Method, system and machine readable storage medium for neural image analysis

Abstract

Methods, systems, and machine-readable storage media for neural image analysis are provided. The method includes: receiving a two-dimensional (2D) input image depicting a person in a first pose; receiving a target image corresponding to the second posture; processing the 2D input image to generate soft feature pooling features; processing the target image to generate a target soft intrinsic distance; and decoding features of the 2D input image based on a combination of the soft feature pooling features and a target soft intrinsic distance to generate a decoded image depicting the person in the second pose.

Description

[0001] This application is a divisional application of the Chinese patent application with an application date of November 13, 2020, international application number PCT / EP2020 / 082047, invention name “Image Generation Using Surface-Based Neural Synthesis”, and application number 202080078647.X entering the Chinese national phase.

[0002] Priority claim

[0003] This application claims the benefit of priority to U.S. Provisional Application No. 62 / 936,328, filed on November 15, 2019, which is incorporated herein by reference in its entirety. Technical Field

[0004] The present disclosure relates to synthesizing images using continuous surface-level parameterization of objects. Background Art

[0005] Modern user devices provide messaging applications that allow users to exchange messages with each other. Such messaging applications have recently begun to incorporate graphics in such communications. Graphics may include avatars or cartoons that mimic the user's actions. BRIEF DESCRIPTION OF THE DRAWINGS

[0006] In the drawings, which are not necessarily drawn to scale, like numerals may describe similar components in different views. To easily identify the discussion of any particular element or action, the highest digit or digits in a reference numeral refer to the figure number in which the element is first introduced. Some non-limiting examples are shown in the figures of the accompanying drawings, in which:

[0007] Figure 1 is a diagrammatic representation of a networked environment in which the present disclosure may be deployed, according to some examples.

[0008] Figure 2 is a diagrammatic representation of a messaging system having both client-side and server-side functionality, according to some examples.

[0009] Figure 3 is a diagrammatic representation of data structures maintained in a database according to some examples.

[0010] Figure 4 is a diagrammatic representation of messages according to some examples.

[0011] Figure 5 and Figure 6 is a diagrammatic representation of operations performed by a gesture generation system, according to some examples.

[0012] Figure 7 Here is a diagrammatic representation of the diffusion and attention layers based on some examples.

[0013] Figure 8 is a graphical representation of statistical modeling based on the locations and features of some examples.

[0014] Figure 9 Diagrammatic representations of graphical user interfaces according to some examples are presented.

[0015] Figure 10A and Figure 10B is a flow diagram illustrating example operation of a messaging application server according to an example.

[0016] Figure 11 is a diagrammatic representation of a machine in the form of a computer system according to some examples, within which a set of instructions may be executed, for causing the machine to perform any one or more of the methodologies discussed herein.

[0017] Figure 12 is a block diagram illustrating a software architecture in which examples may be implemented. DETAILED DESCRIPTION

[0018] The following description includes systems, methods, techniques, instruction sequences, and computing machine program products for implementing the illustrative examples of the present disclosure. In the following description, for illustrative purposes, many specific details are set forth to provide an understanding of various examples. However, it will be apparent to those skilled in the art that examples can be practiced without these specific details. Generally, known instruction instances, protocols, structures, and techniques do not need to be shown in detail.

[0019] Neural image synthesis can be controlled by conditioning the network on a given signal, which can be a classification label, text, or layout constraints indicated by another image. Typical systems learn fully disentangled image synthesis by combining a 3D dataset and a 2D dataset or an unstructured 2D image set, allowing explicit control of camera, shape, and lighting parameters. When focusing on humans, the conditioning signal can include body keypoints, semantic parts, or DensePose level information. Typically, this is done by encoding this information as an additional input channel that is concatenated with the RGB image and fed to a CNN, or by using pose information in conjunction with a spatial transformer that densely warps the RGB values ​​or neuron activations.

[0020] The disclosed examples improve the accuracy of neural synthesis models by exploiting continuous, surface-level parameterization of object category shapes (with a particular focus on humans), and make such models more controllable. Specifically, according to the disclosed examples, a charted, UV coordinate-based representation of humans is used to improve image synthesis in terms of both controllability and realism. The charting is integrated into two complementary approaches to image synthesis: parametric generative models, such as principal component analysis (PCA) or autoencoders, in which explicit image encoding determines image synthesis; and descriptive models, in which images are synthesized by moment matching to be statistically indistinguishable from a target signal. As a parametric model, the disclosed examples use a semantic conditioning signal to modulate the behavior of the decoder through adaptive instance normalization. As a descriptive model, a general transfer method is used, which applies a coloring and whitening (CWT) transform to match the style signal and the Gram Matrix of the content.

[0021] Specifically, a 2D input image is received and features of the input image are obtained. These features are pooled and assigned to different channels based on region-specific appearance information (e.g., hair color and style around the face or display type around the feet). This information is compressed, and a membership-weighted estimate of the mean and variance is applied to each channel. As an example, each row of the matrix corresponds to a human joint. To construct the matrix, features defined for the image and each joint are obtained, and the disclosed example emphasizes those features that may include the joint in the corresponding matrix row. A target image is received, such as an image depicting different poses. The target image is processed by obtaining pixel values ​​for different regions of the image based on a dense pose function. As an example, a region of the target image corresponding to a shoulder landmark is analyzed to obtain a set of pixel values ​​associated with that region of the target image. These pixel values ​​are designated as target soft intrinsic distances. These soft intrinsic distances are concatenated with the pooled features of the input image and decoded to generate a decoded image in which the input image appears together with the features of the target image. For example, an input image depicting a person in one pose is decoded to generate an output image in which the person is depicted in another pose.

[0022] According to a first aspect, the present disclosure describes a computer-implemented neural image synthesis method comprising: receiving a two-dimensional continuous surface representation of a three-dimensional object, the continuous surface comprising a plurality of landmark locations; determining a first set of soft membership functions based on the landmark locations and the relative locations of points in the two-dimensional continuous surface representation; receiving a two-dimensional input image comprising an object of the same type as the three-dimensional object; extracting a plurality of features from the input image using a feature recognition model; generating an encoded feature representation of the extracted features using the first set of soft membership functions; generating a dense feature representation of the extracted features based on the encoded representation using a second set of soft membership functions; and processing the second set of soft membership functions and the dense feature representation using a neural image decoder model to generate an output image.

[0023] A two-dimensional continuous surface representation of a three-dimensional object comprises a mapping of the surface of the three-dimensional object onto a two-dimensional planar area. It may also be referred to as "charting". The representation is continuous in the sense that the two-dimensional representation of the object is not separated - that is, the different elements of the object are not divided into separate two-dimensional representations. In some examples, such a mapping is obtained by effectively "unfolding" and "flattening" the three-dimensional surface into two dimensions. This unfolding process may not completely fill the area used for the two-dimensional continuous surface representation - the remainder of the area may be referred to as "background". The two-dimensional continuous surface representation itself may be in the form of an image. An example of such a two-dimensional continuous surface representation is a UV map that can be generated from the three-dimensional object using UV unfolding techniques. Alternatively, other equivalent representations may be used.

[0024] Determining the first set of soft membership functions may include determining a distance between the landmark location and a plurality of points in the two-dimensional continuous surface representation; and assigning each point in the plurality of points to the landmark based on the determined distance.

[0025] A set of soft membership functions is a function that associates a point (e.g., a pixel) in a two-dimensional continuous surface representation with one or more of the landmarks. In practice, the soft membership function corresponds to an area that roughly locates a position on an object. A background membership function can also be included in a set that assigns points in the two-dimensional continuous surface representation that are determined not to be on an object to a background label. In some examples, the soft feature representation assigns each point to its nearest landmark. In other examples, the soft feature representation assigns a set of weights to each point, each weight being associated with a different landmark and based on the distance to the landmark, e.g., the greater the distance, the smaller the weight.

[0026] Determining landmark locations may include using a landmark recognition model.

[0027] Landmarks (also referred to as "object landmarks") can represent key points of an object. Landmarks can be specific to the object type. For example, in an example where the object is a human body, the landmarks can be key points of the human body, such as joints, facial features, etc. Object landmarks can be manually marked / located or can be marked / located using a landmark recognition / localization model (e.g., a neural network trained to locate landmarks in a two-dimensional continuous surface representation of a specific object type).

[0028] The neural image decoder model may include a convolutional neural network conditioned on a two-dimensional continuous surface representation.

[0029] Generating an encoded feature representation of the extracted features using the first set of soft membership functions includes performing a membership-weighted estimate of a mean and a variance for each channel of the extracted features. The encoded representation may include an estimate of a mean and a variance for each channel of the extracted features, i.e., feature statistics near a landmark.

[0030] Generating a dense feature representation of the extracted features from the encoded representation using the second set of soft membership functions may include applying a dual operation to the membership-weighted estimates of the mean and variance of each channel of the extracted features.

[0031] Dense feature representations can also be referred to as "feature fields" or "pixelated representations." Soft feature unpooling can be used to generate dense feature representations; it is the inverse / dual operation of soft feature pooling. Unpooling effectively distributes the features in the encoded feature representation over corresponding regions of the image. In other words, unpooling spreads the encoding back into the image domain.

[0032] In some examples, the first set of soft membership functions and the second set of soft membership functions are the same. The method may also include: generating a three-dimensional model / representation of the three-dimensional object based on the input image; and generating a two-dimensional continuous surface representation based on the three-dimensional model / representation. The method may also include: modifying values ​​in the encoded representation before generating the dense feature representation.

[0033] The first set of soft membership functions is reused to generate a dense feature representation that allows shape and appearance information in the input image to be disentangled, thereby providing a means for independently controlling / changing shape and appearance during image generation.

[0034] A three-dimensional model / representation, such as a DensePose representation, can be determined from the input image and then used to generate a two-dimensional continuous surface representation, such as using UV unwrapping. The two-dimensional continuous surface representation then corresponds to the input image.

[0035] A two-dimensional continuous surface representation of the three-dimensional object may be generated from the input image. The method may further include: receiving another two-dimensional input image, the further input image including another object of the same type as the three-dimensional object; generating another two-dimensional continuous surface representation of the three-dimensional object from the further two-dimensional input image, the further continuous surface including a plurality of landmark locations; and determining a second set of soft membership functions based on the landmark locations and the relative locations of points in the further two-dimensional continuous surface representation.

[0036] Determining a first set of soft membership functions based on an input image and determining a second set of soft membership functions based on a different input image allows pose / style information to be transferred from one image to another. For example, a (first) input image may include an image of an object (e.g., a human) in a first pose and is used to determine the first set of soft membership functions. A second input image may include an image of an object (e.g., a human) in a second pose and is used to determine the second set of soft membership functions. Extracting features from the first image and generating an encoded representation of these features using the first set of soft membership functions can associate the features of the first image with object landmarks. Unpooling the encoded representation of the features of the first image using the second set of soft membership functions to generate a dense representation can actually transfer the features of the first image to an image having the pose of the second image.

[0037] Generating the further two-dimensional continuous surface representation from the further (ie second) image may be performed in the same way as generating the two-dimensional continuous surface representation from the input (ie first) image.

[0038] The input image may include an image of an object in a first pose, and the other input image may include an image of another object in a second pose. The second image may include a portion corresponding to an unseen portion of the first image. Generating a two-dimensional continuous surface representation from the input image may include generating a portion of the two-dimensional continuous surface representation corresponding to the unseen portion of the first image from the encoded representation using a learned attention mechanism. The learned attention mechanism may be based on a first set of soft membership functions.

[0039] When the input (i.e., first) image and the additional (i.e., second) image show objects in different poses, there may be areas of the second image that are not present in the first image and therefore cannot be directly migrated from the first image. These areas can be interpolated into the first image using a learned attention mechanism. A learned model such as a neural network or matrix with learned weights, parameters, and / or components can be used to diffuse the observed features (i.e., extracted features) of the first image across landmarks to generate a set of diffuse features. An attention mechanism is used to generate a set of refined features by combining the set of diffuse features with a set of observed features to prevent the set of diffuse features from overwriting the observed features.

[0040] According to another aspect, the present disclosure describes a computer-implemented style transfer method comprising: determining a set of content features based on a source image using an encoder neural network; determining a set of style features based on a style image using an encoder neural network; determining position-dependent content features using joint statistics of position and content features in a region of the source image; determining position-dependent style features using joint statistics of position and style features in a region of the style image; generating a set of transformed content features based on the set of position-dependent content features based on the joint statistics of position and content features; generating a set of transformed style features based on the set of transformed content features based on the joint statistics of position and style features; and generating an output image based on the set of transformed style features and the set of transformed content features using a decoder neural network.

