsic characteristic.

[0009] The method may further comprise evaluating a loss function based on a difference between the 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. [0010] The subject may be a user of a client device having a webcam directed to the user. The method may further comprise identifying the selected observation point as a location of the webcam directed to the user from the client device.

[0011] The viewer may be using a network-coupled client device, and providing an image of the three-dimensional model of the subject may comprise streaming a video with multiple images of the three-dimensional model of the subject to the network-coupled client device.

[0012] The subject may be a user of a client device having an immersed reality application running therein, and the method may further comprise identifying the selected observation point as a location, within the immersed reality application, where the viewer is located.

[0013] In a second aspect, there is provided a system comprising: a memory storing multiple instructions and one or more processors configured to execute the instructions to cause the system to perform the method of the first aspect.

[0014] Also described is a computer program product comprising instructions which, when the program is executed by a computer, cause the computer to carry out the method of the first aspect.

[0015] Also described is a computer-readable storage medium comprising instructions which, when executed by a computer, cause the computer to carry out the method of the first aspect. The medium may be non-transitory.

[0016] In a third aspect, there is provided a computer-implemented method for training a model to provide a view of a subject to an auto stereoscopic display in a virtual reality headset, the method comprising: collecting, from a face of multiple users, multiple ground-truth images; rectifying the ground-truth images with stored, calibrated stereoscopic pairs of images; generating, with a three-dimensional face model, multiple synthetic views of subjects, wherein the synthetic views of subjects include an interpolation of multiple feature maps projected along different directions corresponding to multiple views of the subjects; and training the three-dimensional face model based on a difference between the ground-truth images and the synthetic views of subjects.

[0017] Generating multiple synthetic views may comprise projecting image features from each of the ground-truth images along a selected observation direction and concatenating multiple feature maps produced by each of the ground-truth images in a permutation invariant combination, each of the ground-truth images having an intrinsic characteristic.

[0018] Training the three-dimensional face model may comprise updating at least one in a set of learnable weights for each of multiple features in the feature maps based on a value of a loss function indicative of the difference between the ground-truth images and the synthetic views of subjects.

[0019] Training the three-dimensional face model may comprise training a background value for each of multiple pixels in the ground-truth images based on a pixel background value projected from the multiple ground-truth images.

[0020] The method may further comprise interpolating the feature maps by averaging multiple feature vectors from multiple cameras to form a camera summarized feature vector over different directions at a desired point.

[0021] Training the three-dimensional face model may comprise generating a background model using specific features for each of multiple cameras collecting the multiple ground-truth images.

[0022] Also described is a system comprising: a memory storing multiple instructions and one or more processors configured to execute the instructions to cause the system to perform the method of the third aspect.

[0023] Also described is a computer program product comprising instructions which, when the program is executed by a computer, cause the computer to carry out the method of the third aspect.

[0024] Also described is a computer-readable storage medium comprising instructions which, when executed by a computer, cause the computer to carry out the method of the third aspect. The medium may be non-transitory.

BRIEF DESCRIPTION OF THE FIGURES

[0025] FIG. 1 illustrates an example architecture suitable for providing a real-time, clothed subject animation in a virtual reality environment. [0026] FIG. 2 is a block diagram illustrating an example server and client from the architecture of FIG. 1.

[0027] FIG. 3 illustrates a block diagram of a model architecture used for a 3D rendition of a portion of a face of a VR/AR headset user.

[0028] FIGS. 4A-4C illustrate volumetric avatars computed given only two views as input.

[0029] FIG. 5 illustrates different techniques: Reality Capture, Neural Volumes, Globally conditioned, Neural Radiance Field (NeRF), and pixel-aligned techniques, compared to a ground-truth identity.

[0030] FIG. 6 illustrates generated alpha/normals/avatar in the canonical viewpoint using eNerf and pixel-aligned avatars compared to the ground truth.

[0031] FIG. 7 illustrates predicted texture with respect to the number of views.

[0032] FIG. 8 illustrates a background ablation result.

[0033] FIG. 9 illustrates the sensitivity of the pixel aligned features to the choice of the employed feature extractor, including a shallow convolutional network.

[0034] FIG. 10 illustrates a camera-aware feature summarization strategy.

[0035] FIG. 11 illustrates a flowchart in a method for rendering a three-dimensional (3D) view of a portion of a user’s face from multiple, two-dimensional (2D) images of a portion of the user’s face.

[0036] FIG. 12 illustrates a flowchart in a method for training a model to render a three- dimensional (3D) view of a portion of a user’s face from multiple, two-dimensional (2D) images of a portion of the user’s face.

[0037] FIG. 13 illustrates a computer system configured to perform at least some of the methods for using an AR or VR device.

[0038] In the figures, like elements are labeled likewise, according to their description, unless explicitly stated otherwise.

DETAILED DESCRIPTION OF THE FIGURES

[0039] In the following detailed description, numerous specific details are set forth to provide a full understanding of the present disclosure. It will be apparent, however, to one ordinarily skilled in the art, that 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 so as not to obscure the disclosure.

