JP7784082B2 - 3D human model output device, program, and 3D human model output method - Google Patents

Description

This disclosure relates to a three-dimensional human model generation device, and more specifically to a device that enables computer graphics designers to generate desired three-dimensional human models with minimal effort.

In the technical field of computer graphics, there is an ever-increasing demand for the use of a large number of polygons to display people in high detail in three dimensions. However, displaying people in high detail in three dimensions requires a great deal of effort. In order to meet this growing demand and make it possible to provide more high-quality three-dimensional human models, there is a need to reduce the effort required to generate each three-dimensional human model.

One technology that can reduce the amount of work required is one that can automatically generate data related to the color and texture of the surface of three-dimensional model data (also called "texture") by performing a coloring process based on a trained neural network model that has been trained in advance using example training data (see Patent Document 1). According to the technology described in Patent Document 1, texture is automatically generated using a trained neural network model, which can reduce the amount of work required to set the color and texture to be applied to the surface of a three-dimensional model.

Japanese Patent Application Laid-Open No. 2020-013390

When generating 3D character models, it is important that the designer can easily generate a 3D character model that meets the client's needs, matching the attributes and shapes desired by the client. However, the technology in Patent Document 1 only automatically generates data related to the color and texture of the 3D model surface, and there is still room for improvement in terms of automating the process of meeting the client's needs.

More specifically, in the neural network machine learning using training data described in Patent Document 1, the designer prepares model training data. A neural network that generates a three-dimensional human model equipped with the attributes desired by the client of the designer requires training data corresponding to each of the client's desired attributes. This requires a great deal of effort from the designer to prepare a large amount of training data. In order for designers to generate the desired three-dimensional human model with minimal effort, it is important to reduce the effort required for neural network training.

This disclosure has been made in light of these circumstances, and its purpose is to provide a 3D character model generation device that enables computer graphics designers to generate the 3D character models desired by clients with as little effort as possible.

As a result of extensive research into solving the above problems, the inventors of the present application discovered that the above objectives could be achieved by performing unsupervised learning of neural networks, leading to the completion of this disclosure. Specifically, this disclosure provides the following:

The first disclosure relates to an evaluation learning unit that can use person model configuration information that can configure a three-dimensional human model, evaluation information for each of a plurality of types of three-dimensional human models configured using the person model configuration information, and attribute information for the plurality of types of three-dimensional human models to machine-learn evaluation criteria for the plurality of types of three-dimensional human models using a first neural network; a generation unit that can generate a plurality of types of generated three-dimensional human models using a second neural network and the attribute information; an evaluation unit that can evaluate the plurality of types of generated three-dimensional human models using the evaluation criteria and the attribute information; The present invention provides a 3D human model generation device comprising: an output unit capable of outputting a single type of output 3D human model evaluated by the evaluation unit, or multiple types of output 3D human models evaluated, from among multiple types of generated 3D human models; a receiving unit capable of receiving user evaluation information regarding the output 3D human model; an update unit capable of updating the human model configuration information using the output 3D human model and updating the evaluation information using the user evaluation information; and a machine learning unit capable of performing unsupervised machine learning for at least the second neural network.

According to the first disclosure, the first neural network can be made to learn by machine learning evaluation criteria related to the attributes desired by a user (a graphic designer who is the user of the 3D human model) in a 3D human model. The generation unit can generate multiple 3D human models to be generated that have the attributes desired by the user. The output unit can then select and output multiple types of output 3D human models from the multiple 3D human models generated by the generation unit that satisfy the evaluation criteria related to the attributes, i.e., that are recognized as desired 3D human models. Because a series of processes related to the output of the output 3D human models is performed automatically, the designer can output the 3D human model desired by the user (the client) with less effort.

In the evaluation learning unit, the character model configuration information, evaluation information, and attribute information serve as training data that serves as a model for machine learning. Because the update unit can update the character model configuration information and evaluation information, designers can prepare training data with little effort. In other words, designers can obtain evaluation criteria that better reflect the attributes that users (those who order designers) desire in 3D character models with little effort. This allows designers to output the 3D character models desired by users (those who order designers) with little effort.

According to the first disclosure, the generation unit can use the machine-learned second neural network to generate a more detailed 3D human model desired by the user (the person ordering the designer). Unsupervised machine learning can be performed without externally provided model training data, so a more detailed 3D human model desired by the user (the person ordering the designer) can be generated without increasing the designer's workload.

Therefore, according to the first disclosure, it is possible to provide a 3D human model generation device that enables a designer working in computer graphics to generate the 3D human model desired by the user (orderer) with as little effort as possible.

The second disclosure provides a 3D human model generation device in addition to the first disclosure, in which the machine learning unit is capable of performing the machine learning using the unsupervised learning for both the first neural network and the second neural network.