[0041] The method can provide an enhancement to other descriptive methods such as whitening and colorization transformations by considering non-stationary patterns in the input image using position-dependent style features and position-dependent content features.

[0042] A feature recognition neural network can be used to determine content features and style features from the source image and style image, respectively. An example of such a network is the VGG network, but other feature recognition networks can alternatively be used. The decoder neural network can be a neural network trained to reproduce an image from the feature map produced by the feature recognition neural network. The feature recognition neural network and the decoder neural network can together form an autoencoder system.

[0043] As described above in relation to the first aspect, the style image and / or the source image comprises a continuous two-dimensional representation of a three-dimensional object.

[0044] The joint statistics of the position and content features include a content feature mean, a content position mean, and a covariance between the content feature and the content position, and wherein determining the position-dependent content feature includes determining a conditional model of the content feature conditioned on the position. The conditional model of the content feature may include a position-dependent content mean and a conditional content covariance. Generating a set of transformed content features from the set of position-dependent content features may include centering the position-dependent content features based on the position-dependent content mean and applying a whitening transform based on the conditional content covariance.

[0045] The joint statistics of the position and style features may include a style feature mean, a style position mean, and a covariance between the style feature and the style position, and determining the position-dependent style feature includes determining a conditional model of the style feature conditioned on the position. The conditional model of the style feature may include a position-dependent style mean and a conditional style covariance. Generating a set of transformed style features from a set of position-dependent content features may include adding the position-dependent content features to the position-dependent style mean and applying a shading transformation based on the conditional style covariance.

[0046] Models such as multivariate Gaussian models can be used to capture the dependency of extracted content / style features on continuous position coordinates. This model can be used to replace static content / style features used in other descriptive synthesis models, such as whitening and colorization transformations. Whitening refers to reducing the style features present in the content features. Colorization refers to adding style features from the style image to the (whitened) content features of the source image.

[0047] The method may further include: mapping the source / style image to an embedding using the trained model; and determining joint statistics of location and content / style features in a region of the source / style image based on the embedding.

[0048] Using embeddings of position coordinates rather than the position coordinates themselves allows for capturing complex spatial dependencies. For example, in the human body, mirror symmetry is expected to exist about the vertical axis rather than the horizontal axis. Therefore, features will be more correlated in the horizontal direction than in the vertical direction. The model can be trained based on a loss function that penalizes distances in the vertical direction more than in the horizontal direction when generating a mapping.

[0049] Networked computing environment

[0050] Figure 1 1 is a block diagram illustrating an example messaging system 100 for exchanging data (e.g., messages and associated content) over a network. The messaging system 100 includes multiple instances of a client device 102, each of which hosts a number of applications including a messaging client 104 and other external applications 109 (e.g., third-party applications). Each messaging client 104 is communicatively coupled to other instances of the messaging client 104 (e.g., hosted on respective other client devices 102), a messaging server system 108, and external application servers 110 via a network 112 (e.g., the Internet). The messaging client 104 can also communicate with locally hosted third-party applications 109 using an application programming interface (API).

[0051] The messaging clients 104 are able to communicate and exchange data with other messaging clients 104 and messaging server systems 108 via the network 112. The data exchanged between the messaging clients 104 and between the messaging clients 104 and messaging server systems 108 includes functions (e.g., commands for initiating functions) and payload data (e.g., text, audio, video, or other multimedia data).

[0052] The messaging server system 108 provides server-side functionality to certain messaging clients 104 via the network 112. Although certain functionality of the messaging system 100 is described herein as being performed by the messaging client 104 or by the messaging server system 108, the location of certain functionality within the messaging client 104 or the messaging server system 108 may be a design choice. For example, it may be technically preferable to initially deploy certain technologies and functionality within the messaging server system 108 but later migrate the technologies and functionality to the messaging client 104 if the client device 102 has sufficient processing power.

[0053] The messaging server system 108 supports various services and operations provided to the messaging clients 104. Such operations include sending data to the messaging clients 104, receiving data from the messaging clients 104, and processing data generated by the messaging clients 104. The data may include message content, client device information, geolocation information, media enhancements and overlays, message content persistence conditions, social network information, and live event information as examples. The data exchange within the messaging system 100 is initiated and controlled by functionality available through the user interface (UI) of the messaging clients 104.

[0054] Turning now specifically to the messaging server system 108, an application program interface (API) server 116 is coupled to the application server 114 and provides a programming interface to the application server 114. The application server 114 is communicatively coupled to a database server 120, which facilitates access to a database 126 that stores data associated with messages processed by the application server 114. Similarly, a web server 128 is coupled to the application server 114 and provides a web-based interface to the application server 114. To this end, the web server 128 handles incoming network requests via the Hypertext Transfer Protocol (HTTP) and several other related protocols.

[0055] The application program interface (API) server 116 receives and sends message data (e.g., commands and message payloads) between the client device 102 and the application server 114. In particular, the application program interface (API) server 116 provides a set of interfaces (e.g., routines and protocols) that the messaging client 104 can call or query to invoke functionality of the application server 114. An application program interface (API) server 116 exposes various functions supported by the application server 114, including: account registration; login functionality; sending messages from a particular messaging client 104 to another messaging client 104 via the application server 114; sending media files (e.g., images or videos) from the messaging client 104 to the messaging server 118; and setting up collections of media data (e.g., stories) for access by another messaging client 104, retrieving a friend list of a user of the client device 102, retrieving such collections, retrieving messages and content, adding and removing entities (e.g., friends) to and from an entity graph (e.g., a social graph), locating friends in a social graph, and opening application events (e.g., related to a messaging client 104).

[0056] The application server 114 hosts several server applications and subsystems, including, for example, a messaging server 118, an image processing server 122, and a social network server 124. The messaging server 118 implements several message processing technologies and functions that, in particular, relate to the aggregation and other processing of content (e.g., text and multimedia content) included in messages received from multiple instances of the messaging client 104. As will be described in further detail, text and media content from multiple sources can be aggregated into content collections (e.g., called stories or galleries). These collections are then made available to the messaging client 104. Given the hardware requirements for other processor- and memory-intensive processing of data, such processing can also be performed on the server side by the messaging server 118.

[0057] The application server 114 also includes an image processing server 122 that is dedicated to performing various image processing operations, typically with respect to images or videos within the payload of messages sent from or received at the messaging server 118. Figure 5 The detailed functions of the image processing server 122 are shown and described. The image processing server 122 is used to implement the 3D body model generation system 230 ( Figure 2 ) 3D body model generation operation.

[0058] In one example, the image processing server 122 detects a person in an input 2D image. The image processing server 122 also receives a target image of the person or another person that presents a different pose than the person in the 2D image. The image processing server 122 uses the target image to generate an output image that depicts the person in the input 2D image presenting the pose of the person in the target image. In some cases, the image processing server 122 uses adaptive instance normalization (CHAIN) of graphics to perform this pose transfer. Specifically, adaptive instance normalization (AdaIN) modifies the statistics of each channel c in the feature map through a parameter function such as a multilayer perceptron:

[0059]

[0060] Among them, x i represents the activation at position i, μ, σ are calculated by standard instance normalization, and γ, β are the multiplicative gain terms and additive gain terms predicted by the side branches to appropriately modify the behavior of the network. AdaIN is applied to multiple stages of the decoder and shows that the values ​​of γ, β at different network depths provide a natural disentanglement of the structural hierarchy. The image processing server 122 according to some examples determines the spatially varying instance normalization parameters γ, β modulated by the continuous surface representation. Specifically, CHAIN ​​uses both the input image and the surface-based interpretation to construct the conditioning signal:

[0061]

[0062] The conditioning signal is designed to disentangle shape and appearance. This allows the image processing server 122 to synthesize the pose of a person with the clothing of another person, or to easily perform appearance inpainting, in a second stage. The conditioning signal is constructed by first deriving local shape and appearance descriptors and then fusing them into a dense conditioning signal consisting of c for each channel. i , b i Perform regression CNN processing. Figure 5 The training of the CHAIN ​​system implemented by the image processing server 122 is discussed.

[0063] The social network server 124 supports various social networking functions and services and makes these functions and services available to the messaging server 118. To this end, the social network server 124 maintains and accesses an entity graph 308 (e.g., Figure 3 Examples of functions and services supported by the social network server 124 include identifying other users in the messaging system 100 who have relationships with or are "following" a particular user, and identifying interests and other entities of a particular user.

[0064] Returning to the messaging client 104, the features and functionality of the external resource (e.g., a third-party application 109 or applet) are made available to the user via the interface of the messaging client 104. The messaging client 104 receives a user selection of an option to launch or access features of an external resource (e.g., a third-party resource) (e.g., an external app 109). The external resource can be a third-party application (external app 109) installed on the client device 102 (e.g., a "local app"), or a small-scale version of a third-party application (e.g., a "mini-app") hosted on the client device 102 or remote from the client device 102 (e.g., on a third-party server 110). The small-scale version of the third-party application includes a subset of the features and functionality of the third-party application (e.g., a full-scale local version of a third-party standalone application) and is implemented using a markup language document. In one example, the small-scale version of the third-party application (e.g., a "mini-app") is a web-based markup language version of the third-party application and is embedded in the messaging client 104. In addition to using markup language documents (eg, .*ml files), applets may also incorporate scripting languages ​​(eg, .*js files or .json files) and style sheets (eg, .*ss files).

[0065] In response to receiving a user selection of an option to launch or access a feature of an external resource (external app 109), the messaging client 104 determines whether the selected external resource is a web-based external resource or a locally installed external application. In some cases, an external application 109 installed locally on the client device 102 can be launched independently of and separately from the messaging client 104, for example, by selecting an icon corresponding to the external application 109 on a home screen of the client device 102. A small-scale version of such an external application can be launched or accessed via the messaging client 104, and in some examples, no portion or a limited portion of the small-scale external application can be accessed from outside the messaging client 104. The small-scale external application can be launched by the messaging client 104 receiving a markup language document associated with the small-scale external application from the external application server 110 and processing such a document.

[0066] In response to determining that the external resource is a locally installed external application 109, the messaging client 104 instructs the client device 102 to launch the external application 109 by executing locally stored code corresponding to the external application 109. In response to determining that the external resource is a web-based resource, the messaging client 104 communicates with the external application server 110 to obtain a markup language document corresponding to the selected resource. The messaging client 104 then processes the obtained markup language document to present the web-based external resource within the user interface of the messaging client 104.

[0067] The messaging client 104 can notify the user of the client device 102 or other users associated with such user (e.g., "friends") of activities occurring in one or more external resources. For example, the messaging client 104 can provide participants in a conversation (e.g., a chat session) within the messaging client 104 with notifications regarding current or recent use of an external resource by one or more members of the user group. One or more users can be invited to join an active external resource or to launch an external resource that was recently used but is currently inactive (in a friend group). The external resource can provide participants in the conversation (each using a corresponding messaging client 104) with the ability to share items, conditions, states, or locations within the external resource with one or more members of the user group entering the chat session. The shared items can be interactive chat cards that members of the chat can interact with, for example, to launch a corresponding external resource, view specific information within the external resource, or be directed to a specific location or state within the external resource. Within a given external resource, a response message can be sent to the user on the messaging client 104. The external resource may selectively include different media items in the response based on the current context of the external resource.

[0068] The messaging client 104 can present a list of available external resources (e.g., third-party or external applications 109 or applets) to the user to launch or access a given external resource. The list can be presented in a context-sensitive menu. For example, the icons representing different external applications in the external applications 109 (or applets) can change based on how the user launches the menu (e.g., from a conversational interface or from a non-conversational interface).

[0069] System Architecture

[0070] Figure 21 is a block diagram illustrating further details regarding the messaging system 100 according to some examples. Specifically, the messaging system 100 is shown as including a messaging client 104 and an application server 114. The messaging system 100 includes several subsystems that are supported on the client side by the messaging client 104 and on the server side by the application server 114. These subsystems include, for example, a transient timer system 202, a collection management system 204, an enhancement system 208, a mapping system 210, a gaming system 212, and an external resource system 220.

[0071] The transient timer system 202 is responsible for implementing temporary or time-limited access to content by the messaging client 104 and the messaging server 118. The transient timer system 202 includes several timers that selectively enable access (e.g., for presentation and display) of messages and associated content via the messaging client 104 based on duration and display parameters associated with the message or collection of messages (e.g., a story). Additional details regarding the operation of the transient timer system 202 are provided below.