General Overview

[0040] Virtual telepresence applications try to represent with high accuracy and fidelity the human head. Human heads are challenging to model and render due to their complex geometry and appearance properties: sub-surface scattering of skin, fine-scale surface detail, thin- structured hair, and the human eyes as well as the teeth are both specular and translucent. Existing approaches include complex and expensive multi-view capture rigs (with up to hundreds of cameras) to reconstruct even a person-specific model of a human head. Currently, high-quality approaches are employed volumetric models rather than a textured mesh, since they can better leam to represent fine structures on the face like hair, which is critical to achieving a photorealistic appearance. Volumetric models typically employ a global code to represent facial expressions or only work for static scenes. While such architectures achieve impressive rendering quality, they are difficult to adapt to a multi-identity setting. A global code, as is used to control expression, is not sufficient for modeling identity variation across subjects. Some attempts to solve this problem include implicit models to represent scenes and objects. These models have the advantage that the scene is represented as a parametric function in a continuous space, which allows for fine-grained inference of geometry and texture. However, these approaches fail to model view-dependent effects presented for example with hair with a textured surface. The approach can generalize across objects, but only at low resolutions and can only handle purely Lambertian surfaces, which is not sufficient for human heads. One drawback of the above methods is that they are trained to model only a single scene or object. Methods which can generate multiple objects are typically limited in terms of quality and resolution of the predicted texture and geometry. For example, methods such as Scene Representation Networks (SRNs), which generate a set of weights from a global image encoding (e.g., a single latent code vector per image), have difficulty generalizing to local changes (e.g., facial expressions) and fail to recover high-frequency details even when these are visible in the input images. This is because the global latent code summarizes information in the image and must discard some information to generate a compact encoding of the data.

[0041] To solve the above problem in the field of immersive reality applications for computer networks, embodiments as disclosed herein implement predicted volumetric avatars of the human head from a limited number of inputs. To achieve this, embodiments as disclosed herein enable model generalization across multiple identities by a parameterization that combines neural radiance fields with local, pixel-aligned features extracted directly from model inputs. This approach results in shallow and simple networks that can be implemented in real-time immersive applications. In some embodiments, models trained on a photometric re-rendering loss function may not use explicit 3D supervision to render a subject-based avatar in real time. Models as disclosed herein generate faithful facial expressions in a multi-identity setting, and are thus applicable in the field of real-time group-immersive applications. Embodiments as disclosed herein generalize to multiple, unseen identities and expressions in real-time, and provide a good representation of temporal image sequences.

[0042] Some embodiments include a pixel-aligned volumetric avatar (PVA) model for the estimation of a volumetric 3D avatar using only a few input images of a human head. The PVA model is able to generalize to unseen identities in real-time. To improve generalization across identities, a PVA model parameterizes the volumetric model via local, pixel-aligned features extracted from the input images. Thus, the PVA model can synthesize novel views for unseen identities and expressions while preserving high frequency details in the rendered avatar. In addition, some embodiments include a pixel-aligned radiance field that predicts implicit shape and appearance from a sparse set of posed images for any point in space, in any direction of view.

Example System Architecture

[0043] FIG. 1 illustrates an example architecture 100 suitable for accessing a volumetric avatar model engine. Architecture 100 includes servers 130 communicatively coupled with client devices 110 and at least one database 152 over a network 150. One of the many servers 130 is configured to host a memory including instructions which, when executed by a processor, cause the server 130 to perform at least some of the steps in methods as disclosed herein. In some examples, the processor is configured to control a graphical user interface (GUI) for the user of one of client devices 110 accessing the volumetric avatar model engine with an immersive reality application. Accordingly, the processor may include a dashboard tool, configured to display components and graphic results to the user via the GUI. For purposes of load balancing, multiple servers 130 can host memories including instructions to one or more processors, and multiple servers 130 can host a history log and a database 152 including multiple training archives used for the volumetric avatar model engine. Moreover, in some examples, multiple users of client devices 110 may access the same volumetric avatar model 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 devices 110 and servers 130 may communicate with each other via network 150 and resources located therein, such as data in database 152.

[0044] Servers 130 may include any device having an appropriate processor, memory, and communications capability for hosting the volumetric avatar model engine including multiple tools associated with it. The volumetric avatar model engine may be accessible by various clients 110 over network 150. Clients 110 can be, for example, desktop computers, mobile computers, tablet computers (e.g., including e-book readers), mobile devices (e.g, a smartphone or PDA), or any other device having appropriate processor, memory, and communications capabilities for accessing the volumetric avatar model engine on one or more of servers 130. In some examples, client devices 110 may include VR/ AR headsets configured to run an immersive reality application using a volumetric avatar model supported by one or more of servers 130. Network 150 can include, for example, any one or more of a local area tool (LAN), a wide area tool (WAN), the Internet, and the like. Further, network 150 can include, but is not limited to, any one or more of the following tool topologies, including a bus network, a star network, a ring network, a mesh network, a star-bus network, tree or hierarchical network, and the like.

[0045] 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 over network 150 via respective communications modules 218-1 and 218-2 (hereinafter, collectively referred to as “communications modules 218”). Communications modules 218 are configured to interface with network 150 to send and receive information, such as data, requests, responses, and commands to other devices via network 150. Communications modules 218 can be, for example, modems or Ethernet cards, and may include radio hardware and software for wireless communications (e.g, via electromagnetic radiation, such as radiofrequency -RF-, near field communications -NFC-, WiFi, and BlueTooth radio technology). A user may interact with client device 110 via an input device 214 and an output device 216. Input device 214 may include a mouse, a keyboard, a pointer, a touchscreen, a microphone, a joystick, a virtual joystick, and the like. In some examples, input device 214 may include cameras, microphones, and sensors, such as touch sensors, acoustic sensors, inertial motion units -IMUs- and other sensors configured to provide input data to a VR/AR headset. For example, in some examples, input device 214 may include an eye tracking device to detect the position of a user’s pupil in a VR/AR headset. Output device 216 may be a screen display, a touchscreen, a speaker, and the like. Client device 110 may include a memory 220- 1 and a processor 212-1. Memory 220-1 may include an application 222 and a GUI 225, configured to run in client device 110 and couple with input device 214 and output device 216. Application 222 may be downloaded by the user from server 130 and may be hosted by server 130. In some examples, client device 110 is a VR/AR headset and application 222 is an immersive reality application.