According to the second disclosure, because machine learning is performed through unsupervised learning, the evaluation learning unit can have the first neural network learn evaluation criteria based on updated character model configuration information and evaluation information, even without training data provided by the designer. Therefore, a three-dimensional character model desired by the user (client) can be output in a more vivid state without increasing the designer's effort.

The third disclosure provides a 3D human model generation device according to the first or second disclosure, wherein the unsupervised learning is learning using a generative adversarial network.

According to the third disclosure, machine learning can be performed using a generative adversarial network, with the second neural network serving as the generator of the generative adversarial network and the first neural network serving as the classifier of the generative adversarial network. In machine learning using a generative adversarial network, the generator generates data and the classifier classifies the generated data. The generator then performs machine learning with the goal of having the classifier classify the generated data as correct data.

In machine learning using a generative adversarial network, where the generator is a second neural network related to the generation of a generated 3D human model and the classifier is a first neural network that has learned evaluation criteria for 3D human models through machine learning, when the classifier identifies the generated data as correct, this is equivalent to the first neural network that learned the evaluation criteria through machine learning evaluating the generated 3D human model generated by the second neural network as a 3D human model that has the attributes desired by the user (orderer). In other words, the second neural network performs machine learning with the goal of generating a generated 3D human model that will be evaluated by the first neural network as having the attributes desired by the user (orderer). This enables the generation unit using the second neural network to generate the desired 3D human model in a more detailed state.

The machine learning using the above-mentioned generative adversarial network is unsupervised learning, so it does not increase the designer's workload. Therefore, according to the third disclosure, it is possible to provide a 3D human model generation device that enables designers involved in computer graphics to generate the 3D human model desired by the user (client) with minimal effort.

The fourth disclosure provides a 3D human model generation device according to any one of the first to third disclosures, wherein the attribute information includes information about the characteristics of the person.

According to the fourth disclosure, the first neural network can be trained to learn evaluation criteria related to the characteristics of a person desired by a user (client). This allows the first neural network to train evaluation criteria for a 3D human model by machine learning, so as to focus on and evaluate aspects of the 3D human model that correspond to the characteristics of a person desired by the user (client who orders a 3D human model from a designer). Furthermore, the designer can easily generate and output a 3D human model that captures the characteristics of a person desired by the user (client) of the 3D human model, via information related to the characteristics of the person contained in the attribute information.

The fifth disclosure provides a 3D human model generation device according to any one of the first to fourth disclosures, wherein the attribute information includes information regarding the purpose of use of the 3D human model.

According to the fifth disclosure, the first neural network can be trained by machine learning evaluation criteria related to the intended use of the 3D human model. Furthermore, the generation unit can generate a 3D human model to be generated according to the intended use. Therefore, the designer can generate the 3D human model desired by the user (client) simply by specifying the intended use, without having to go to the trouble of specifying attribute information for the 3D human model according to the intended use one by one. Therefore, designers working in computer graphics can generate the 3D human model desired by the user (client) with less effort.

The sixth disclosure provides a 3D human model generation device according to any one of the first to fifth disclosures, wherein the attribute information includes identification information that enables identification of the single type of output 3D human model or multiple types of output 3D human models.

According to the sixth disclosure, instead of using updated character model configuration information, it is possible to associate evaluation information updated by user evaluation information with the output 3D character model using identification information included in attribute information. This allows the first neural network to machine-learn evaluation criteria related to the output 3D character model without updating the character model configuration information. If the character model configuration information is not updated, the designer will not have to wait for updates to the character model configuration information, reducing the designer's workload.

The seventh disclosure changes the category of the configuration related to the first disclosure.

This disclosure provides a 3D character model generation device that enables computer graphics designers to generate the 3D character models desired by users (orderers) of 3D character models with as little effort as possible.

1 is a block diagram showing an example of the configuration of a three-dimensional human model generating device according to an embodiment of the present invention; FIG. 10 is a diagram illustrating an example of a character model information table. FIG. 10 is a diagram illustrating an example of a neural network table. 10 is a flowchart illustrating an example of a three-dimensional human model generation process using the three-dimensional human model generation device of the present embodiment. 10 is a flowchart illustrating an example of machine learning processing when a generative adversarial network is used. FIG. 10 is an explanatory diagram showing an example of generating a three-dimensional human model using attribute information. FIG. 10 is an explanatory diagram showing an example of machine learning using user evaluation information.

Below, an example of a preferred embodiment for implementing this disclosure will be described with reference to the drawings. Note that this is merely an example, and the technical scope of this disclosure is not limited to this example.