[0072] The collection management system 204 is responsible for managing groups and collections of media (e.g., collections of text, image, video, and audio data). Collections of content (e.g., messages including images, videos, text, and audio) can be organized into "event galleries" or "event stories." Such collections can be made available for a specified time period, such as the duration of an event related to the content. For example, content related to a concert can be made available as a "story" for the duration of the concert. The collection management system 204 can also be responsible for publishing an icon to the user interface of the messaging client 104 that provides notification of the existence of a particular collection.

[0073] In addition, the collection management system 204 also includes a curation interface 206 that allows a collection manager to manage and curate specific content collections. For example, the curation interface 206 enables an event organizer to curate a content collection related to a specific event (e.g., removing inappropriate content or redundant messages). In addition, the collection management system 204 uses machine vision (or image recognition technology) and content rules to automatically curate content collections. In some examples, users can be paid compensation for including user-generated content in a collection. In such cases, the curation management system 204 operates to automatically pay such users for using their content.

[0074] The enhancement system 208 provides various functions that enable users to enhance (e.g., annotate or otherwise modify or edit) media content associated with a message. For example, the enhancement system 208 provides functions related to generating and publishing media overlays for messages processed by the messaging system 100. The enhancement system 208 is operable to supply media overlays or enhancements (e.g., image filters) to the messaging client 104 based on the geographic location of the client device 102. In another example, the enhancement system 208 is operable to supply media overlays to the messaging client 104 based on other information such as social network information of the user of the client device 102. Media overlays can include audio and visual content and visual effects. Examples of audio and visual content include pictures, text, logos, animations, and sound effects. Examples of visual effects include color overlays. Audio and visual content or visual effects can be applied to media content items (e.g., photos) at the client device 102. For example, a media overlay can include text, graphic elements, or images that can be overlaid on a photo taken by the client device 102. In another example, the media overlay includes a location identification overlay (e.g., Venice Beach), a live event name, or a business name overlay (e.g., Beach Cafe). In another example, the enhancement system 208 uses the geolocation of the client device 102 to identify a media overlay that includes the name of a business at the geolocation of the client device 102. The media overlay may include other tags associated with the business. The media overlay may be stored in the database 126 and accessed through the database server 120.

[0075] In some examples, the enhancement system 208 provides a user-based publishing platform that enables a user to select a geographic location on a map and upload content associated with the selected geographic location. The user can also specify the environment in which a particular media overlay should be provided to other users. The enhancement system 208 generates a media overlay including the uploaded content and associates the uploaded content with the selected geographic location.

[0076] In other examples, the augmentation system 208 provides a merchant-based publishing platform that enables merchants to select specific media overlays associated with a geographic location via a bidding process. For example, the augmentation system 208 associates the media overlay of the highest-bidding merchant with the corresponding geographic location for a predefined amount of time. The augmentation system 208 communicates with the image processing server 122 to automatically select and activate an augmented reality experience related to the image captured by the client device 102. Once the user selects the augmented reality experience when scanning an image using a camera in the user environment, one or more images, videos, or augmented reality graphic elements are retrieved and presented as an overlay on the scanned image. In some cases, the camera is switched to a frontal perspective (e.g., the front-facing camera of the client device 102 is activated in response to activation of a specific augmented reality experience), and an image from the front-facing camera of the client device 102, rather than the rear-facing camera of the client device 102, begins to be displayed on the client device 102. One or more images, videos, or augmented reality graphic elements are retrieved and presented as an overlay on the image captured and displayed by the front-facing camera of the client device 102.

[0077] The mapping system 210 provides various geolocation capabilities and supports the presentation of map-based media content and messages by the messaging client 104. For example, the mapping system 210 enables the display of user icons or avatars (e.g., stored in the profile data 316) on a map to indicate the current or past locations of the user's "friends" in the context of the map, as well as media content generated by such friends (e.g., a collection of messages including photos and videos). For example, a message posted by a user to the messaging system 100 from a particular geolocation can be displayed to the particular user's "friends" at that particular location in the context of the map on the map interface of the messaging client 104. A user can also share his or her location and status information with other users of the messaging system 100 via the messaging client 104 (e.g., using an appropriate status avatar), where the location and status information is similarly displayed to selected users in the context of the map interface of the messaging client 104.

[0078] The gaming system 212 provides various gaming functions within the context of the messaging client 104. The messaging client 104 provides a gaming interface that provides a list of available games (e.g., web-based games or web-based applications) that can be launched by a user within the context of the messaging client 104 and played with other users of the messaging system 100. The messaging system 100 also enables a particular user to invite other users to play a particular game by sending an invitation to such other users from the messaging client 104. The messaging client 104 also supports both voice messaging and text messaging (e.g., chat) within the context of game play, provides leaderboards for games, and also supports the provision of in-game rewards (e.g., coins and items).

[0079] The external resource system 220 provides an interface for the messaging client 104 to communicate with the external application server 110 to launch or access external resources. Each external resource (app) server 110 hosts, for example, an application based on a markup language (e.g., HTML5) or a small-scale version of an external application (e.g., a game, utility, payment, or ride-sharing application external to the messaging client 104). The messaging client 104 can launch a web-based resource (e.g., an application) by accessing an HTML5 file from the external resource (app) server 110 associated with the web-based resource. In some examples, the application hosted by the external resource server 110 is programmed in JavaScript using a software development kit (SDK) provided by the messaging server 118. The SDK includes an application programming interface (API) with functions that can be called or activated by a web-based application. In some examples, the messaging server 118 includes a JavaScript library that provides access to a given third-party resource for certain user data of the messaging client 104. HTML5 is used as an example technology for programming games, but applications and resources programmed based on other technologies can be used.

[0080] To integrate the SDK's functionality into a web-based resource, the SDK is downloaded from the messaging server 118 by the external resource (app) server 110 or received in other ways by the external resource (app) server 110. Once downloaded or received, the SDK is included as part of the application code of the web-based external resource. The code of the web-based resource can then call or activate certain functions of the SDK to integrate the features of the messaging client 104 into the web-based resource.

[0081] The SDK stored on the messaging server 118 effectively provides a bridge between external resources (e.g., a third party or external application 109 or applet and the messaging client 104). This provides users with a seamless experience of communicating with other users on the messaging client 104 while also preserving the look and feel of the messaging client 104. In order to bridge the communication between the external resources and the messaging client 104, in some examples, the SDK facilitates communication between the external resource server 110 and the messaging client 104. In some examples, the WebViewJavaScriptBridge running on the client device 102 establishes two one-way communication channels between the external resources and the messaging client 104. Messages are sent asynchronously between the external resources and the messaging client 104 via these communication channels. Each SDK function call is sent as a message and a callback. Each SDK function is implemented by building a unique callback identifier and sending a message with the callback identifier.

[0082] By using the SDK, not all information from the messaging client 104 is shared with the external resource server 110. The SDK limits which information is shared based on the needs of the external resource. In some examples, each external resource server 110 provides an HTML5 file corresponding to a web-based external resource to the messaging server 118. The messaging server 118 can add a visual representation (e.g., box art or other graphics) of the web-based external resource in the messaging client 104. Once the user selects the visual representation through the GUI of the messaging client 104 or instructs the messaging client 104 to access a feature of the web-based external resource, the messaging client 104 obtains the HTML5 file and instantiates the resources required to access the feature of the web-based external resource.

[0083] The messaging client 104 presents a graphical user interface (e.g., a login page or title screen) of the external resource. During, before, or after presenting the login page or title screen, the messaging client 104 determines whether the external resource launched has previously been authorized to access the user data of the messaging client 104. In response to determining that the external resource launched has previously been authorized to access the user data of the messaging client 104, the messaging client 104 presents another graphical user interface of the external resource including the functions and features of the external resource. In response to determining that the external resource launched has previously not been authorized to access the user data of the messaging client 104, after a threshold time period (e.g., 3 seconds) of displaying the login page or title screen of the external resource, the messaging client 104 slides upward a menu for authorizing the external resource to access the user data (e.g., animating the menu to appear from the bottom of the screen to the middle or other portion of the screen). The menu identifies the type of user data that the external resource will be authorized to use. In response to receiving the user selection of the accept option, messaging client 104 adds the external resource to the list of authorized external resources and allows the external resource to access the user data from messaging client 104. In some examples, the external resource is authorized by messaging client 104 to access the user data according to the OAuth 2 framework.

[0084] The messaging client 104 controls the type of user data shared with the external resource based on the type of external resource that is authorized. For example, an external resource comprising a full-scale external application (e.g., a third party or external application 109) is provided with access to a first type of user data (e.g., only a two-dimensional avatar of the user with or without different avatar characteristics). As another example, an external resource comprising a small-scale version of an external application (e.g., a web-based version of a third-party application) is provided with access to a second type of user data (e.g., payment information, a two-dimensional avatar of the user, a three-dimensional avatar of the user, and avatars with various avatar characteristics). Avatar characteristics include different ways to customize the look and feel of an avatar, such as different poses, facial features, clothing, etc.

[0085] The pose generation system 230 generates and implements a CHAIN ​​model to synthesize a person's pose with another person's clothing and / or synthesize a person's pose in an input image with a different pose depicted in a target image.

[0086] Data Architecture

[0087] Figure 3is a diagram illustrating a data structure 300 that may be stored in the database 126 of the messaging server system 108, according to certain examples. Although the contents of the database 126 are shown as including several tables, it should be understood that the data may be stored in other types of data structures (e.g., as an object-oriented database).

[0088] The database 126 includes message data stored in the message table 302. For any particular message, the message data includes at least message sender data, message recipient (or receiver) data, and payload. Figure 4 Additional details regarding information that may be included in a message and included in the message data stored in message table 302 are described.

[0089] The entity table 306 stores entity data and is linked (e.g., by reference) to the entity graph 308 and profile data 316. Entities for which records are maintained within the entity table 306 may include individuals, corporate entities, organizations, objects, places, events, and the like. Regardless of the entity type, any entity for which the messaging server system 108 stores data may be an identified entity. Each entity is provided with a unique identifier and an entity type identifier (not shown).

[0090] The entity graph 308 stores information about relationships and associations between entities. Such relationships may be, for example, simply social, professional (e.g., working in a common company or organization), interest-based, or activity-based.

[0091] The profile data 316 stores various types of profile data about a particular entity. Based on the privacy settings specified by the particular entity, the profile data 316 can be selectively used and presented to other users of the messaging system 100. In the case where the entity is a person, the profile data 316 includes, for example, a user name, phone number, address, settings (e.g., notification and privacy settings), and an avatar representation (or a collection of such avatar representations) selected by the user. The particular user can then selectively include one or more of these avatar representations in the content of messages transmitted via the messaging system 100 and on a map interface displayed to other users by the messaging client 104. The collection of avatar representations can include a "status avatar," which presents a graphical representation of a status or activity that the user may choose to convey at a particular time.

[0092] Where the entity is a group, the group's profile data 316 may similarly include one or more avatar representations associated with the group, in addition to the group name, members, and various settings related to the group (eg, notifications).

[0093] The database 126 also stores enhancement data, such as overlays or filters, in an enhancement table 310. The enhancement data is associated with and applied to videos (video data is stored in the video table 304) and images (image data is stored in the image table 312).

[0094] In one example, a filter is an overlay that is displayed over an image or video during presentation to a receiving user. Filters can be of various types, including filters that a user selects from a set of filters presented to a sending user by messaging client 104 while the sending user is composing a message. Other types of filters include geolocation filters (also known as geofilters), which can be presented to a sending user based on geolocation. For example, a geolocation filter specific to a nearby or special location can be presented within a user interface by messaging client 104 based on geolocation information determined by a global positioning system (GPS) unit of client device 102.

[0095] Another type of filter is a data filter, which can be selectively presented to the sending user by the messaging client 104 based on other input or information collected during the message creation process by the client device 102. Examples of data filters include the current temperature at a particular location, the current speed the sending user is traveling, the battery life of the client device 102, or the current time.

[0096] Other augmented data that can be stored in the image table 312 includes augmented reality content items (e.g., corresponding to application lenses or augmented reality experiences). Augmented reality content items can be real-time special effects and sounds that can be added to images or videos. Each augmented reality experience can be associated with one or more tagged images. In some examples, when a tagged image is determined to match a query image received from the client device 102, the corresponding augmented reality experience (e.g., augmented data) of the tagged image is retrieved from the image table 312 and provided to the client device 102.