[0046] Server 130 includes a memory 220-2, a processor 212-2, and communications module 218-2. Hereinafter, processors 212-1 and 212-2, and memories 220-1 and 220-2, will be collectively referred to, respectively, as “processors 212” and “memories 220.” Processors 212 are configured to execute instructions stored in memories 220. In some examples, memory 220-2 includes a volumetric avatar model engine 232. Volumetric avatar model engine 232 may share or provide features and resources to GUI 225, including multiple tools associated with training and using a three-dimensional avatar rendering model for immersive reality applications (e.g., application 222). The user may access volumetric avatar model engine 232 through application 222, installed in a memory 220-1 of client device 110. Accordingly, application 222, including GUI 225, may be installed by server 130 and perform scripts and other routines provided by server 130 through any one of multiple tools. Execution of application 222 may be controlled by processor 212-1.

[0047] In that regard, volumetric avatar model engine 232 may be configured to create, store, update, and maintain a PVA model 240, as disclosed herein. PVA model 240 may include an encoder-decoder tool 242, a ray marching tool 244, and a radiance field tool 246. Encoderdecoder tool 242 collects input images with multiple, simultaneous views of a subject and extracts pixel-aligned features to condition radiance field tool 246 via a ray marching procedure in ray marching tool 244. PVA model 240 can generate novel views of unseen subjects from one or more sample images processed by encoder-decoder tool 242. In some examples, encoder-decoder tool 242 is a shallow (e.g, including a few one- or two-node layers) convolutional network. In some examples, radiance field tool 246 converts three-dimensional location and pixel-aligned features into color and opacity fields that can be projected in any desired direction of view.

[0048] In some examples, volumetric avatar model engine 232 may access one or more machine learning models stored in a training database 252. Training database 252 includes training archives and other data files that may be used by volumetric avatar model engine 232 in the training of a machine learning model, according to the input of the user through application 222. Moreover, in some examples, at least one or more training archives or machine learning models may be stored in either one of memories 220 and the user may have access to them through application 222.

[0049] Volumetric avatar model engine 232 may include algorithms trained for the specific purposes of the engines and tools included therein. The algorithms may include machine learning or artificial intelligence algorithms making use of any linear or non-linear algorithm, such as a neural network algorithm, or multivariate regression algorithm. In some examples, the machine learning model may include a neural network (NN), a convolutional neural network (CNN), a generative adversarial neural network (GAN), a deep reinforcement learning (DRL) algorithm, a deep recurrent neural network (DRNN), a classic machine learning algorithm such as random forest, k-nearest neighbor (KNN) algorithm, k-means clustering algorithms, or any combination thereof. More generally, the machine learning model may include any machine learning model involving a training step and an optimization step. In some examples, training database 252 may include a training archive to modify coefficients according to a desired outcome of the machine learning model. Accordingly, in some examples, volumetric avatar model engine 232 is configured to access training database 252 to retrieve documents and archives as inputs for the machine learning model. In some examples, volumetric avatar model engine 232, the tools contained therein, and at least part of training database 252 may be hosted in a different server that is accessible by server 130 or client device 110.

[0050] FIG. 3 illustrates a block diagram of a model architecture 300 used for a 3D rendition of a face portion of a VR/AR headset user. Model architecture 300 is a pixel aligned volumetric avatar (PVA) model. PVA model 300 is learned from a multi-view image collection that produces multiple, 2D input images 301-1, 301-2, and 301-n (hereinafter, collectively referred to as “input images 301”). Each of input images 301 is associated with a camera view vector, Vi (e.g. , vi, V2 and v_n), which indicates the direction of view of the user’s face for that particular image. In some examples, input images 301 are collected simultaneously, or quasi- simultaneously, so that the different view vectors, Vi, point to the same volumetric representation of a subject. Each of vectors vi is a known viewpoint 311, associated with camera intrinsic parameters, Ki, and rotation, Ri (e.g., {Ki, [R|t]i}). Camera intrinsic parameters Ki may include brightness, color mapping, sensor efficiency and other cameradependent parameters. Rotation, Ri indicates the orientation (and distance) of the subject’s head relative to the camera. The different camera sensors have a slightly different response to the same incident radiance despite the fact that they are the same camera model. If nothing is done to address this, the intensity differences end up baked into the scene representation N, which will cause the image to unnaturally brighten or darken from certain viewpoints. To address this, we learn a per-camera bias and gain value. This allows the system to have an ‘easier’ way to explain this variation in the data.

[0051] The value of ‘n’ is purely exemplary, as anyone with ordinary skills would realize that any number, n, of input images 301 can be used. PVA model 300 produces a volumetric rendition 321 of the headset user. Volumetric rendition 321 is a 3D model (e.g., “avatar”) that can be used to generate a 2D image of the subject from the target viewpoint. This 2D image changes as the target viewpoint changes (e.g., as the viewer moves around the headset user). [0052] PVA model 300 includes a convolutional encoder-decoder 310A, a ray marching stage 310B, and a radiance field stage 3 IOC (hereinafter, collectively referred to as “PVA stages 310”). PVA model 300 is trained with input images 301 selected from a multi-identity training corpus, using gradient descent. Accordingly, PVA model 300 includes a loss function defined between predicted images from multiple subjects and the corresponding ground truth. This enables PVA model 300 to render accurate volumetric renditions 321 independently of the subject.

[0053] Convolutional encoder-decoder network 310A takes input images 301 and produces pixel-aligned feature maps 303-1, 303-2, and 303-n (hereinafter, collectively referred to as “feature maps 303,” f(i)). Ray marching stage 310B follows each of the pixels along a ray in target viewj, defined by {Kj , [R|t]j } , accumulating color, c, and optical density (“opaqueness”) produced by radiance field stage 310C at each point. Radiance field stage 310C (N) converts 3D location and pixel-aligned features to color and opacity, to render a radiance field 315 (c, o).