<Three-dimensional human model generating device 1>
1 is a block diagram showing an example of the configuration of a three-dimensional human model generating device 1 (hereinafter also simply referred to as "generating device 1") according to this embodiment. The generating device 1 is configured to include at least a control unit 11, a storage unit 12, a display unit 13, and an input unit 14.

[Control unit 11]
The control unit 11 includes a CPU (Central Processing Unit), a RAM (Random Access Memory), a ROM (Read Only Memory), and the like.

The control unit 11 loads a specified program, and the memory unit 12, display unit 13, and input unit 14 work together as necessary to realize the software configuration elements of the generation device 1, such as the evaluation learning unit 111, generation unit 112, evaluation unit 113, output unit 114, reception unit 115, update unit 116, and machine learning unit 117.

[Storage unit 12]
The storage unit 12 is a device in which data and files are stored, and has a data storage unit such as a hard disk, semiconductor memory, recording medium, memory card, etc. The storage unit 12 may have a mechanism that enables connection to storage devices or storage systems such as NAS (Network Attached Storage), SAN (Storage Area Network), cloud storage, file server, and distributed file system via a network.

The memory unit 12 stores a control program executed by the microcomputer, a character model information table 121, a neural network table 122, etc. Signals and information are input to the memory unit 12 from the output unit 114 of the control unit 11. In addition, signals and information are transmitted from the memory unit 12 to the receiving unit 115 of the control unit 11.

[Personal Model Information Table 121]
Fig. 2 is a diagram showing an example of the character model information table 121. The character model information table 121 shown in Fig. 2 stores information linking an "ID" for identifying character model information, "character model configuration information" having the shape, the surface texture, and the surface color of the three-dimensional character model, "evaluation information" regarding the three-dimensional character model configured by the character model configuration information, and "attribute information" of the three-dimensional character model.

Furthermore, the character model information table 121 is configured to allow the addition of information linking an ID, character model configuration information, and attribute information. Because the character model information table 121 is configured to allow the addition of information, the update unit 116 can update the character model configuration information using the output 3D character model.

(Person model configuration information)
The character model configuration information is information that can configure a three-dimensional character model. The three-dimensional character model is not particularly limited as long as it is a three-dimensional model that visually represents a person or a part of a person. The three-dimensional character model may be either male or female, and may include at least one of three-dimensional models of clothing, accessories, and props worn by the person. The format of the character model configuration information is not particularly limited, and character model configuration information in a conventional format can be stored.

The character model configuration information preferably includes at least one of the following: information about the three-dimensional shape of the three-dimensional character model, information about the texture of the surface of the three-dimensional character model, and information about the color of the surface of the three-dimensional character model. By including this information, model training data can be provided to the first neural network N1, which will be described later. This allows the first neural network N1 to machine-learn evaluation criteria related to at least one of the three-dimensional shape of the three-dimensional character model, the surface color of the three-dimensional character model, and the surface color of the three-dimensional character model.

Furthermore, by including various information related to these three-dimensional human models, the human model information table 121 can provide model training data to the second neural network N2, which will be described later. This allows the second neural network N2 to machine-learn a generation algorithm related to at least one of the three-dimensional shape of the three-dimensional human model, the surface color of the three-dimensional human model, and the surface color of the three-dimensional human model. For example, the row with ID "M1" stores human model configuration information that configures a three-dimensional human model with a shape of an elongated face and surface color and texture of light-colored hair, beard, and light-colored skin.

The character model configuration information may include information that can be used to configure an output three-dimensional character model to be output by the output unit 114, which will be described later. The information that can configure an output three-dimensional character model is not particularly limited as long as it can configure an output three-dimensional character model, and similar to the character model configuration information, configuration information that can configure an output three-dimensional character model in a format of conventional technology can be stored. Because the character model configuration information may include information that can configure an output three-dimensional character model, the update unit 116, which will be described later, can update the character model configuration information using the output three-dimensional character model.

(Evaluation information)
The evaluation information is user evaluation information on a three-dimensional human model configured based on the human model configuration information. In the example shown in Fig. 2, an evaluation indicating the degree to which the three-dimensional human model conforms to the evaluation criteria desired by the user, on a scale of 100 points, is stored as the evaluation information. The format of the evaluation information is not limited to an evaluation indicating the degree to which the three-dimensional human model conforms to the evaluation criteria on a scale of 100 points, and may include, for example, information on the score given by the user to the three-dimensional human model, information listing the attributes that the three-dimensional human model has, information listing the attributes that the three-dimensional human model does not have, etc.

The evaluation criteria may also include a history of three-dimensional character models that the user has purchased since placing an order. The routes by which the user's evaluation is input to the receiving unit 115 include a route from the input unit 14 to the receiving unit 115, and a route by which the evaluation is input to the receiving unit 115 via the external device 120 and the communication unit 15. The user is the person who places an order for a three-dimensional character model with a designer.