[0097] As described above, augmented data includes augmented reality content items, overlays, image transformations, AR images, and similar terms that refer to modifications that can be applied to image data (e.g., video or images). This includes real-time modifications that modify the image as it is captured using the device sensors (e.g., one or more cameras) of the client device 102 and then displayed on the screen of the client device 102 with the modifications. This also includes modifications to stored content (e.g., video clips in a gallery that can be modified). For example, in a client device 102 with access rights to multiple augmented reality content items, a user can use a single video clip with multiple augmented reality content items to see how different augmented reality content items will modify the stored clip. For example, by selecting different augmented reality content items for the same content, multiple augmented reality content items that apply different pseudo-random motion models can be applied to the same content. Similarly, real-time video capture can be used with the modifications shown to show how the video image currently captured by the sensor of the client device 102 will modify the captured data. Such data may simply be displayed on the screen without being stored in memory, or content captured by the device sensors may be recorded and stored in memory with or without modification (or both). In some systems, a preview feature may show how different augmented reality content items will look in different windows on the display simultaneously. For example, this may enable viewing multiple windows with different pseudo-random animations on the display simultaneously.

[0098] Thus, using data from an augmented reality content item and various systems or other such transformation systems that use this data to modify content can involve detecting objects (e.g., faces, hands, bodies, cats, dogs, surfaces, objects, etc.), tracking such objects as they leave, enter, and move around the field of view in a video frame, and modifying or transforming such objects as they are tracked. In various examples, different methods for implementing such transformations can be used. Some examples can involve generating a three-dimensional mesh model of one or more objects and using transformations of the model and animated textures within the video to implement the transformations. In other examples, tracking of points on an object can be used to place an image or texture (which can be two-dimensional or three-dimensional) at the tracked location. In yet further examples, neural network analysis of video frames can be used to place an image, model, or texture within content (e.g., an image or frame of a video). Thus, an augmented reality content item refers to both the images, models, and textures used to create the transformations within the content, as well as the additional modeling and analysis information required to implement such transformations using object detection, tracking, and placement.

[0099] Real-time video processing can be performed using any type of video data (e.g., video streams, video files, etc.) stored in the memory of any type of computerized system. For example, a user can load a video file and store it in the device's memory, or a video stream can be generated using the device's sensors. Furthermore, computer animation models can be used to process any object, such as human faces and body parts, animals, or inanimate objects (e.g., chairs, cars, or other objects).

[0100] In some examples, when a specific modification is selected along with the content to be transformed, the elements to be transformed are identified by a computing device. If the elements to be transformed are present in a frame of a video, the elements to be transformed are detected and tracked. Elements of an object are modified according to the modification request, thereby transforming the frame of the video stream. Transforming the frames of the video stream can be performed using different methods for different types of transformations. For example, for frame transformations primarily involving variations of elements of an object, feature points are calculated for each element of the object (e.g., using an active shape model (ASM) or other known methods). A mesh based on the feature points is then generated for each of at least one element of the object. This mesh is used in subsequent stages of tracking the elements of the object in the video stream. During the tracking process, the mesh referenced for each element is aligned with the position of each element. Additional points are then generated on the mesh. A first set of first points is generated for each element based on the modification request, and a second set of points is generated for each element based on the set of first points and the modification request. The frame of the video stream can then be transformed by modifying the elements of the object based on the set of first and second points and the mesh. In such methods, the background of the modified object can also be altered or distorted by tracking and modifying the background.

[0101] In some examples, a transformation that changes some areas of an object using the elements of the object can be performed by calculating feature points for each element of the object and generating a grid based on the calculated feature points. Points are generated on the grid, and then various areas based on these points are generated. The elements of the object are then tracked by aligning the area of ​​each element with the position of each of at least one element, and the properties of the area can be modified based on the modification request, thereby transforming the frame of the video stream. Depending on the specific modification request, the properties of the mentioned area can be transformed in different ways. Such modifications can involve: changing the color of the area; removing at least part of the area from the frame of the video stream; including one or more new objects in the area based on the modification request; and modifying or distorting elements of the area or object. In various examples, any combination of such modifications or other similar modifications can be used. For certain models to be animated, some feature points can be selected as control points to be used to determine the entire state space of options for model animation.

[0102] In some examples of computer animation models that use face detection to transform image data, faces are detected on an image using a specific face detection algorithm (e.g., Viola-Jones). An active shape model (ASM) algorithm is then applied to the facial region of the image to detect facial feature reference points.

[0103] Other methods and algorithms suitable for face detection can be used. For example, in some examples, features are located using landmarks, which represent distinguishable points that are present in most of the images considered. For example, for facial landmarks, the location of the left eye pupil can be used. If the initial landmarks are not recognizable (for example, if the person has an eye patch), secondary landmarks can be used. Such a landmark identification process can be used for any such object. In some examples, a set of landmarks forms a shape. The shape can be represented as a vector using the coordinates of the points in the shape. One shape is aligned with another shape using a similarity transformation (allowing translation, scaling, and rotation) that minimizes the average Euclidean distance between the shape points. The average shape is the average of the aligned training shapes.

[0104] In some examples, a landmark search begins with an average shape aligned with the position and size of a face determined by a global face detector. This search then iterates through the steps of proposing tentative shapes by adjusting the positioning of shape points through template matching of the image texture surrounding each point, and then fitting the tentative shapes to a global shape model until convergence occurs. In some systems, individual template matches are unreliable, and the shape model pools the results of weak template matches to form a stronger overall classifier. This entire search is repeated at each level of the image pyramid, from coarse to fine resolution.

[0105] The transformation system can capture an image or video stream on a client device (e.g., client device 102) and perform complex image manipulations locally on the client device 102 while maintaining an appropriate user experience, computational time, and power consumption. Complex image manipulations can include size and shape changes, emotion transfer (e.g., changing a face from a frown to a smile), state transfer (e.g., aging a subject, reducing apparent age, changing gender), style transfer, application of graphical elements, and any other suitable image or video manipulations enabled by a convolutional neural network that has been configured to execute efficiently on the client device 102.

[0106] In some examples, a computer-animated model for transforming image data can be used by a system in which a user can capture an image or video stream of the user (e.g., a selfie) using a client device 102 having a neural network operating as part of a messaging client 104 operating on the client device 102. A transformation system operating within the messaging client 104 determines the presence of a face within the image or video stream and provides a modification icon associated with the computer-animated model to transform the image data, or the computer-animated model can be presented in association with an interface described herein. The modification icon includes changes that can be the basis for modifying the user's face within the image or video stream as part of the modification operation. Once the modification icon is selected, the transformation system initiates a process of transforming the user's image to reflect the selected modification icon (e.g., generating a smiley face on the user). Once the image or video stream is captured and the specified modification is selected, the modified image or video stream can be presented in a graphical user interface displayed on the client device 102. The transformation system can implement a complex convolutional neural network on a portion of the image or video stream to generate and apply the selected modification. That is, once a modification icon is selected, the user can capture an image or video stream and be presented with the modified results in real time or near real time. Furthermore, when the video stream is being captured, the modification can be persistent and the selected modification icon remains toggled. Machine-taught neural networks can be used to implement such modifications.

[0107] The graphical user interface presenting the modifications performed by the transformation system can provide the user with additional interactive options. Such options can be based on the interface used to initiate content capture and selection of a particular computer animation model (e.g., initiated from a content creator user interface). In various examples, the modifications can be persistent after the initial selection of the modification icon. The user can turn the modifications on or off by tapping or otherwise selecting a face modified by the transformation system, and store it for later review or browsing to other areas of the imaging application. In the case of multiple faces being modified by the transformation system, the user can globally turn the modifications on or off by tapping or selecting a single face modified and displayed within the graphical user interface. In some examples, each face in a group of multiple faces can be modified individually, or such modifications can be individually toggled by tapping or selecting each face or a series of faces displayed within the graphical user interface.

[0108] The story table 314 stores data about a collection of messages and associated image, video, or audio data, which are compiled into a collection (e.g., a story or gallery). The creation of a particular collection can be initiated by a particular user (e.g., each user whose record is maintained in the entity table 306). A user can create a "personal story" in the form of a collection of content that has been created and sent / broadcasted by the user. To this end, the user interface of the messaging client 104 may include a user-selectable icon that enables the sending user to add specific content to his or her personal story.

[0109] A collection may also constitute a "live story" which is a collection of content from multiple users that is created manually, automatically, or using a combination of manual and automatic techniques. For example, a "live story" may constitute a curated stream of user-submitted content from various locations and events. A user whose client device has location services enabled and who is at a public location event at a particular time may be presented with an option, for example, via a user interface of the messaging client 104, to contribute content to a particular live story. Live stories may be identified to a user by the messaging client 104 based on his or her location. The end result is a "live story" told from a community perspective.

[0110] Another type of content collection is called a "location story," which enables users whose client devices 102 are located within a specific geographic location (e.g., on a college or university campus) to contribute to a particular collection. In some embodiments, contributions to location stories may require a second degree of authentication to verify that the end user belongs to a specific organization or other entity (e.g., is a student on a university campus).

[0111] As mentioned above, video table 304 stores video data, which, in one example, is associated with messages whose records are maintained within message table 302. Similarly, image table 312 stores image data associated with messages whose message data is stored in entity table 306. Entity table 306 may associate various enhancements from enhancement table 310 with the various images and videos stored in image table 312 and video table 304.

[0112] Data communication architecture

[0113] Figure 4is a schematic diagram illustrating the structure of a message 400 according to some examples, the message 400 being generated by the messaging client 104 for transmission to another messaging client 104 or the messaging server 118. The content of a particular message 400 is used to populate a message table 302 stored within the database 126, which is accessible by the messaging server 118. Similarly, the content of the message 400 is stored in memory as "in-flight" or "in-transit" data of the client device 102 or application server 114. The message 400 is shown as including the following example components:

[0114] Message identifier 402 : A unique identifier that identifies the message 400 .

[0115] Message text payload 404 : Text to be generated by the user via the user interface of the client device 102 and included in the message 400 .

[0116] • Message image payload 406: Image data captured by the camera component of the client device 102 or retrieved from the memory component of the client device 102 and included in the message 400. The image data for a sent or received message 400 may be stored in the image table 312.

[0117] • Message video payload 408: Video data captured by the camera component or retrieved from the memory component of the client device 102 and included in the message 400. The video data for a message 400 sent or received may be stored in the video table 304.

[0118] • Message audio payload 410 : audio data captured by a microphone or retrieved from a memory component of the client device 102 and included in the message 400 .

[0119] Message enhancement data 412: Enhancement data (e.g., filters, stickers, or other annotations or enhancements) representing enhancements to be applied to the message image payload 406, message video payload 408, or message audio payload 410 of the message 400. Enhancement data for a sent or received message 400 may be stored in the enhancement table 310.

[0120] • Message duration parameter 414: A parameter value indicating the amount of time, in seconds, that the content of the message (e.g., message image payload 406, message video payload 408, message audio payload 410) is to be presented to or accessible to the user via the messaging client 104.

[0121] Message geolocation parameters 416: Geolocation data (e.g., latitude and longitude coordinates) associated with the content payload of the message. Multiple message geolocation parameter 416 values ​​may be included in the payload, with each of these parameter values ​​being associated with a content item included in the content (e.g., a specific image in the message image payload 406, or a specific video in the message video payload 408).

[0122] Message story identifier 418: An identifier value that identifies one or more content collections (e.g., the "stories" identified in story table 314) with which a particular content item in message image payload 406 of message 400 is associated. For example, an identifier value can be used to associate multiple images within message image payload 406 with each of multiple content collections.

[0123] Message tags 420: Each message 400 may be tagged with a plurality of tags, each of which indicates the subject matter of the content included in the message payload. For example, where a particular image included in the message image payload 406 depicts an animal (e.g., a lion), a tag value may be included within the message tags 420 indicating the relevant animal. Tag values ​​may be manually generated based on user input, or may be automatically generated using, for example, image recognition.

[0124] • Message sender identifier 422: An identifier (eg, a messaging system identifier, an email address, or a device identifier) ​​that indicates the user of the client device 102 on which the message 400 was generated and from which the message 400 was sent.

[0125] • Message recipient identifier 424: Indicates an identifier (eg, a messaging system identifier, email address, or device identifier) ​​of the user of the client device 102 to which the message 400 is addressed.

[0126] The content (e.g., value) of each component of message 400 may be a pointer to a location in a table where the content data value is stored. For example, the image value in message image payload 406 may be a pointer to a location in image table 312 (or the address of a location in image table 308). Similarly, the value in message video payload 408 may point to data stored in video table 304, the value stored in message enhancement data 412 may point to data stored in enhancement table 310, the value stored in message story identifier 418 may point to data stored in story table 314, and the values ​​stored in message sender identifier 422 and message recipient identifier 424 may point to user records stored in entity table 306.