[0054] Input images 301 are 3D objects having a height (h) and a width (w) corresponding to the 2D image collected by a camera along direction vi, and a depth of 3 layers for each color pixel R, G, B. Feature maps 303 are 3D objects having dimensions h xw x d. Encoder-decoder network 310A encodes input images 301 using learnable weights 320-1, 320-2... 320-n (hereinafter, collectively referred to as “learnable weights 320”). Ray marching stage 310B performs world to camera projections 323, bilinear interpolations 325, positional encoding 327, and feature aggregation 329.

[0055] In some examples, for a conditioning view Vi G R ^hxwx3 feature maps 303 may be defined as functions

[0056] where Φ (X): R³ → R ^6x1 is the positional encoding of a point 330 (X G R ³) with 2 x I different basis functions. Point 330 (X), is a point along a ray directed from a 2D image of the subject to a specific viewpoint 331, ro. Feature maps 303 (f⁽ⁱ⁾ ∈ R^hxwxd) are associated with a camera position vector, vi, where d is the number of feature channels, h and w are image height and width, and fx ∈ R^d is an aggregated image feature associated with point X. For each feature map f⁽ⁱ⁾, ray marching stage 310B obtains fx G R^d by projecting 3D point X along the ray using camera intrinsic (K) and extrinsic (R, t) parameters of that particular viewpoint, [0057] where Π is a perspective projection function to camera pixel coordinates, and F(f, x) is the bilinear interpolation 325 of f at pixel location x. Ray marching stage 31 OB combines pixel-aligned features f⁽ⁱ⁾x from multiple images for radiance field stage 310C.

[0058] For each given training image vj with camera intrinsics Kj and rotation and translation Rj, tj , the predicted color of a pixel p G R² for a given viewpoint in the focal plane of the camera and center 331 (ro) G R³ is obtained by marching rays into the scene using the camera- to-world projection matrix, with the direction of the rays given by,

[0059] Ray marching stage 310B accumulates radiance and opacity values along a ray 335 defined by r(t) = r₀ + td for t G [t_near, t_far] as follows:

(6)

[0060] Where,

(7)

[0061] In some examples, ray marching stage 310B uniformly samples a set of n_s points t ~ [t_near, t_far] . Setting X = r(t) the quadrature rule may be used to approximate integrals 6 and 7. A function I_α(p) may be defined as

(8)

[0062] where αi = 1 - exp(-δi -σi) with δi being the distance between the i+l-th and i-th sample point along ray 335. [0063] In a multi-view setting with known camera viewpoints, Vi, and a fixed number of conditioning views ray marching stage 310B aggregates the features by simple concatenation. Concretely, for n conditioning images {vi}ⁿi=i with corresponding rotation and translation matrices given by {Ri}ⁿi=i and {ti}ni=l, using features {f⁽ⁱ⁾x }ⁿi=i for each point X as in Eq. (3), ray marching stage 310B generates the final feature as follows,

[0064] Where © represents a concatenation along the depth dimension. This preserves feature information from viewpoints, {vi}ⁿi=i, helping PVA model 300 to determine the best combination and employ the conditioning information.

[0065] In some examples, PVA model 300 is agnostic to viewpoint and number of conditioning views. Simple concatenation as above is insufficient in this case, since the number of conditioning views may not be known a priori, leading to different feature dimensions (d) during inference time. To summarize features for a multi-view setting, some examples include a permutation invariant function G: R^nxd R^d such that for any permutation ψ,

[0066] A simple permutation invariant function for feature aggregation is the mean of the sampled feature maps 303. This aggregation procedure may be desirable when depth information during training is available. However, in the presence of depth ambiguity (e.g, for points that are projected onto feature map 303 before sampling), the above aggregation may lead to artifacts. To avoid this, some examples consider camera information to include effective conditioning in radiant field stage 310C. Accordingly, some examples include a conditioning function network N_cf : R^d+7 R^d that takes the feature vector, f⁽ⁱ⁾x, and the camera information (ci) and produces a camera summarized feature vector f'⁽ⁱ⁾x. These modified vectors are then averaged over multiple, or all, conditioning views, as follows

[0067] The advantage of this approach is that the camera summarized features can take likely occlusions into account before the feature average is performed. The camera information is encoded as a 4D rotation quaternion and a 3D camera position. [0068] Some examples may also include a background estimation network, N_bg, to avoid learning parts of the background in the scene representation. Background estimation network, N_bg, may be defined as: N_bg: R^nc:→ R^hxwx3 to learn a per-camera fixed background. In some examples, radiant field stage 310C may use N_bg to predict the final image pixels as:

[0069] with Ibg = Ibg +N_bg(ci) for camera Ci where Ī_bg is an initial estimate of the background extracted using inpainting, and lex is as defined by Eq. (8). These inpainted backgrounds are often noisy, leading to ‘halo’ effects around the head of the person. To avoid this, N_bg model learns the residual to the inpainted background. This has the advantage of not needing a high capacity network to account for the background.

[0070] For ground-truth target images vj, PVA model 300 trains both radiance field stage 310C and feature extraction network using a simple photo-metric reconstruction loss:

[0071] FIGS. 4A-4C illustrate volumetric avatars 421A-1, 421A-2, 421A-3, 421A-4, and 421A-5 (hereinafter, collectively referred to as “volumetric avatars 421A”), 421B-1, 421B-2, 421B-3, 421B-4, and 421B-5 (hereinafter, collectively referred to as “volumetric avatars 421B”), 421C-1, 421C-2, 421C-3, 421C-4, and 421C-5 (hereinafter, collectively referred to as “volumetric avatars 421C”). Volumetric avatars 421A, 421B, and 421C (hereinafter, collectively referred to as “volumetric avatars 421”) are high fidelity reproductions of diverse subjects, obtained using only two subject views as input.