(Attribute information)
The attribute information is information about the attributes of the three-dimensional human model configured by the human model configuration information. The format of the attribute information is not particularly limited, and may include any information related to the attributes of the three-dimensional human model.

The attribute information preferably includes information related to "characteristics desired by the user." In the example shown in Figure 2, information related to character characteristics such as "light-colored hair," "long, thin face," and "beard" is stored as attribute information linked to ID "M1." By including information related to character characteristics, the first neural network N1, described below, can learn by machine learning evaluation criteria related to the character characteristics desired by the user (client). This allows the first neural network N1 to learn by machine learning evaluation criteria for the 3D character model so that the evaluation focuses on parts corresponding to the character characteristics desired by the user who uses the 3D character model (the client who orders the 3D character model from the designer). Furthermore, via the information related to character characteristics included in the attribute information, the designer can easily generate and output a 3D character model that captures the character characteristics desired by the user (client) of the 3D character model.

The attribute information preferably includes information regarding the intended use of the three-dimensional human model. In the example shown in Figure 2, information regarding the intended use of the three-dimensional human model, "for advertising in North America," is stored as attribute information linked to ID "M1." By including information regarding the intended use of the three-dimensional human model, the first neural network N1, described below, can learn by machine learning the evaluation criteria regarding the intended use of the three-dimensional human model desired by the user (client). The intended use includes the region in which the advertisement will be made, such as "for advertising in North America," "for advertising in Asia," "for advertising in Europe," or "for advertising in Africa."

As a result, the first neural network N1 can machine-learn evaluation criteria for three-dimensional human models, so as to evaluate whether the three-dimensional human model is suitable for the intended use desired by the user who will use the three-dimensional human model (the client who orders the three-dimensional human model from the designer). Furthermore, the designer can easily generate and output a three-dimensional human model that is suitable for the intended use desired by the user (client) of the three-dimensional human model, using the information regarding the intended use of the three-dimensional human model contained in the attribute information.

By including information about the intended use of the 3D character model in the attribute information, the generation device 1 can receive the intended use of the 3D character model via the attribute information. This means that the designer can generate a 3D character model to be generated that meets the intended use, simply by specifying the intended use, without having to go to the trouble of specifying attribute information for the 3D character model in accordance with the intended use. This means that designers working in computer graphics can generate the 3D character model desired by the user (client) with less effort.

The attribute information preferably includes identification information that makes the output 3D human model identifiable. In the example shown in FIG. 2, information that makes the output 3D human model identifiable, "output 3D human model: P1," is stored as attribute information linked to ID "M5." This makes it possible to associate the evaluation information updated by the user evaluation information with the output 3D human model using the identification information included in the attribute information, instead of the updated human model configuration information. This allows the first neural network N1 to machine-learn evaluation criteria for the generated 3D model without updating the human model configuration information stored in the human model information table 121. Not updating the human model configuration information eliminates the need for the designer to wait for updates to the human model configuration information, reducing the designer's workload.

In the character model information table 121 shown in FIG. 2, IDs are stored, making it easy to obtain and update the information stored in the character model information table 121. Since evaluation information is stored, model training data can be provided to the first neural network N1, which will be described later. This allows the first neural network N1 to learn evaluation criteria for three-dimensional character models by machine learning. Since attribute information is stored, the evaluation learning unit 111 can allow the first neural network N1 to learn evaluation criteria related to the attributes that a client, who is a user of the three-dimensional character model, requests of the three-dimensional character model from a graphic designer.

The generation unit 112 can generate multiple 3D human model outputs that have the attributes that the user desires in a 3D human model. The output unit 114 can then select and output multiple types of 3D human model outputs that satisfy the evaluation criteria for attributes, i.e., are recognized as the desired 3D human model, from the multiple 3D human model outputs generated by the generation unit 112.

For example, the row for ID "M1" stores character model configuration information for a three-dimensional character model having a long, thin face and a surface color and texture of light hair, a beard, and light skin, evaluation information "90" for the three-dimensional character model configured using this character model configuration information, and attribute information for the three-dimensional character model configured using this character model configuration information: "light hair," "long, thin face," "beard," "light skin," and "for advertising in North America." In this way, the character model information table 121 can store character model configuration information, evaluation information, and attribute information in association with each other.

[Neural Network Table 122]
3 is a diagram showing an example of the neural network table 122. The neural network table 122 stores at least a first neural network N1 (corresponding to ID "N1") related to evaluation criteria for three-dimensional human models and a second neural network N2 (corresponding to ID "N2") related to generation of three-dimensional human models.