[0127] Figure 55 is a diagrammatic representation of operations 500 performed by pose generation system 230 according to some examples. Specifically, operations 500 represent how a CHAIN ​​implementation of pose generation system 230 may be trained. To train pose generation system 230, an input image 520 is received, which may be a 2D image of a person in a given pose. A target image 510 may include an image that is the same as or similar to input image 520. A feature extractor 550 is applied to input image 520 to generate a set of soft feature pools 552. Soft feature pools 552 may be a matrix of feature sets extracted from input image 520 arranged based on the expected positioning of different joints.

[0128] As an example, the first row of the matrix may include a set of features corresponding to the shoulder joint. Features located closer to the expected location of the shoulder joint are assigned a greater weight than features located farther away from the expected location of the shoulder joint. Surface coordinate representations 530 of different landmarks or joints are received and used to obtain features or pixel values ​​of the input image 520 and assign features or pixel values ​​based on the landmarks or joints. The surface coordinate representation 530 may be received as a two-dimensional continuous surface representation of a three-dimensional object and include multiple landmark locations. As an example, the ellipse-like region on the upper right side of the estimated soft intrinsic distance representation 540 corresponds to the upper right joint in the surface coordinate representation 530 and the shoulder of the input image 520 in the image coordinate representation. A weighted average of the pixel values ​​in this particular region is calculated and assigned to one or more rows of the matrix to generate the soft feature pooling 552. Thus, the system calculates an average over the entire region rather than looking at a specific location and assigning that location as a joint.

[0129] The pose generation system 230 uses dense pose analysis to locate regions within the human body. Within these regions, the pose generation system 230 averages the features for each region separately, and this average is used as the appearance code for the image 520. Each tile in the estimated soft intrinsic distances 540 (a first set of soft membership functions) represents the likelihood that a pixel value or feature of the input image 520 corresponds to a specific joint based on the surface coordinate representation 530. In some cases, the features for each joint or matrix row are unpooled by copying the average of the features across the entire image domain to generate soft feature unpooling 554. This creates a dense feature representation of the extracted features using a second set of soft membership functions. This averaging is guided by a magnitude determined by the map of estimated soft intrinsic distances 540. In effect, a given feature is copied and pasted everywhere, but scaled down in regions away from the joints. That is, for appearance adjustment, the pose generation system 230 uses the same P' soft membership function to compile image information into a compact, controllable appearance code, which is then broadcast back to the original image coordinates and concatenated to the soft membership functions.

[0130] In one example, given an RGB image, the pose generation system 230 uses DensePose to obtain its UV mapping and soft-assigns each object pixel to a control point based on the Euclidean distance between its respective UV values ​​(e.g., using a first set of soft membership functions to generate encoded features representing the extracted features). A channel for background pixels is introduced, resulting in a tensor of P'=P+1 soft membership functions or regions that roughly locate each pixel on the human body. This information allows CHAIN ​​to modulate the behavior of the network differently based on the specific body region corresponding to pixel i.

[0131] In some examples, for the collection phase, the pose generation system 230 processes the RGB image using a fully convolutional encoder that produces K = 32 eigenvalues ​​at its top layer for each of the P' membership functions, for a total of K x P' channels. Each of these channels encodes different region-specific appearance information (e.g., hair color and style around the face, or shoe type around the foot). This information is compressed region by region by performing a membership-weighted estimate of the mean and variance of each channel:

[0132]

[0133] where i refers to the neuron location, Indicates the soft membership function with respect to region p expectations, and C k，p is the response to the kth channel of region p. Both the feature mean and variance are encoded as they are commonly used to describe texture properties.

[0134] Gathering this information provides a compact, D = 2K × P'-dimensional code of the appearance in terms of feature statistics around a sparse set of control points. These can be swapped or attenuated between regions of different people, allowing controllable image synthesis, for example, by gradually interpolating between different features or clothing. By again relying on soft membership functions, this sparse representation is spread into a densely defined signal. In particular, the sparse code is distributed over the regions for which it is most responsible:

[0135]

[0136] Note that π p The function is defined over the entire image domain and effectively "smears" the kth channel of the pth region over the portion of the image it occupies while smoothly blending it with channels from neighboring regions. These operations provide an appearance conditioning signal and can be understood as implementing surface-driven image encoding, such as encoding geometry and appearance 560.

[0137] The estimated soft intrinsic distance 540 is concatenated with the soft feature unpooled features 554 to provide an encoded geometry and appearance representation 560. The encoded geometry and appearance representation 560 is applied to an Adain network 570 to generate an output image.

[0138] In an example, the feature extractor 550 is implemented as a neural network (e.g., a convolutional neural network) to generate soft feature pooling 552. The feature extractor 550 can be trained on a set of training images. For example, a first training image can be received and processed by the feature extractor 550. The feature extractor 550 uses the surface coordinate representation 530 to generate soft feature pooling 552. The soft feature pooling 552 is processed to generate soft feature de-pooling 554. The same first training image is also processed as a target image to obtain an estimated soft intrinsic distance 540. The encoded geometry and appearance 560 are generated and used by the Adain network 570 to generate an output image. The output image is compared with the first training image to calculate the loss. Based on the deviation and loss between the two images, the parameters of the feature extractor 550 are updated and used to process the second training image. Once the loss is within a threshold or the stopping criterion is met, the training ends. That is, during training, the same image is used to perform pooling and de-pooling of features.

[0139] Figure 6 6 is a diagrammatic representation of operations 600 performed by the pose generation system 230 according to some examples. Operations 600 are performed after the feature extractor 550 has been trained. That is, rather than inputting the same image as the input image and the target image, a first image depicting a person in a given pose is received as the input image. A target image depicting the same person or a different person in another pose is received. The trained feature extractor 550 processes the input image to obtain soft feature pooling. The target image is processed to generate target soft intrinsic distances. The target soft intrinsic distances are concatenated with the unpooled features of the input image to provide the encoded target geometry and input appearance. This information is processed by the decoder (e.g., Adain network 570) to generate a decoded output image.

[0140] That is, the encoder uses two signals to drive the output. Figure 6 As shown, the input image is processed to encode its appearance. For pooling, the source image is used, and for unpooling, the target image is used. For example, pose generation system 230 obtains a weighted average of features in a given region. Then, pose generation system 230 takes the average vector and places the features of that region on the target image. In other words, the features of the input image are smeared according to the features of the target image.

[0141] In some cases, the feature extractor 550 is further trained based on paired images depicting people in different poses. For example, a first training image depicting a person in a first pose is received as an input image. A second training image depicting a person in a second pose is received as a target image. The first training image is processed by the feature extractor 550 to generate a soft feature pool, which is then unpooled using features from the second training image. The unpooled features are concatenated with the target soft intrinsic distances and processed by the decoder to generate an output image. A loss is calculated between the generated output image and the second training image, and this loss is used to update the parameters of the feature extractor 550 and / or the Adaline network 570.

[0142] In some cases, not all parts of the target image are observed in the input image. For example, in one image, pose generation system 230 only observes the side of a person, while in another image, pose generation system 230 only observes the front of the person. In this case, the UV value of each pixel in the input image is mapped to distinguish those pixels that are observed from those that are not. Figure 7 The diffusion and attention layer 710 shown in is a trained network that can be included as part of the feature extractor 550 that diffuses visible information to unseen regions while using features from unobserved regions by the feature extractor 550 and / or the Adain network 570.

[0143] The goal of this task is to use separate appearance and pose donors and synthesize a person with the appearance of the appearance donor and the pose of the pose donor. The diffusion and attention layer or network 710 estimates the missing feature values ​​of the unobserved parts and then combines them with the observed feature values. In particular, the encoded features are updated by a residual branch modulated by an attention signal, which allows the pose generation system 230 to make stronger changes to the unobserved features and keep the fully observed features unchanged:

[0144] Diffusion: F D =WF O

[0145] Note: F′=AF D +(1-A)F O

[0146] A p =σ( <w p , C>)

[0147] Among them, W, w p is the parameter of end-to-end learning, F O Indicates the observed encoding, F Dis the result of diffusing information across control points, A is the attention signal indicating whether the encoding should be updated or retained, and C is the P-dimensional vector obtained from the normalized area across regions of the membership function,

[0148]

[0149] By w p , each region p can learn its own attention function, allowing, for example, a smaller head region to have a lower threshold to consider it occluded compared to a larger torso region. The diffusion and attention layer or network 710 is also trained based on paired training images to learn and update the diffusion and attention parameters (e.g., W, w p ). The attention signal A measures whether a pixel is observed. The attention signal controls whether diffuse features or observed features should be used. When a feature is visible, the feature or expected feature is used in the feature extractor 550, and when the feature is not visible, the diffuse feature is used instead. In some cases, soft feature pooling marks a feature as a diffuse feature or a visible feature, which is then used in the soft feature unpooling operation.

[0150] Figure 8 is a graphical representation 800 of position and feature statistical modeling according to some examples. In some cases, the feature extractor 550 also implements position and feature statistical modeling.

[0151] Position and feature statistics modeling collects the first and second order moments of the filter bank for a reference dataset (or a single image), and then synthesizes a new signal that reproduces the same statistics. Depending on whether the new signal is randomly initialized or constrained to remain close to some content signal, texture synthesis or style transfer is provided respectively. Instead of influencing the decoder network based on the style signal, the statistics of the network features are adjusted so that their Gram matrices match. This operation is supplemented by a decoder that learns a mapping from arbitrary network features to the input image. Therefore, any feature-level operation has a direct image counterpart that can be computed efficiently.

[0152] The disclosed system incorporates cartographic information into an efficient feed-forward system for style transfer that relies on Gaussian density modeling and subsequent matching ("coloring") of feature statistics. This allows the feature extractor 550 to condition the synthesis on continuous coordinates and facilitates texture modeling and inpainting that takes surface locations into account.

[0153] Whitening and color transformation can be implemented by the feature extractor 550 to match the statistics of the VGG network features of the content image with the style image. In particular, Represents the empirical feature mean and covariance of the content signal c and represents the empirical feature mean and covariance of the style signal, with the goal of matching the content feature statistics with those of the target signal. To this end, the feature extractor 550 first centers and rotates the content features using a whitening transformation:

[0154]

[0155] Among them, E c 、D c Depend on In the second step, the feature extractor 550 transforms the features again using the shading transformation:

[0156]

[0157] Among them, now, It can be easily verified has zero mean and unit covariance, and Mean with style signal and covariance This process is efficient because it only requires two matrix diagonalizations instead of gradient descent.By considering the dependence of texture statistics on position, the feature extractor 550 can handle the non-stationarity of stylization.

[0158] A multivariate Gaussian model is used to capture the dependence of the feature activation vector f on continuous position coordinates x. These can be pixel locations, or in this example, chart coordinates associated with a given observation. Specifically, the feature extractor 550 considers:

[0159]

[0160] The Gaussian mean and covariance matrix of the source and style images are estimated by accumulating local second-order statistics in each region. Using this factorization, spatially correlated feature activations can be obtained through the conditional distribution Modeling, where

[0161]

[0162] The position-corrected counterparts of the whitening and shading transforms can be obtained as follows:

[0163]

[0164] in, and The conditional feature covariance matrix is ​​used, and the centering transformation depends on the specific feature location x.

[0165] In some cases, a 2D chart coordinate x is mapped by a learnable embedding x'=φ(x) implemented via a two-layer MLP with 256 hidden units and 16 output vectors. The feature extractor 550 uses a joint feature texture model when modeling the appearance in the intrinsic UV coordinates. There is mirror symmetry around the central vertical axis (capturing the left / right symmetry of the appearance), and features should be more correlated in the horizontal direction rather than the vertical direction (because the texture changes more dramatically from pants / skirt to top / shirt). Based on these observations, the embedding network minimizes the following objectives:

[0166]

[0167] in, The penalty distance on the (vertical) u-axis is larger than the penalty distance on the (horizontal) v-axis and captures symmetry around the center.

[0168] As an example, Figure 8 As shown, an input image 810 is received. Along a horizontal axis, as shown by 820, a first portion of the input image (e.g., the portion where the torso appears) is copied across the image domain. That is, the pixels of the first portion are mirrored across the entire image space of that horizontal axis portion. Along a horizontal axis, as shown by 822, a second portion of the input image (e.g., the portion where the legs appear) is copied across the image domain. That is, the pixels of the second portion are mirrored across the entire image space of that horizontal axis portion. In this way, if a given texture appears in the target image, that texture will be statistically modeled and appear in the same portion of the output image when applied to the input image. In this way, for example, a shirt pattern will be visible across different poses of the target and input images.