[0072] Volumetric avatars 421 illustrate that a PVA model as disclosed herein can generate multiple views of different subject avatars given only two views as input, from a large number of novel viewpoints.

[0073] FIG. 5 illustrates avatars 521A-1, 521A-2, 521A-3, 521A-4 (hereinafter, collectively referred to as “reality capture avatars 521A”), avatars 521B-1, 521B-2, 521B-3, 521B-4 (hereinafter, collectively referred to as “neural volumes avatars 521B”), avatars 521C-1, 521C- 2, 521 C-3, 521 C-4 (hereinafter, collectively referred to as “globally conditioned, cNeRF avatars 521C”), and avatars 521D-1, 521D-2, 521D-3, 521D-4 (hereinafter, collectively referred to as “volumetric avatars 521D”). Avatars 521A, 521B, 521C, and 521D will be hereinafter collectively referred to as “avatars 521.” Avatars 521 are obtained from two input images of the novel identities and compared to ground-truth images 501-1, 501-2, 501-3, and 501-4 (hereinafter, collectively referred to as “ground-truth images 501”), as input to compute the reconstruction. [0074] Reality capture avatars 521A are obtained with a structure-from-motion (SFM) and multi-view stereo (MVS) algorithm that reconstructs a 3D model from a set of captured images. Neural volumes avatars 521B are obtained with a voxel-based inference method that globally encodes dynamic images of a scene and decodes a voxel grid and a warp field that represents the scene. cNeRF avatars 521 C are a variant of NeRF algorithms with global identity conditioning (cNeRF). In some examples, cNeRF avatars 521C employ a VGG network to extract a single 64D feature vector for each training identity, and condition a NeRF model additionally on this input. Volumetric avatars 521D are obtained with a PVA model, as disclosed herein.

[0075] Volumetric avatars 521D are more complete reconstructions than reality capture avatars 521A, which typically use many more ground-truth images 501 to obtain a good reconstruction. Volumetric avatars 521D also leads to more detailed reconstruction than cNerf avatars 521 C due to the pixel-aligned features in PVA models as disclosed herein, which provide a more complete information for the model at test time.

[0076] Table 1 below is a comparison of the performance volumetric avatars 521D with NV avatars 521B and cNeRF avatars 521C using different metrics. A structural similarity index (SSIM) preferably has a maximum value of one (1), and a learned perceptual image patch similarity (LPIPS) metric and a mean squared error (MSE) metric preferably have lower values.

[0077] A PVA model as disclosed herein resolve certain shortcomings of global identity encoding methods trained in a scene specific manner like Neural Volumes and cNeRF, which do not generalize well to unseen identities. For example, cNeRF avatar 521 C has the facial features smoothed out and some of the local details of unseen identities are lost (like facial hair in 521C-3 and 521C-4, and hair length in 521 C-2), since this model relies heavily on the learned global prior. Reality capture avatar 521A fails to capture the head structure as there are no prior models built into the SfM+MVS framework, leading to incomplete reconstructions. A large number of images would be required to faithfully reconstruct a novel identity for RC 521A models. Neural volumes avatars 521B generate better textures because of the generated warp field which accounts for some degree of local information. However, neural volume avatars 521B use an encoder configured with a global encoding and projects test time identities into the nearest training time identity, leading to inaccurate avatar predictions. Volumetric avatars 521D reconstruct volumetric heads from just two example viewpoints, along with the structure of the hair.

[0078] FIG. 6 illustrates generated 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 621 A-l, 621 A- 2, and 621A-3 (hereinafter, collectively referred to as “avatars 621 A”) using eNerf, and associated ground-truth images 601-1, 601-2, and 601-3 for three different subjects. 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 “avatars 62 IB”) for pixel-aligned avatars obtained with a PVA model, as disclosed herein.

[0079] Avatars 621B are well suited to capture expression information, compared to avatars 621A, which have a harder time generalizing facial expressions to novel identities. To obtain 621 A avatars, aNeRF model is conditioned on a one-hot expression code and one-hot identity information (eNeRF) on test time identities. However, despite having seen all the identities during training, avatars 621B better generalize to dynamic expressions for multiple identities, compared to avatars 621 A. Because the PVA model leverages local features for conditioning, avatars 621B capture dynamic effects on a specific identity (both geometry and texture) better than avatars 621 A.

[0080] FIG. 7 illustrates 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 avatars 721.

[0081] Avatars 721 and normal views 733 illustrate different views of the subject than captured in images 701-1, 701-2, 701-3, 701-4, and 701-5 (hereinafter, collectively referred to as “ground-truth images 701”). Because PVA models learn shape priors from training identities, normal views 733 are consistent with the identity of ground-truth images 701. However, when extrapolating to extreme views (733-1 and 721-1), artifacts appear in the parts of the face that are unseen in the “conditioning” ground-truth images 701. This is due to the inherent depth ambiguity due to projection of the sample points onto ground-truth image 701- 1. Addition of a second view (e.g., ground-truth image 701-2) already significantly reduces these artifacts in normal view 733-2 and avatar 721-2, as the PVA model now has more information regarding features from different views and thus depth information. In general, a PVA model as disclosed herein can achieve a large degree of view extrapolation with just two conditioning views.

[0082] FIG. 8 illustrates background ablation results for 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”).

[0083] FIG. 9 illustrates the sensitivity of a PVA model to the choice of the employed feature extractor 921A (“hour glass network”), 921B (“UNet”), and 921C (shallow convolutional network), based on a conditioning view 901. 921A and 921B are reliable feature extractors. In some examples, shallow encoder-decoder architecture 921 C performs may be desirable, as it preserves more of the local information without having to encode all the pixel level information into a bottleneck layer.