The format of the first neural network N1 and the second neural network N2 is not particularly limited, and may include a data structure containing information regarding an activation function that determines the presence or absence of a signal output by an artificial neuron (also referred to as a node) of the neural network, or the strength of the signal; information regarding a weight matrix (also referred to as a weight parameter) that represents the weighting of signals input to the artificial neuron of the neural network (also referred to as synaptic connection strength or simply connection strength); information regarding a bias vector that assigns a reference weight (also referred to as a bias parameter or simply bias) to signals input to the artificial neuron of the neural network; information regarding the connection relationships between the artificial neurons of the neural network; etc.

The types of the first neural network N1 and the second neural network N2 stored in the neural network table 122 are not particularly limited, and may include, for example, a feedforward neural network (also referred to as a FFNN), a convolutional neural network (CNN, or ConvNet), a deep stacking network (DSN), an RBF network (also referred to as a radial basis function network), a recurrent neural network (RNN), and a modular neural network. Various prior art neural networks, such as the first neural network N1 and the second neural network N2, can be stored in the neural network table 122.

Looking more closely at the configuration of the neural network table 122, the neural network table 122 shown in Figure 3 stores an "ID" that identifies the neural network and the neural network's weight matrices A and B, linked together. Storing the ID makes it easy to obtain and update the information stored in the neural network table 122. Storing the neural network's weight matrices A and B makes it possible to generate or evaluate neural networks, and to have the neural network perform machine learning.

The row with ID "N1" stores a first neural network N1 represented by weight matrix A1 (a matrix including elements a111 to a179) and weight matrix B1 (a matrix including elements b111 to b194). By storing weight matrix A1 and weight matrix B1, the evaluation learning unit 111 can have the first neural network N1 learn the evaluation criteria by machine learning. Furthermore, by storing weight matrix A1 and weight matrix B1, the evaluation unit 113 can evaluate the generated 3D human model.

The row with ID "N2" stores a second neural network N2 represented by weight matrix A2 (a matrix including elements a211 to a249) and weight matrix B2 (a matrix including elements b211 to b297). By storing weight matrix A2 and weight matrix B2 in the neural network table 122, the generation unit 112 can generate a generated 3D human model. Furthermore, by storing weight matrix A2 and weight matrix B2, the machine learning unit 117 can perform machine learning on the second neural network N2.

By storing the first neural network N1 in the neural network table 122, the evaluation learning unit 111 can have the first neural network N1 learn evaluation criteria for 3D human models. By storing the second neural network N2 in the neural network table 122, the generation unit 112 can generate multiple types of generated 3D models. Furthermore, by storing the second neural network N2 in the neural network table 122, the machine learning unit 117 can perform unsupervised machine learning on the second neural network N2. Machine learning is a method of learning without providing correct answers to the training data.

[Display section 13]
There is no particular limitation on the type of the display unit 13. Examples of the display unit 13 include a monitor, a touch panel, a projector, and a video card that displays the output 3D human model on the external device 120. The display unit 13 receives signals and information output from the output unit 114 of the control unit 11.

[Input unit 14]
There is no particular limitation on the type of input unit 14. Examples of the input unit 14 include a keyboard, a mouse, a touch panel, and a communication device that receives input from the external device 120. The signals and information input from the input unit 14 are input to the receiving unit 115 of the control unit 11. The user's order information input to the input unit 14 is sent to the receiving unit 115.

[Communication unit 15]
The generation device 1 may include a communication unit 15 that communicates between the generation device 1 and the external device 120. The communication unit 15 and the external device 120 are connected via a network 123. The network 123 may be constructed wirelessly or wired. The external device 120 may be either singular or plural. By including the communication unit 15, the generation device 1 can transmit the output 3D human model output by the output unit 114 to the external device 120. By including the communication unit 15, the generation device 1 can receive commands related to the generation of the 3D human model from the external device 120. By including the communication unit 15, the generation device 1 can receive and use human model configuration information, evaluation information, attribute information, user evaluation information, etc. from the external device 120.

[External device 120]
The external device 120 may be either a mobile terminal or a fixed terminal. The external device 120 is operated by a user and includes a control circuit 125, an input unit 126, a display unit 127, a memory unit 128, and a communication unit 129. The control circuit 125 is a computer having an input port, an output port, and a central processing circuit. The control circuit 125 is connected to the input unit 126, the display unit 127, the memory unit 128, and the communication unit 129 so as to be able to transmit and receive signals. Examples of the input unit 126 include a keyboard, a mouse, a touch panel, and a communication device that receives input from the external device 120. Examples of the display unit 127 include a monitor, a touch panel, a projector, and a video card that displays a 3D human model to be output on the external device 120. The memory unit 128 is a device that stores data and files and includes a data storage unit such as a hard disk, semiconductor memory, recording medium, or memory card. The communication unit 129 is connected to the network 123.