[0169] Figure 9 is a diagrammatic representation of a graphical user interface 900 according to some examples. Figure 9 As shown, an input image 910 is received and a target image 920 is obtained. The pose generation system 230 processes the input image 910 to generate pooled features 552. The pose generation system 230 obtains the target soft intrinsic distance from the target image 920 and uses the target soft intrinsic distance from the target image 920 to unpool or unpool the soft features of the pooled features 552. This generates an encoded target geometry and input appearance, which is provided to the CHAIN ​​decoder to generate an output image 930. As an example, a person in a first pose facing away from the camera is received as an input image, and the target image may depict the same person in a second pose (e.g., a side view) with different clothing (e.g., a shirt). The pose generation system 230 outputs an image 930 in which the person in the first pose depicted in the input image 910 is depicted in the pose of the target image 920 and with the clothing of the person in the target image 920.

[0170] Figure 10A 1000 is a flowchart illustrating example operations of the messaging client 104 in performing process 1000, according to an example. Process 1000 can be implemented in computer-readable instructions for execution by one or more processors, such that the operations of process 1000 can be performed in part or in whole by functional components of the messaging server system 108; therefore, process 1000 is described below by way of example with reference thereto. However, in other examples, at least some of the operations of process 1000 can be deployed on various other hardware configurations. The operations in process 1000 can be performed in parallel in any order, or can be skipped or omitted entirely.

[0171] At operation 1001, the image processing server 122 receives a two-dimensional continuous surface representation of a three-dimensional object, the continuous surface including a plurality of landmark locations. For example, the gesture generation system 230 receives the surface coordinate representation 530 ( Figure 5 ).

[0172] At operation 1002 , the image processing server 122 determines a first set of soft membership functions based on landmark locations and relative locations of points in a two-dimensional continuous surface representation. For example, the pose generation system 230 calculates an estimated soft intrinsic distance 540 based on an input dense pose or target image.

[0173] At operation 1003 , the image processing server 122 receives a two-dimensional input image, the input image including an image of an object. For example, the gesture generation system 230 receives the input image 520 .

[0174] At operation 1004 , the image processing server 122 extracts a plurality of features from the input image using a feature recognition model. For example, the pose generation system 230 applies the feature extractor 550 to generate the soft feature pooling 552 .

[0175] At operation 1005 , the image processing server 122 generates encoded feature representations of the extracted features using the first set of soft membership functions. For example, the pose generation system 230 generates the soft feature pooling 552 using the surface coordinate representation 530 .

[0176] At operation 1006, the image processing server 122 generates a dense feature representation of the extracted features based on the encoded representation using the second set of soft membership functions. For example, the pose generation system 230 performs soft feature unpooling 554 using the input dense pose or features of the target image (e.g., the estimated soft intrinsic distance 540).

[0177] At operation 1007, the image processing server 122 uses the neural image decoder model to process the second set of soft membership functions and the dense feature representation to generate an output image. For example, the pose generation system 230 uses the Adain decoder 570 to process the encoded geometry and appearance 560 information to generate an output image.

[0178] Figure 10B 1000 is a flowchart illustrating example operations of the messaging client 104 in performing process 1000, according to an example. Process 1000 can be implemented in computer-readable instructions for execution by one or more processors, such that the operations of process 1000 can be performed in part or in whole by functional components of the messaging server system 108; therefore, process 1000 is described below by way of example with reference thereto. However, in other examples, at least some of the operations of process 1000 can be deployed on various other hardware configurations. The operations in process 1000 can be performed in parallel in any order, or can be skipped or omitted entirely.

[0179] At operation 1011 , the image processing server 122 uses an encoder neural network to determine a set of content features from a source image. For example, the pose generation system 230 processes the input image to generate a set of features, such as soft feature pooling 552 .

[0180] At operation 1012, the image processing server 122 determines a set of style features from the style image using an encoder neural network. For example, the pose generation system 230 processes the target image to identify style features (e.g., patterns on a shirt or pants in the image).

[0181] At operation 1013, the image processing server 122 uses the combined statistics of position and content features in the region of the source image to determine a position-dependent content feature. For example, the pose generation system 230 identifies regions in the source image where stylistic features should be replicated throughout, such as a torso region that should have features replicated horizontally but not vertically. As another example, a head region is a position-dependent content feature that should not have a style applied to it.

[0182] At operation 1014, the image processing server 122 uses the joint statistics of the position and style features in the region of the style image to determine the position-dependent style features. For example, the pose generation system 230 identifies the region in the source image where the style features should be replicated, e.g., the torso region should have features replicated horizontally but not vertically. Specifically, each region 820 and 822 replicates the features of the style image 810 horizontally across its region.

[0183] At operation 1015 , the image processing server 122 generates a set of transformed content features from the set of location-related content features based on joint statistics of location and content features.

[0184] At operation 1016, the image processing server 122 generates a set of transformed style features from the set of transformed content features based on the joint statistics of the position and style features. For example, the pose generation system 230 applies the style of the style image in one region (e.g., the shirt region) but not in another region (e.g., the head region) to the source image.

[0185] At operation 1017 , the image processing server 122 generates another output image based on the transformed set of style features and the transformed set of content features using the decoder neural network.

[0186] Machine Architecture

[0187] Figure 11The illustrative embodiment of the present invention is a diagrammatic representation of a machine 1100 in which instructions 1108 (e.g., software, programs, applications, applet, apps, or other executable code) may be executed for causing the machine 1100 to perform any one or more of the methodologies discussed herein. For example, the instructions 1108 may cause the machine 1100 to perform any one or more of the methodologies described herein. The instructions 1108 transform a general-purpose, unprogrammed machine 1100 into a specialized machine 1100 that is programmed to perform the functions described and illustrated in the manner described. The machine 1100 may operate as a standalone device or may be coupled (e.g., using a network) to other machines. In a networked deployment, the machine 1100 may operate in the capacity of a server or a client machine in a server-client network environment, or as a peer machine in a peer-to-peer (or distributed) network environment. The machine 1100 may include, but is not limited to, a server computer, a client computer, a personal computer (PC), a tablet computer, a laptop computer, a netbook, a set-top box (STB), a personal digital assistant (PDA), an entertainment media system, a cellular phone, a smartphone, a mobile device, a wearable device (e.g., a smart watch), a smart home device (e.g., a smart appliance), other smart devices, a network device, a network router, a network switch, a network bridge, or any machine capable of executing instructions 1108 that specify actions to be taken by the machine 1100, either sequentially or otherwise. Furthermore, while only a single machine 1100 is shown, the term "machine" should also be construed to include a collection of machines that individually or jointly execute instructions 1108 to perform any one or more of the methods discussed herein. For example, the machine 1100 may include the client device 102 or any of several server devices that form part of the messaging server system 108. In some examples, the machine 1100 may also include both a client system and a server system, wherein certain operations of a particular method or algorithm are performed on the server side, and wherein certain operations of a particular method or algorithm are performed on the client side.

[0188] The machine 1100 may include a processor 1102, a memory 1104, and input / output (I / O) components 1138 that may be configured to communicate with each other via a bus 1140. In an example, the processor 1102 (e.g., a central processing unit (CPU), a reduced instruction set computing (RISC) processor, a complex instruction set computing (CISC) processor, a graphics processing unit (GPU), a digital signal processor (DSP), an application specific integrated circuit (ASIC), a radio frequency integrated circuit (RFIC), another processor, or any suitable combination thereof) may include, for example, a processor 1106 that executes instructions 1108 and a processor 1110. The term "processor" is intended to include a multi-core processor that may include two or more independent processors (sometimes referred to as "cores") that may execute instructions concurrently. Although Figure 11 Multiple processors 1102 are shown, but the machine 1100 may include a single processor with a single core, a single processor with multiple cores (e.g., a multi-core processor), multiple processors with a single core, multiple processors with multiple cores, or any combination thereof.

[0189] The memory 1104 includes a main memory 1112, a static memory 1114, and a storage unit 1116, all of which are accessible by the processor 1102 via the bus 1140. The main memory 1104, the static memory 1114, and the storage unit 1116 store instructions 1108 that implement any one or more of the methods or functions described herein. The instructions 1108 may also reside, completely or partially, within the main memory 1112, within the static memory 1114, within the machine-readable medium 1118 within the storage unit 1116, within at least one of the processors 1102 (e.g., within a cache memory of the processor), or within any suitable combination thereof during execution thereof by the machine 1100.

[0190] The I / O components 1138 may include various components for receiving input, providing output, generating output, transmitting information, exchanging information, capturing measurements, etc. The specific I / O components 1138 included in a particular machine will depend on the type of machine. For example, a portable machine such as a mobile phone may include a touch input device or other such input mechanism, while a headless server machine may not include such a touch input device. It should be appreciated that the I / O components 1138 may include Figure 11Many other components are not shown in the drawings. In various examples, the I / O components 1138 may include user output components 1124 and user input components 1126. The user output components 1124 may include visual components (e.g., displays such as plasma display panels (PDPs), light emitting diode (LED) displays, liquid crystal displays (LCDs), projectors, or cathode ray tubes (CRTs)), acoustic components (e.g., speakers), tactile components (e.g., vibration motors, resistance mechanisms), other signal generators, etc. The user input components 1126 may include alphanumeric input components (e.g., keyboards, touch screens configured to receive alphanumeric input, optical keyboards, or other alphanumeric input components), point-based input components (e.g., mice, touch pads, trackballs, joysticks, motion sensors, or other pointing instruments), tactile input components (e.g., physical buttons, touch screens or other tactile input components that provide location and / or force of touch or touch gestures), audio input components (e.g., microphones), etc.

[0191] In another example, the I / O component 1138 may include a biometric component 1128, a motion component 1130, an environmental component 1132, or a position component 1134, as well as various other components. For example, the biometric component 1128 includes components for detecting expressions (e.g., hand expressions, facial expressions, voice expressions, body gestures, or eye tracking), measuring biosignals (e.g., blood pressure, heart rate, body temperature, sweat, or brain waves), identifying people (e.g., voice recognition, retinal recognition, facial recognition, fingerprint recognition, or electroencephalogram-based recognition), etc. The motion component 1130 includes an acceleration sensor component (e.g., an accelerometer), a gravity sensor component, and a rotation sensor component (e.g., a gyroscope).

[0192] Environmental components 1132 include, for example, one or more cameras (with still image / photo and video capabilities), an illumination sensor component (e.g., a photometer), a temperature sensor component (e.g., one or more thermometers that detect ambient temperature), a humidity sensor component, a pressure sensor component (e.g., a barometer), an acoustic sensor component (e.g., one or more microphones that detect background noise), a proximity sensor component (e.g., an infrared sensor that detects nearby objects), a gas sensor (e.g., a gas detection sensor that detects concentrations of hazardous gases for safety or measures pollutants in the atmosphere), or other components that can provide indications, measurements, or signals corresponding to the surrounding physical environment.

[0193] With respect to cameras, client device 102 can have a camera system that includes, for example, a front-facing camera on the front surface of client device 102 and a rear-facing camera on the rear surface of client device 102. The front-facing camera can, for example, be used to capture still images and videos of a user of client device 102 (e.g., a "selfie"), which can then be enhanced with the enhancement data (e.g., filters) described above. For example, the rear-facing camera can be used to capture still images and videos in a more conventional camera mode, which are similarly enhanced with the enhancement data. In addition to the front-facing camera and the rear-facing camera, client device 102 can also include a 360° camera for capturing 360° photos and videos.

[0194] Additionally, the camera system of the client device 102 may include dual rear cameras (e.g., a main camera and a depth sensing camera), or even triple, quad, or quintuple rear camera configurations on the front and back sides of the client device 102. For example, these multi-camera systems may include a wide-angle camera, an ultra-wide-angle camera, a telephoto camera, a macro camera, and a depth sensor.

[0195] The location component 1134 includes a positioning sensor component (e.g., a GPS receiver component), an altitude sensor component (e.g., an altimeter or barometer that detects air pressure, from which altitude can be obtained), an orientation sensor component (e.g., a magnetometer), and the like.

[0196] Various technologies can be used to implement communications. The I / O components 1138 also include a communications component 1136 that is operable to couple the machine 1100 to the network 1120 or device 1122 via corresponding couplings or connections. For example, the communications component 1136 may include a network interface component or other suitable device that interfaces with the network 1120. In other examples, the communications component 1136 may include a wired communications component, a wireless communications component, a cellular communications component, a near field communications (NFC) component, a Components (e.g. Low power consumption), Device 1122 may be another machine or any of a variety of peripheral devices (eg, a peripheral device coupled via USB).