[0084] FIG. 10 illustrates a camera-aware feature summarization strategy. For a 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 (avatar 1021A-1) or with (avatar 1021B-1), camera specific information. Likewise, for a 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 (avatar 1021A-2) or with (avatar 1021B-2), camera specific information. And for a 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 (avatar 1021A-3) or with (avatar 1021B-3), camera specific information (cf. camera intrinsic (K) and extrinsic (R, t) parameters, Eq. (3)).

[0085] Particularly, without the camera information, avatars 1021 A-1, 1021 A-2, and 1021 A- 3 present streaking in the generated images due to inconsistent averaging of information from different viewpoints (particularly in avatars 1021 A-l and 1021 A-2).

[0086] FIG. 11 illustrates a flowchart in a method 1100 for rendering a three-dimensional (3D) view of a portion of a user’s face from multiple, two-dimensional (2D) images of a portion of the user’s face. Steps in method 1100 may be performed at least partially by a processor executing instructions stored in a memory, wherein the processor and the memory are part of a client device or a VR/AR headset as disclosed herein (e.g, memories 220, processors 212, client devices 110). In yet other examples, at least one or more of the steps in 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 are remotely located in a cloud server, and the headset device is communicatively coupled to the cloud server via a communications module coupled to a network (cf. server 130, communications modules 218). In some examples, method 1100 may be performed using a volumetric avatar model engine configured to train a PVA model including an encoder-decoder tool, a ray marching tool, and a radiance field tool, these tools including a neural network architecture in a machine learning or artificial intelligence algorithm, as disclosed herein (e.g., volumetric avatar model engine 232, PVA model 240, encoder-decoder tool 242, ray marching tool 244, and radiance field tool 246). In some examples, methods consistent with the present disclosure may include at least one or more steps from method 1100 performed in a different order, simultaneously, quasi- simultaneously, or overlapping in time.

[0087] Step 1102 includes receiving multiple two-dimensional images having at least two or more fields of view of a subject.

[0088] Step 1104 includes extracting multiple image features from the two-dimensional images using a set of learnable weights. In some examples, step 1104 includes extracting intrinsic properties of a camera used to collect the two-dimensional images.

[0089] Step 1106 includes projecting the image features along a direction between a three- dimensional model of the subject and a selected observation point for a viewer. In some examples, step 1106 includes comprises interpolating a feature map associated with a first direction with a feature map associated with a second direction. In some examples, step 1106 includes aggregating the image features for multiple pixels along the direction between the three-dimensional model of the subject and the selected observation point. In some examples, step 1106 includes concatenating multiple feature maps produced by each of multiple cameras in a permutation invariant combination, each of the multiple cameras having an intrinsic characteristic.

[0090] In some examples, the subject is a user of a client device having a webcam directed to the user, and step 1106 includes identifying the selected observation point as a location of the webcam directed to the user from the client device. In some examples, the subject is a user of a client device having an immersed reality application running therein, and step 1106 further includes identifying the selected observation point as a location, within the immersed reality application, where the viewer is located. [0091] Step 1108 includes providing, to the viewer, an image of the three-dimensional model of the subject. In some examples, step 1108 includes evaluating a loss function based on a difference between the 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 streaming a video with multiple images of the three-dimensional model of the subject to the network-coupled client device.

[0092] FIG. 12 illustrates a flowchart in a method 1200 for training a model to render a three- dimensional (3D) view of a portion of a user’s face from multiple, two-dimensional (2D) images of a portion of the user’s face. Steps in method 1200 may be performed at least partially by a processor executing instructions stored in a memory, wherein the processor and the memory are part of a client device or a VR/AR headset as disclosed herein (e.g., memories 220, processors 212, client devices 110). In yet other examples, at least one or more of the steps in 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 are remotely located in a cloud server, and the headset device is communicatively coupled to the cloud server via a communications module coupled to a network (cf. server 130, communications modules 218). In some examples, method 1200 may be performed using a volumetric avatar model engine configured to train a PVA model including an encoder-decoder tool, a ray marching tool, and a radiance field tool, these tools including a neural network architecture in a machine learning or artificial intelligence algorithm, as disclosed herein (e.g., volumetric avatar model engine 232, PVA model 240, encoder-decoder tool 242, ray marching tool 244, and radiance field tool 246). In some examples, methods consistent with the present disclosure may include at least one or more steps from method 1200 performed in a different order, simultaneously, quasi-simultaneously, or overlapping in time.

[0093] Step 1202 includes collecting, from a face of multiple users, multiple ground-truth images.

[0094] Step 1204 includes rectifying the ground-truth images with stored, calibrated stereoscopic pairs of images.

[0095] Step 1206 includes generating, with a three-dimensional face model, multiple synthetic views of subjects, wherein the synthetic views of subjects include an interpolation of multiple feature maps projected along different directions corresponding to multiple views of the subjects. In some examples, step 1206 includes projecting image features from each of the ground-truth images along a selected observation direction and concatenating multiple feature maps produced by each of the ground-truth images in a permutation invariant combination, each of the ground-truth images having an intrinsic characteristic. In some examples, step 1206 further includes interpolating the feature maps by averaging multiple feature vectors from multiple cameras to form a camera summarized feature vector over different directions at a desired point.

[0096] Step 1208 includes training the three-dimensional face model based on a difference between the ground-truth images and the synthetic views of subjects. In some examples, step 1208 includes updating at least one in a set of learnable weights for each of multiple features in the feature maps based on a value of a loss function indicative of the difference between the ground-truth images and the synthetic views of subjects. In some examples, step 1208 includes training a background value for each of multiple pixels in the ground-truth images based on a pixel background value projected from the multiple ground-truth images. In some examples, step 1208 includes generating a background model using specific features for each of multiple cameras collecting the multiple ground-truth images.