<Main Flowchart of Generation Process Executed by Generation Device 1>
4 is an example of a main flowchart showing the procedure of the generation process in this modified example. A preferred procedure of the generation process performed by the generation device 1 will be described below with reference to FIG.

[Step S1: Machine Learning of Evaluation Criteria]
First, the control unit 11 executes the evaluation learning unit 111 in cooperation with the memory unit 12, refers to the character model information table 121, and executes a process of having the first neural network N1 learn the evaluation criteria for multiple types of three-dimensional character models by machine learning (step S1).

The algorithm used to machine-learn the evaluation criteria in step S1 is not particularly limited, and any known machine learning algorithm for supervised learning of neural networks can be used, such as backpropagation, stochastic gradient descent such as the Widrow-Hoff method (also known as the delta rule), gradient descent, online learning, batch learning, logistic function, sigmoid function, maximum function, etc.

The control unit 11 executes the evaluation learning unit 111 and acquires character model configuration information, evaluation information, and attribute information from the character model information table 121. The control unit 11 then causes the first neural network N1 stored in the neural network table 122 to machine-learn evaluation criteria using the character model configuration information, evaluation information, and attribute information, and updates the first neural network N1 stored in the neural network table 122. In other words, the control unit 11 acquires the character model configuration information, evaluation information, and attribute information from the character model information table 121, and causes the first neural network N1 to machine-learn evaluation criteria using the character model configuration information, evaluation information, and attribute information as example training data.

As a result, the first neural network N1 can learn by machine learning the evaluation criteria related to the attributes that the client, a graphic designer who is the user of the 3D human model, desires in the 3D human model. After completing the processing of step S1, the control unit 11 makes the decision of step S2.

[Step S2: Determine whether attribute information has been received]
The control unit 11 executes the generation unit 112 in cooperation with the storage unit 12 and the input unit 14, and determines whether or not the receiving unit 115 has received the attribute information (step S2). If the attribute information has been received, the control unit 11 determines Yes in step S2 and performs the process of step S3. If the attribute information has not been received, the control unit 11 determines No in step S2 and performs the process of step S7.

[Step S3: Generate a 3D human model to be generated]
The control unit 11 executes the generation unit 112 in cooperation with the storage unit 12, and acquires the second neural network N2 stored in the neural network table 122. Then, the control unit 11 generates a plurality of types of three-dimensional human models to be generated using the attribute information received in step S2 and the second neural network N2 (step S3).

The generation unit 112 can generate multiple random 3D human model instances using the attribute information received in step S2 and the second neural network N2 represented by weighting matrices A2 and B2. This allows the designer to generate multiple 3D human model instances with the attributes desired by the user (orderer) of the 3D human model with minimal effort. After processing step S3, the control unit 11 proceeds to step S4.

[Step S4: Evaluate the generated 3D human model]
The control unit 11 executes the evaluation unit 113 in cooperation with the storage unit 12, and acquires the first neural network N1 stored in the neural network table 122. Then, the control unit 11 uses the attribute information received in step S2 and the first neural network N1 to evaluate the multiple types of generated 3D human models generated in step S3 (step S4).

In this way, the control unit 11 evaluates the multiple types of generated 3D human models generated in step S3 using the attribute information received in step S2 and the first neural network N1 that learned the evaluation criteria by machine learning in step S1. Therefore, the multiple types of generated 3D human models generated by the generation unit 112 are given an evaluation regarding the attributes indicated by the attribute information received in step S2. After processing step S4, the control unit 11 proceeds to step S5.

[Step S5: Output the output 3D human model]
The control unit 11 executes the output unit 114 in cooperation with the memory unit 12 and the display unit 13, and outputs, as an output three-dimensional character model, a three-dimensional character model that has undergone a predetermined evaluation in step S4 from among the multiple types of three-dimensional character models generated in step S3 (step S5).

The output unit 114 can select and output 3D human models that have been evaluated based on the attributes indicated in the attribute information received in step S2. Because the series of processes related to the output of the 3D human models is performed automatically, the designer can output the 3D human model desired by the user (client) with less effort. After processing step S5, the control unit 11 proceeds to step S6.

[Step S6: Update person model configuration information]
The control unit 11 executes the update unit 116 in cooperation with the memory unit 12, links the new ID with the character model configuration information including information capable of configuring the output three-dimensional character model output in step S5, and attribute information including information capable of identifying the output three-dimensional character model, and adds them to the character model information table 121 (step S6).

The updating unit 116 links the new ID with the character model configuration information including information capable of configuring the output 3D human model output in step S5, and attribute information including information capable of identifying the output 3D human model, and adds the linked information to the character model information table 121, thereby updating the character model configuration information stored in the character model information table 121. This allows the designer to prepare, with little effort, training data that will serve as a model for use in the processing performed in step S1 and the processing performed in step S9, described below. In other words, the designer can, with little effort, obtain evaluation criteria that better reflect the attributes that the user (the person ordering the designer) desires in a 3D human model. This allows the designer to, with little effort, output the 3D human model that the user (the person ordering the designer) desires.