[0197] In addition, the communication component 1136 can detect the identifier or include a component that is operable to detect the identifier. For example, the communication component 1136 can include a radio frequency identification (RFID) tag reader component, an NFC smart tag detection component, an optical reader component (e.g., an optical sensor for detecting one-dimensional bar codes such as Universal Product Code (UPC) bar codes, multi-dimensional bar codes such as Quick Response (QR) codes, Aztec codes, Data Matrix, Dataglyph, MaxiCode, PDF417, Ultra Code, UCC RSS-2D bar codes, and other optical codes) or an acoustic detection component (e.g., a microphone for identifying tagged audio signals). In addition, various information can be obtained via the communication component 1136, such as location positioning via Internet Protocol (IP), location information via Internet Protocol (IP) geolocation, ... Positioning by signal triangulation, positioning via detection of NFC beacon signals, etc., which can indicate a specific position.

[0198] Various memories (e.g., main memory 1112, static memory 1114, and memory of processor 1102) and storage unit 1116 may store one or more sets of instructions and data structures (e.g., software) implemented or used by any one or more of the methods or functions described herein. When executed by processor 1102, these instructions (e.g., instructions 1108) cause various operations to implement the disclosed examples.

[0199] Instructions 1108 may be sent or received over network 1120 via a network interface device (e.g., a network interface component included in communications component 1136) using a transmission medium and using any of a number of well-known transmission protocols (e.g., Hypertext Transfer Protocol (HTTP)). Similarly, instructions 1108 may be sent or received via a coupling (e.g., a peer-to-peer coupling) with device 1122 using a transmission medium.

[0200] Software Architecture

[0201] Figure 1212 is a block diagram 1200 illustrating a software architecture 1204 that can be installed on any one or more of the devices described herein. The software architecture 1204 is supported by hardware such as a machine 1202 including a processor 1220, memory 1226, and I / O components 1238. In this example, the software architecture 1204 can be conceptualized as a stack of layers, where each layer provides specific functionality. The software architecture 1204 includes layers such as an operating system 1212, libraries 1210, frameworks 1208, and applications 1206. In operation, the applications 1206 invoke API calls 1250 through the software stack and receive messages 1252 in response to the API calls 1250.

[0202] The operating system 1212 manages hardware resources and provides common services. The operating system 1212 includes, for example, a kernel 1214, services 1216, and drivers 1222. The kernel 1214 acts as an abstraction layer between the hardware layer and other software layers. For example, the kernel 1214 provides memory management, processor management (e.g., scheduling), component management, networking and security settings, and other functions. Services 1216 can provide other common services to other software layers. Drivers 1222 are responsible for controlling or interfacing with the underlying hardware. For example, drivers 1222 may include display drivers, camera drivers, or Low-power drivers, Flash drivers, serial communication drivers (e.g., USB drivers), drivers, audio drivers, power management drivers, etc.

[0203] The libraries 1210 provide a common low-level infrastructure used by the applications 1206. The libraries 1210 may include system libraries 1218 (e.g., C standard libraries) that provide functions such as memory allocation functions, string manipulation functions, mathematical functions, etc. In addition, the libraries 1210 may include API libraries 1224, such as media libraries (e.g., libraries for supporting the presentation and manipulation of various media formats, such as Moving Picture Experts Group 4 (MPEG4), Advanced Video Coding (H.264 or AVC), Moving Picture Experts Group Layer 3 (MP3), Advanced Audio Coding (AAC), Adaptive Multi-Rate (AMR) audio codec, Joint Photographic Experts Group (JPEG or JPG), or Portable Network Graphics (PNG)), graphics libraries (e.g., an OpenGL framework for rendering graphical content on a display in two dimensions (2D) and three dimensions (3D), database libraries (e.g., SQLite providing various relational database functions), web libraries (e.g., WebKit providing web browsing functions), etc. The library 1210 may also include various other libraries 1228 to provide many other APIs to the application 1206 .

[0204] The framework 1208 provides a common high-level infrastructure used by the applications 1206. For example, the framework 1208 provides various graphical user interface (GUI) functions, advanced resource management, and advanced positioning services. The framework 1208 can provide a wide range of other APIs that can be used by the applications 1206, some of which may be specific to a particular operating system or platform.

[0205] In an example, applications 1206 may include a home application 1236, a contacts application 1230, a browser application 1232, a book reader application 1234, a location application 1242, a media application 1244, a messaging application 1246, a game application 1248, and a variety of other applications such as external applications 1240. Applications 1206 are programs that perform functions defined in the program. Various programming languages ​​may be used to create one or more of the applications 1206 structured in various ways, such as an object-oriented programming language (e.g., Objective-C, Java, or C++) or a procedural programming language (e.g., C or assembly language). In a specific example, external applications 1240 (e.g., those written by an entity other than the vendor of a particular platform using ANDROID) may be used to create a program that is not a part of the platform. TM or IOS TM Software Development Kit (SDK) can be used to develop applications on platforms such as IOS TM ANDROID TM 、 Mobile software running on the mobile operating system of the phone or another mobile operating system. In this example, the external application 1240 can activate the API call 1250 provided by the operating system 1212 to facilitate the functions described herein.

[0206] Glossary

[0207] "Carrier signal" refers to any intangible medium that can store, encode, or carry instructions for execution by a machine and includes digital or analog communication signals or other intangible media to facilitate communication of such instructions. Instructions may be sent or received over a network using a transmission medium via a network interface device.

[0208] "Client Device" refers to any machine that interfaces with a communications network to obtain resources from one or more server systems or other client devices. A client device may be, but is not limited to, a mobile phone, desktop computer, laptop computer, portable digital assistant (PDA), smartphone, tablet computer, ultrabook, netbook, laptop computer, multiprocessor system, microprocessor-based or programmable consumer electronics, game console, set-top box, or any other communications device that a user may use to access a network.

[0209] "Communications network" means one or more parts of a network, which may be an ad hoc network, an intranet, an extranet, a virtual private network (VPN), a local area network (LAN), a wireless LAN (WLAN), a wide area network (WAN), a wireless WAN (WWAN), a metropolitan area network (MAN), the Internet, a part of the Internet, a part of the Public Switched Telephone Network (PSTN), a Plain Old Telephone Service (POTS) network, a cellular telephone network, a wireless network, The coupling may be a network, another type of network, or a combination of two or more such networks. For example, the network or a portion of the network may include a wireless network or a cellular network, and the coupling may be a code division multiple access (CDMA) connection, a global system for mobile communications (GSM) connection, or other type of cellular or wireless coupling. In this example, the coupling may implement any of various types of data transmission technologies, such as single carrier radio transmission technology (1xRTT), evolution data optimized (EVDO) technology, general packet radio service (GPRS) technology, enhanced data rates for GSM evolution (EDGE) technology, the third generation partnership project (3GPP) including 3G, fourth generation wireless (4G) networks, universal mobile telecommunications system (UMTS), high speed packet access (HSPA), world wide interoperability for microwave access (WiMAX), long term evolution (LTE) standards, other data transmission technologies defined by various standards setting organizations, other long distance protocols, or other data transmission technologies.

[0210] "Component" means a device, physical entity, or logic having boundaries defined by function or subroutine calls, branch points, APIs, or other techniques provided for partitioning or modularizing specific processing or control functions. A component can be combined with other components via its interfaces to execute a machine process. A component can be a packaged functional hardware unit designed for use with other components and can be part of a program that generally performs a specific one of the related functions.

[0211] Components may constitute software components (e.g., code implemented on a machine-readable medium) or hardware components. A "hardware component" is a tangible unit that is capable of performing certain operations and may be configured or arranged in some physical manner. In various examples, one or more computer systems (e.g., stand-alone computer systems, client computer systems, or server computer systems) or one or more hardware components of a computer system (e.g., a processor or group of processors) may be configured by software (e.g., an application or application portion) to be hardware components that operate to perform certain operations as described herein.

[0212] Hardware components can also be implemented mechanically, electronically, or in any suitable combination thereof. For example, a hardware component can include a dedicated circuit or logic that is permanently configured to perform certain operations. A hardware component can be a dedicated processor, such as a field programmable gate array (FPGA) or an application specific integrated circuit (ASIC). A hardware component can also include a programmable logic or circuit that is temporarily configured to perform certain operations by software. For example, a hardware component can include software executed by a general-purpose processor or other programmable processor. Once configured by such software, the hardware component becomes a specific machine (or a specific component of a machine) that is uniquely customized to perform the configured function and is no longer a general-purpose processor. It will be understood that the decision to implement a hardware component mechanically, in a dedicated and permanently configured circuit, or in a temporarily configured circuit (e.g., configured by software) can be driven by cost and time considerations. Accordingly, the phrase "hardware component" (or "hardware-implemented component") should be understood to include a tangible entity, i.e., an entity that is physically constructed, permanently configured (e.g., hardwired) or temporarily configured (e.g., programmed) to operate in some way or perform certain operations described herein.

[0213] Considering an example where hardware components are temporarily configured (e.g., programmed), each of the hardware components need not be configured or instantiated at any one time. For example, where the hardware components include a general-purpose processor that is configured by software to become a special-purpose processor, the general-purpose processor can be configured to become a different special-purpose processor (e.g., including different hardware components) at different times. Thus, the software configures one or more specific processors to, for example, constitute a specific hardware component at one time and to constitute a different hardware component at a different time.

[0214] Hardware components can provide information to other hardware components and receive information from other hardware components. Therefore, described hardware components can be considered to be coupled in communication. In the case of multiple hardware components simultaneously, communication can be realized by (for example, by suitable circuit and bus) between two or more hardware components or in the middle of carrying out signal transmission. In the example that multiple hardware components are configured or instantiated at different times, communication between such hardware components can be realized, for example, by storing information in a memory structure that multiple hardware components can access and retrieving information in the memory structure. For example, a hardware component can perform an operation and the output of the operation is stored in a memory device coupled in communication with it. Then, other hardware components can access the memory device at a subsequent time to retrieve the stored output and process it. Hardware components can also initiate communication with input devices or output devices, and can operate on resources (for example, the collection of information).

[0215] The various operations of the example methods described herein may be performed at least in part by one or more processors, one or more of which are temporarily configured (e.g., by software) or permanently configured to perform the relevant operations. Whether temporarily or permanently configured, such processors may constitute processor-implemented components that operate to perform one or more operations or functions described herein. As used herein, "processor-implemented components" refer to hardware components implemented using one or more processors. Similarly, the methods described herein may be at least partially implemented by processors, wherein the specific one or more processors are examples of hardware. For example, at least some of the operations of the methods may be performed by one or more processors 1102 or processor-implemented components. In addition, one or more processors may also operate to support the execution of relevant operations in a "cloud computing" environment or as a "software as a service" (SaaS) operation. For example, at least some of the operations may be performed by a group of computers (as an example of a machine including a processor), wherein the operations may be accessed via a network (e.g., the Internet) and via one or more appropriate interfaces (e.g., APIs). The execution of certain operations may be distributed among the processors, not only residing within a single machine, but also deployed on several machines. In some examples, the processor or processor-implemented components may be located in a single geographic location (e.g., in a home environment, an office environment, or a server farm). In other examples, the processor or processor-implemented components may be distributed across several geographic locations.

[0216] "Computer-readable storage media" refers to both machine storage media and transmission media. Thus, the term encompasses both storage devices / medium and carrier / modulated data signals. The terms "machine-readable medium," "computer-readable medium," and "device-readable medium" mean the same thing and may be used interchangeably in this disclosure.

[0217] An "ephemeral message" is a message that is accessible for a limited duration. An ephemeral message can be text, an image, a video, or the like. The access time for an ephemeral message can be set by the sender. Alternatively, the access time can be a default setting or a setting specified by the recipient. Regardless of the setting technique, the message is ephemeral.

[0218] “Machine storage media” refers to a single or multiple storage devices and media (e.g., centralized or distributed databases, and associated caches and servers) that store executable instructions, routines, and data. Thus, the term should be taken to include, but is not limited to, solid-state memory and optical and magnetic media, including memory internal or external to the processor. Specific examples of machine storage media, computer storage media, and device storage media include: non-volatile memory, including, for example, semiconductor memory devices such as erasable programmable read-only memory (EPROM), electrically erasable programmable read-only memory (EEPROM), FPGAs, and flash memory devices; magnetic disks, such as internal hard disks and removable disks; magneto-optical disks; and CD-ROM and DVD-ROM disks. The terms “machine storage media,” “device storage media,” and “computer storage media” mean the same thing and are used interchangeably in this disclosure. The terms “machine storage media,” “computer storage media,” and “device storage media” expressly exclude carrier waves, modulated data signals, and other such media, at least some of which are encompassed by the term “signal media.”