Hardware Overview

[0097] FIG. 13 is a block diagram illustrating an exemplary computer system 1300 with which headsets and other client devices 110, and methods 1100 and 1200 can be implemented. In certain aspects, computer system 1300 may be implemented using hardware or a combination of software and hardware, either 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, a phablet, a smartphone, a feature phone, a server computer, or otherwise. A server computer may be located remotely in a data center or be stored locally.

[0098] Computer system 1300 includes a bus 1308 or other communication mechanism for communicating information, and a processor 1302 (e.g., processors 212) coupled with bus 1308 for processing information. By way of example, the computer system 1300 may be implemented with one or more processors 1302. Processor 1302 may be a general-purpose microprocessor, a microcontroller, a Digital Signal Processor (DSP), an Application Specific Integrated Circuit (ASIC), a Field Programmable Gate Array (FPGA), a Programmable Logic Device (PLD), a controller, a state machine, gated logic, discrete hardware components, or any other suitable entity that can perform calculations or other manipulations of information.

[0099] Computer system 1300 can include, in addition to hardware, code that creates an execution environment for the computer program in question, e.g., code that constitutes processor firmware, a protocol stack, a database management system, an operating system, or a combination of one or more of them stored in an included memory 1304 (e.g., memories 220), such as a Random Access Memory (RAM), a flash memory, a Read-Only Memory (ROM), a Programmable Read-Only Memory (PROM), an Erasable PROM (EPROM), registers, a hard disk, a removable disk, a CD-ROM, a DVD, or any other suitable storage device, 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.

[00100] The instructions may be stored in the memory 1304 and implemented in one or more computer program products, e.g., one or more modules of computer program instructions encoded on a computer-readable medium for execution by, or to control the operation of, the computer system 1300, and according to any method well known to those of skill in the art, including, but not limited to, computer languages such as data-oriented languages (e.g , SQL, dBase), system languages (e.g., C, Objective-C, C++, Assembly), architectural languages (e.g, Java, .NET), and application languages (e.g., PHP, Ruby, Perl, Python). Instructions may also be implemented in computer languages such as array languages, aspect-oriented languages, assembly languages, authoring languages, command line interface languages, compiled languages, concurrent languages, curly-bracket languages, dataflow languages, data-structured languages, declarative languages, esoteric languages, extension languages, fourth-generation languages, functional languages, interactive mode languages, interpreted languages, iterative languages, list-based languages, little languages, logic-based languages, machine languages, macro languages, metaprogramming languages, multiparadigm languages, numerical analysis, non-English-based languages, object-oriented class-based languages, object-oriented prototype-based languages, off-side rule languages, procedural languages, reflective languages, rule-based languages, scripting languages, stack-based languages, synchronous languages, syntax handling languages, visual languages, wirth languages, and xml-based languages. Memory 1304 may also be used for storing temporary variable or other intermediate information during execution of instructions to be executed by processor 1302.

[00101] A computer program as discussed herein does not necessarily correspond to a file in a file system. A program can 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 program in question, or in multiple coordinated files (e.g, files that store one or more modules, subprograms, or portions of code). A computer program can be deployed to be executed on one computer or on multiple computers that are located 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.

[00102] Computer system 1300 further includes a data storage device 1306 such as a magnetic disk or optical disk, coupled with bus 1308 for storing information and instructions. Computer system 1300 may be coupled via input/output module 1310 to various devices. Input/output module 1310 can be any input/output module. Exemplary input/output modules 1310 include data ports such as USB ports. The input/output module 1310 is configured to connect to a communications module 1312. Exemplary communications modules 1312 include networking interface cards, such as Ethernet cards and modems. In certain aspects, input/output module 1310 is configured to connect to a plurality of devices, such as an input device 1314 and/or an output device 1316. Exemplary input devices 1314 include a keyboard and a pointing device, e.g., a mouse or a trackball, by which a consumer can provide input to the computer system 1300. Other kinds of input devices 1314 can be used to provide for interaction with a consumer as well, such as a tactile input device, visual input device, audio input device, or brain-computer interface device. For example, feedback provided to the consumer can be any form of sensory feedback, e.g., visual feedback, auditory feedback, or tactile feedback; and input from the consumer can be received in any form, including acoustic, speech, tactile, or brain wave input. Exemplary output devices 1316 include display devices, such as an LCD (liquid crystal display) monitor, for displaying information to the consumer.

[00103] According to one aspect of the present disclosure, headsets and client devices 110 can be implemented, at least partially, using a computer system 1300 in response to processor 1302 executing one or more sequences of one or more instructions contained in memory 1304. Such 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 various aspects of the present disclosure. Thus, aspects of the present disclosure are not limited to any specific combination of hardware circuitry and software.

[00104] Various 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 consumer interface or a Web browser through which a consumer can interact with an implementation of the subject matter described in this specification, or any combination of one or more such back end, middleware, or front end components. The components of the system can be interconnected by any form or medium of digital data communication, e.g., a communication network. The communication network can include, for example, any one or more of a LAN, a WAN, the Internet, and the like. Further, the communication network can include, but is not limited to, for example, any one or more of the following network topologies, including a bus network, a star network, a ring network, a mesh network, a star-bus network, tree or hierarchical network, or the like. The communications modules can be, for example, modems or Ethernet cards.

[00105] Computer system 1300 can include clients and servers. A client and server are generally 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 can be, for example, and without limitation, a desktop computer, laptop computer, or tablet computer. Computer system 1300 can also be embedded in another device, for example, and without limitation, a mobile telephone, a PDA, a mobile audio player, a Global Positioning System (GPS) receiver, a video game console, and/or a television set top box.