In this way, the control unit 11 determines whether or not attribute information has been received in step S2, and performs subsequent processing. When attribute information has not been received, the control unit 11 can skip the processing executed in steps S3 to S5, which generate, evaluate, and output a three-dimensional person model according to the attribute information. The control unit 11 can also skip step S6, which relates to updating the person model configuration information. This reduces the amount of processing executed by the control unit 11. After processing step S6, the control unit 11 proceeds to step S7.

[Step S7: Determine whether evaluation information has been received from the user]
The control unit 11 executes the receiving unit 115 in cooperation with the storage unit 12 and the input unit 14 to determine whether or not evaluation information has been received from the user (step S7). If evaluation information has been received from the user, the control unit 11 determines Yes in step S7 and proceeds to step S8. If evaluation information has not been received from the user, the control unit 11 determines No in step S7 and returns to step S1.

[Step S8: Update evaluation information]
The control unit 11 executes the update unit 116 in cooperation with the memory unit 12, and updates the evaluation information regarding the output three-dimensional human model output in S5, which is stored in the human model information table 121, based on the evaluation information received from the user (step S8).

By updating the evaluation information stored in the character model information table 121 with the evaluation information, the designer can, with little effort, prepare training data that will serve as a model for use in the processing performed in step S1 and the processing performed in step S9, described below. In other words, the designer can, with little effort, obtain evaluation criteria that better reflect the attributes that the user (the person ordering the designer) desires in a three-dimensional character model. This allows the designer to, with little effort, output the three-dimensional character model that the user (the person ordering the designer) desires. After processing step S8, the control unit 11 proceeds to step S9.

[Step S9: Execute machine learning by unsupervised learning]
The control unit 11 executes the machine learning unit 117 in cooperation with the storage unit 12, and acquires the character model configuration information, evaluation information, and attribute information from the character model information table 121. Then, using the character model configuration information, evaluation information, and attribute information, the control unit 11 causes the second neural network N2 stored in the neural network table 122 to perform unsupervised machine learning to generate a three-dimensional character model, thereby updating the second neural network N2 stored in the neural network table 122 (step S9).

The algorithm used for machine learning in step S9 is not particularly limited as long as it is unsupervised learning, and can be any known machine learning algorithm for unsupervised learning of neural networks, using one or more of cluster analysis, principal component analysis, vector quantization, self-organizing maps, generative adversarial networks (also known as GANs), deep belief networks (also known as DBNs), Hebb's law, etc.

The machine learning performed in step S9 enables the second neural network N2 to generate a more detailed 3D character model related to the attribute information. As a result, in step S3, the generation unit 112 uses the second neural network N2 that has undergone machine learning to generate a more detailed 3D character model desired by the user (the person ordering the designer). The unsupervised machine learning performed in step S9 can be performed without externally provided model training data, so the generation device 1 can generate a more detailed 3D character model desired by the user (the person ordering the designer) without increasing the designer's workload.

Therefore, by performing the processes from step S1 to step S9, it is possible to provide a 3D human model generation device 1 that enables a computer graphics designer to generate the 3D human model desired by the user (client) with as little effort as possible. After performing the process of step S9, the control unit 11 returns to step S1.

In this way, the control unit 11 determines whether or not evaluation information has been received from the user, and when the control unit 11 receives evaluation information from the user, it can perform "unsupervised machine learning." Note that "unsupervised machine learning" will be described later. Furthermore, when the control unit 11 has not received evaluation information from the user, it can repeat the 3D person model generation process without performing "unsupervised machine learning."

It is preferable that the algorithm used for machine learning in step S9 includes a generative adversarial network, as this can reduce the variability in the output from the neural network in response to the input to the neural network, and various functions, including the generation and evaluation of three-dimensional human models, can be realized on the neural network.

When generating a 3D human model using the second neural network N2, the received attribute information corresponds to the input to the second neural network N2, and the generated 3D human model corresponds to the output from the second neural network N2. Because it is possible to reduce the variation in output relative to the input, the second neural network N2 can generate only the desired 3D human model in a precise state for the received attribute information, and avoid generating 3D human models that are similar but not identical to the desired 3D human model. In other words, it becomes possible to generate a 3D human model related to the attribute information in a more precise state.

Figure 5 is a flowchart showing an example of the flow of the machine learning process executed in step S9 when a generative adversarial network is used. Below, with reference to Figure 5, a preferred procedure for the machine learning process executed in step S9 of Figure 4 will be explained, taking as an example a case where the algorithm used to perform the machine learning includes a generative adversarial network.