[0219] “Non-transitory computer-readable storage medium” refers to a tangible medium capable of storing, encoding, or carrying instructions for execution by a machine.

[0220] "Signal medium" refers to any intangible medium that can store, encode, or carry instructions for execution by a machine, and "signal medium" includes digital or analog communication signals or other intangible media to facilitate the transmission of software or data. The term "signal medium" should be deemed to include any form of modulated data signal, carrier wave, etc. The term "modulated data signal" refers to a signal whose one or more characteristics are set or changed in such a manner as to encode information in the signal. The terms "transmission medium" and "signal medium" refer to the same thing and are used interchangeably in this disclosure.

[0221] In particular, the following scheme is proposed in this paper.

[0222] 1. A computer-implemented neural image synthesis method, the method comprising:

[0223] receiving a two-dimensional continuous surface representation of a three-dimensional object, the continuous surface comprising a plurality of landmark locations;

[0224] determining a first set of soft membership functions based on the landmark locations and the relative locations of points in the two-dimensional continuous surface representation;

[0225] receiving a two-dimensional input image, the input image comprising an image of the object;

[0226] extracting a plurality of features from the input image using a feature recognition model;

[0227] generating an encoded feature representation of the extracted features using the first set of soft membership functions;

[0228] generating a dense feature representation of the extracted features from the encoded representation using a second set of soft membership functions;

[0229] processing the second set of soft membership functions and the dense feature representation using a neural image decoder model to generate an output image; and

[0230] The output image is caused to be presented on a client device.

[0231] 2. The method of claim 1 , wherein determining the first set of soft membership functions comprises:

[0232] determining distances between the landmark locations and a plurality of points in the two-dimensional continuous surface representation; and

[0233] Each point in the plurality of points is assigned to a landmark based on the determined distance.

[0234] 3. The method according to solution 1 also includes: using a landmark recognition model to determine the landmark location.

[0235] 4. The method of claim 1 , wherein the neural image decoder model comprises a convolutional neural network conditioned on the two-dimensional continuous surface representation.

[0236] 5. The method of claim 1 , wherein generating an encoded feature representation of the extracted features using the first set of soft membership functions comprises performing a weighted membership estimation of the mean and variance of each channel of the extracted features.

[0237] 6. A method according to Option 5, wherein generating a dense feature representation of the extracted features based on the encoded representation using a second set of soft membership functions includes: applying a dual operation to the membership weighted estimates of the mean and variance of each channel of the extracted features.

[0238] 7. The method of claim 1 , wherein the object is a human body, and wherein the landmarks include joints of the human body.

[0239] 8. The method according to claim 1, wherein the first set of soft membership functions and the second set of soft membership functions are the same.

[0240] 9. The method according to claim 8, further comprising:

[0241] generating a three-dimensional model of the three-dimensional object based on the input image; and

[0242] The two-dimensional continuous surface representation is generated based on the three-dimensional model.

[0243] 10. The method according to claim 9 further includes: modifying the values ​​in the encoded representation before generating the dense feature representation.

[0244] 11. The method of claim 1 , wherein the two-dimensional continuous surface representation of the three-dimensional object is generated based on the input image, and wherein the method further comprises:

[0245] receiving a further two-dimensional input image, the further input image comprising a further object of the same type as the three-dimensional object;

[0246] generating a further two-dimensional continuous surface representation of the three-dimensional object from the further two-dimensional input image, the further continuous surface comprising the plurality of landmark locations; and

[0247] The second set of soft membership functions is determined based on the landmark locations and the relative locations of points in the further two-dimensional continuous surface representation.

[0248] 12. The method of claim 11, wherein the input image comprises an image of the object in a first pose, and the further input image comprises an image of the further object in a second pose, wherein the second image comprises a portion corresponding to an unseen portion of the first image; and

[0249] Generating the two-dimensional continuous surface representation according to the input image includes: using the learned attention mechanism to generate a portion of the two-dimensional continuous surface representation corresponding to the unseen portion of the first image according to the encoded representation.

[0250] 13. A method according to Option 12, wherein the learned attention mechanism is based on the first set of soft membership functions.

[0251] 14. The method according to claim 1, further comprising:

[0252] Determine a set of content features from the source image using an encoder neural network;

[0253] determining a set of style features from the style image using the encoder neural network;

[0254] determining a position-dependent content feature using joint statistics of position and content features in a region of the source image;

[0255] determining a position-dependent style feature using joint statistics of position and style features in a region of the style image;

[0256] generating a set of transformed content features from a set of position-dependent content features based on joint statistics of the position and content features;

[0257] generating a set of transformed style features from the set of transformed content features based on joint statistics of the position and style features; and

[0258] Another output image is generated based on the set of transformed style features and the set of transformed content features using a decoder neural network.

[0259] 15. The method according to claim 14 further comprises: combining the output image with the other output image to generate a combined image.

[0260] 16. A method according to Option 14, wherein the joint statistics of the location and content features include the content feature mean, the content location mean, and the covariance between the content feature and the content location, and wherein determining the location-related content features includes: determining a conditional model of the content features conditional on the location.

[0261] 17. The method according to claim 16, wherein the conditional model of the content feature includes a position-related content mean and a conditional content covariance.

[0262] 18. The method of claim 17, wherein generating the set of transformed content features based on the set of location-dependent content features comprises:

[0263] centering the location-related content features based on the location-related content means; and

[0264] A whitening transform is applied based on the conditional content covariance.

[0265] 19. A system for neural image analysis, comprising:

[0266] A processor configured to perform operations comprising:

[0267] receiving a two-dimensional continuous surface representation of a three-dimensional object, the continuous surface comprising a plurality of landmark locations;

[0268] determining a first set of soft membership functions based on the landmark locations and the relative locations of points in the two-dimensional continuous surface representation;

[0269] receiving a two-dimensional input image, the input image comprising an image of the object;

[0270] extracting a plurality of features from the input image using a feature recognition model;

[0271] generating an encoded feature representation of the extracted features using the first set of soft membership functions;

[0272] generating a dense feature representation of the extracted features from the encoded representation using a second set of soft membership functions;

[0273] processing the second set of soft membership functions and the dense feature representation using a neural image decoder model to generate an output image; and

[0274] The output image is caused to be presented on a client device.

[0275] 20. A non-transitory machine-readable storage medium comprising instructions that, when executed by one or more processors of a machine, cause the machine to perform operations for neural image analysis, the operations comprising:

[0276] receiving a two-dimensional continuous surface representation of a three-dimensional object, the continuous surface comprising a plurality of landmark locations;

[0277] determining a first set of soft membership functions based on the landmark locations and the relative locations of points in the two-dimensional continuous surface representation;

[0278] receiving a two-dimensional input image, the input image comprising an image of the object;

[0279] extracting a plurality of features from the input image using a feature recognition model;

[0280] generating an encoded feature representation of the extracted features using the first set of soft membership functions;

[0281] generating a dense feature representation of the extracted features from the encoded representation using a second set of soft membership functions;

[0282] processing the second set of soft membership functions and the dense feature representation using a neural image decoder model to generate an output image; and

[0283] The output image is caused to be presented on a client device.

[0284] Changes and modifications may be made to the disclosed examples without departing from the scope of the present disclosure. These and other changes or modifications are intended to be included within the scope of the present disclosure as expressed in the following claims.

Claims

1. A method comprising: receiving a two-dimensional (2D) input image, the two-dimensional (2D) input image depicting a person in a first pose; receiving a target image corresponding to a second posture; Processing the 2D input image to generate soft feature pooling features; processing the target image to generate a target soft intrinsic distance; as well as Features of the 2D input image are decoded based on a combination of the soft feature pooling features and the target soft intrinsic distance to generate a decoded image, the decoded image depicting the person in the second pose.

2. The method according to claim 1, further comprising: Determine the first set of soft membership functions by following these steps: determining distances between landmark locations of the 2D input image and a plurality of points in a two-dimensional continuous surface representation; as well as Each point in the plurality of points is assigned to a landmark based on the determined distance.

3. The method according to claim 1, further comprising: Landmark locations were determined using a landmark recognition model.

4. The method according to claim 1, further comprising: The features are decoded by a neural image decoder model comprising a convolutional neural network conditioned on a two-dimensional continuous surface representation of the 2D input image.

5. The method according to claim 1, further comprising: An encoded feature representation of the extracted features of the 2D input image is generated using a first set of soft membership functions by performing membership weighted estimation of the mean and variance of each channel of the extracted features of the 2D input image.

6. The method according to claim 5, further comprising: A dense feature representation of the extracted features is generated from the encoded feature representation using a second set of soft membership functions by applying a dual operation to the membership weighted estimates of the mean and variance of each channel of the extracted features.

7. The method according to claim 1, further comprising: Landmarks are detected, the landmarks including joints of the person.

8. The method according to claim 1, wherein The first set of soft membership functions is the same as the second set of soft membership functions used to decode the features.

9. The method according to claim 8, further comprising: generating a three-dimensional 3D model of the person based on the 2D input image; as well as A 2D continuous surface representation is generated from the 3D model.

10. The method according to claim 9, further comprising: Prior to generating a dense feature representation of the 2D input image, values ​​in the encoded representation of the 2D input image are modified.

11. The method according to claim 1 , further comprising: generating a 2D continuous surface representation of a three-dimensional 3D object from the 2D input image, receiving a further 2D input image, the further 2D input image comprising a further object; generating a further 2D continuous surface representation of the 3D object from the further 2D input image, the further 2D continuous surface comprising a plurality of landmark locations; and A second set of soft membership functions is determined based on the landmark locations and the relative locations of points in the further 2D continuous surface representation.

12. The method according to claim 11, wherein The decoded image includes a portion corresponding to an unseen portion of the 2D input image, and the method further comprises: A portion of a two-dimensional continuous surface representation corresponding to the unseen portion is generated from the encoded representation using the learned attention mechanism.

13. The method according to claim 12, wherein: The learned attention mechanism is based on a first set of soft membership functions.

14. The method according to claim 1, further comprising: Determine a set of content features from the source image using an encoder neural network; determining a set of style features from the style image using the encoder neural network; determining a position-dependent content feature using joint statistics of position and content features in a region of the source image; determining a position-dependent style feature using joint statistics of position and style features in a region of the style image; generating a set of transformed content features from a set of position-dependent content features based on joint statistics of the position and content features; generating a set of transformed style features from the set of transformed content features based on joint statistics of the position and style features; as well as An output image is generated based on the set of transformed style features and the set of transformed content features using a decoder neural network.

15. The method according to claim 14, further comprising: The output image is combined with another output image to generate a combined image.

16. The method according to claim 14, wherein The joint statistics of the position and content features include a content feature mean, a content position mean, and a covariance between the content feature and the content position, and wherein determining the position-related content feature includes determining a conditional model of the content feature conditioned on position.

17. The method according to claim 16, wherein The conditional model of the content features includes a position-dependent content mean and a conditional content covariance.

18. The method according to claim 17, wherein Generating the set of transformed content features based on the set of position-dependent content features comprises: centering the location-related content features based on the location-related content means; and A whitening transform is applied based on the conditional content covariance.

19. A system for neural image analysis, comprising: A processor configured to perform operations comprising: receiving a two-dimensional (2D) input image, the two-dimensional (2D) input image depicting a person in a first pose; receiving a target image corresponding to a second posture; Processing the 2D input image to generate soft feature pooling features; processing the target image to generate a target soft intrinsic distance; and Features of the 2D input image are decoded based on a combination of the soft feature pooling features and the target soft intrinsic distance to generate a decoded image, the decoded image depicting the person in the second pose.

20. A non-transitory machine-readable storage medium comprising instructions that, when executed by one or more processors of a machine, cause the machine to perform operations for neural image analysis, the operations comprising: receiving a two-dimensional (2D) input image, the two-dimensional (2D) input image depicting a person in a first pose; receiving a target image corresponding to a second posture; Processing the 2D input image to generate soft feature pooling features; processing the target image to generate a target soft intrinsic distance; as well as Features of the 2D input image are decoded based on a combination of the soft feature pooling features and the target soft intrinsic distance to generate a decoded image, the decoded image depicting the person in the second pose.

Images