[00106] 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, volatile media, and transmission media. Non-volatile media include, for example, optical or magnetic disks, such as data storage device 1306. Volatile media include dynamic memory, such as memory 1304. Transmission media include coaxial cables, copper wire, and fiber optics, including the wires forming bus 1308. Common forms of machine-readable media include, for example, 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, a RAM, a PROM, an EPROM, a FLASH EPROM, any other memory chip or cartridge, or any other medium from which a computer can read. The machine-readable storage medium can be a machine-readable storage device, a machine-readable storage substrate, a memory device, a composition of matter affecting a machine-readable propagated signal, or a combination of one or more of them.

[00107] To illustrate the interchangeability of hardware and software, items such as the various illustrative blocks, modules, components, methods, operations, instructions, and algorithms have been described generally 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.

[00108] As used herein, the phrase “at least one of’ preceding a series of items, with the terms “and” or “or” to separate any of the items, modifies the list as a whole, rather than each member of the list (e.g., each item). The phrase “at least one of’ does not require selection of at least one item; rather, the phrase allows a meaning that includes at least one of any one of the items, and/or at least one of any combination of the items, and/or at least one of each of the items. By way of example, the phrases “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; any combination of A, B, and C; and/or at least one of each of A, B, and C.

[00109] 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 an aspect, the 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, the subject technology, the disclosure, the present disclosure, and other variations thereof and alike are for convenience and do not imply that a disclosure relating to such phrase(s) is essential to the subject technology or that such disclosure applies to all configurations of the subject technology. A disclosure relating to such phrase(s) may apply to all configurations, or one or more configurations. A disclosure relating to such phrase(s) may provide one or more examples. A phrase such as an aspect or some aspects may refer to one or more aspects and vice versa, and this applies similarly to other foregoing phrases.

[00110] A reference to an element in the singular is not intended to mean “one and only one” unless specifically stated, but rather “one or more.” The term “some” refers to one or more. 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 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 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 is to be construed under the provisions of 35 U.S.C. §112, sixth paragraph, unless the element is expressly recited using the phrase “means for” or, in the case of a method claim, the element is recited using the phrase “step for.”

[00111] 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 particular implementations 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. Moreover, although features may be described above as acting in certain combinations and even initially described 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.

[00112] The subject matter of this specification has been described in terms of particular aspects, but other aspects can be implemented and are within the scope of the following claims. For example, while 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 certain circumstances, multitasking and parallel processing may be advantageous. Moreover, the separation of various system components in the 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.

[00113] The title, background, brief description of the drawings, abstract, and drawings are hereby incorporated into the disclosure and are provided as illustrative examples of the disclosure, not as restrictive descriptions. It is submitted with the understanding that they will not be used to limit the scope or meaning of the claims. In addition, in the detailed description, it can be seen that the description provides illustrative examples and the various features are grouped together in various implementations for the purpose of streamlining the disclosure. The method of disclosure is not to be interpreted as reflecting an intention that the described subject matter requires more features than are expressly recited in each claim. Rather, as the 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 described subject matter.

[00114] 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 to encompass all legal equivalents. Notwithstanding, none of the claims are intended to embrace subject matter that fails to satisfy the requirements of the applicable patent law, nor should they be interpreted in such a way.

Claims

1. A computer-implemented method, comprising: receiving multiple two-dimensional images having at least two or more fields of view of a subject; extracting multiple image features from the two-dimensional images using a set of learnable weights; projecting the image features along a direction between a three-dimensional model of the subject and a selected observation point for a viewer; and providing, to the viewer, an image of the three-dimensional model of the subject.

2. The computer-implemented method of claim 1, wherein extracting image features comprises extracting intrinsic properties of a camera used to collect the two- dimensional images.

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

4. The computer-implemented method of any preceding claim, wherein projecting image features along a direction between a three-dimensional model of the subject and a selected observation point comprises aggregating the image features for multiple pixels along the direction between the three-dimensional model of the subject and the selected observation point.

5. The computer-implemented method of any preceding claim, wherein projecting image features along a direction between a three-dimensional model of the subject and a selected observation point comprises concatenating multiple feature maps produced by each of multiple cameras in a permutation invariant combination, each of the multiple cameras having an intrinsic characteristic.

6. The computer-implemented method of any preceding claim, further comprising evaluating a loss function based on a difference between the 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.

7. The computer-implemented method of any preceding claim, wherein the subject is a user of a client device having a webcam directed to the user, further comprising identifying the selected observation point as a location of the webcam directed to the user from the client device.

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 a video with multiple images of the three-dimensional model of the subject to the network-coupled client device.

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

10. A system, comprising: a memory storing multiple instructions; and one or more processors configured to execute the 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 auto stereoscopic display in a virtual reality headset, comprising: collecting, from a face of multiple users, multiple ground-truth images; rectifying the ground-truth images with stored, calibrated stereoscopic pairs of images; generating, with a three-dimensional face model, multiple synthetic views of subjects, wherein the synthetic views of subjects include an interpolation of multiple feature maps projected along different directions corresponding to multiple views of the subjects; and training the three-dimensional face model based on a difference between the ground-truth images and the synthetic views of subjects.

12. The computer-implemented method of claim 11, wherein generating multiple synthetic views comprises projecting image features from each of the ground-truth images along a selected observation direction and concatenating multiple feature maps produced by each of the ground-truth images in a permutation invariant combination, each of the groundtruth images having an intrinsic characteristic.

13. The computer-implemented method of claim 11 or 12, wherein training the three-dimensional face model comprises updating at least one in a set of learnable weights for each of multiple features in the feature maps based on a value of a loss function indicative of the difference between the ground-truth images and the synthetic views 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 multiple pixels in the ground-truth images based on a pixel background value projected from the multiple ground-truth images; and/or generating a background model using specific features for each of multiple cameras collecting the multiple ground-truth images.

15. The computer-implemented method of any of claims 11 to 14, further comprising interpolating the feature maps by averaging multiple feature vectors from multiple cameras to form a camera summarized feature vector over different directions at a desired point.