[Step S11: Obtain attribute information]
The control unit 11 executes the machine learning unit 117 in cooperation with the storage unit 12, and sets the update count, which indicates the number of times the second neural network N2 has been updated, to 0. The control unit 11 further acquires attribute information (step S11). By setting the update count to 0, the control unit 11 can execute machine learning a predetermined number of times in succession. By acquiring the attribute information, the control unit 11 can execute machine learning according to the attribute information. After processing step S11, the control unit 11 proceeds to step S12.

[Step S12: Generate a 3D human model to be generated]
The control unit 11 cooperates with the storage unit 12 to execute the machine learning unit 117, and generates a person model to be generated using the second neural network N2 (step S12).

In machine learning using a generative adversarial network, a generator generates data and a discriminator discriminates the generated data. The generator then performs machine learning with the goal of having the discriminator discriminate the generated data as correct. The control unit 11 generates a generated person model using the second neural network N2, allowing the second neural network N2 to function as a generator in the generative adversarial network. After processing step S12, the control unit 11 proceeds to step S13.

[Step S13: Evaluate the generated 3D human model]
The control unit 11 cooperates with the storage unit 12 to execute the machine learning unit 117, and evaluates the generated character model generated in step S12 using the first neural network N1 (step S13).

In step S13, the control unit 11 evaluates the generated person model generated in step S12 by the machine learning unit 117 using the first neural network N1. As a result, the first neural network N1 can be used as a classifier in a generative adversarial network. Therefore, the machine learning unit 117 can perform machine learning using a generative adversarial network, with the second neural network N2 as the generator in the generative adversarial network and the first neural network N1 as the classifier in the generative adversarial network. After processing step S13, the control unit 11 makes a judgment in step S14.

[Step S14: Determine whether the attribute-related evaluation criteria are met]
The control unit 11 executes the machine learning unit 117 in cooperation with the storage unit 12 to determine whether the generated 3D human model generated in step S12 satisfies the evaluation criteria for attributes (step S14). If the evaluation criteria for attributes are satisfied, the control unit 11 determines Yes in step S14 and proceeds to step S15. On the other hand, if the evaluation criteria for attributes are not satisfied, the control unit 11 determines No in step S14 and proceeds to step S16.

[Step S15: Update the second neural network N2 to generate more similar 3D human models to be generated]
The control unit 11 executes the machine learning unit 117 in cooperation with the memory unit 12, and updates the second neural network N2 so as to generate more generated three-dimensional character models that are similar to the generated character model generated in step S12 (step S15).

By updating the second neural network N2 so that the machine learning unit 117 generates more 3D character models similar to the 3D character model generated in step S12, the generation unit 112 using the second neural network N2 will generate more 3D character models similar to the 3D character model that received a predetermined evaluation in step S13. In other words, the generation unit 112 will be able to generate more detailed 3D character models that the user (orderer) desires. After processing step S15, the control unit 11 proceeds to step S17.

[Step S16: Update the second neural network N2 to generate fewer similar 3D human models]
The control unit 11 executes the machine learning unit 117 in cooperation with the memory unit 12, and updates the second neural network N2 so as to generate fewer generated three-dimensional character models that are similar to the generated character model generated in step S12 (step S16).

By updating the second neural network N2 so that the machine learning unit 117 generates fewer 3D character models similar to the generated character model generated in step S12, the generation unit 112 using the second neural network N2 will generate fewer 3D character models similar to the generated 3D character models that did not receive the specified evaluation in step S13. In other words, the generation unit 112 will be able to generate more detailed 3D character models that the user (orderer) desires.

In this way, in step S14, the control unit 11 determines whether the 3D human model to be generated generated in step S12 satisfies the evaluation criteria for attributes. The machine learning unit 117 can then update the second neural network N2 so as to change the rate at which similar 3D human models are generated, depending on whether the evaluation criteria for attributes are satisfied. After processing step S15 or step S16, the control unit 11 proceeds to step S17.

[Step S17: Increment the update count by 1]
The control unit 11, in cooperation with the storage unit 12, executes the machine learning unit 117 and increments the update count by 1 (step S17). By incrementing the update count by 1, the number of times the second neural network N2 has been updated is reflected in the update count. After processing step S17, the control unit 11 proceeds to step S18.

[Step S18: Determine whether the number of updates is equal to or greater than a predetermined number]
The control unit 11 executes the machine learning unit 117 in cooperation with the storage unit 12 to determine whether the number of updates is equal to or greater than a predetermined number (step S18). If the number of updates is less than the predetermined number, the control unit 11 determines No in step S18.
The process returns to step S12. If the number of updates is equal to or greater than the predetermined number, the control unit 11 determines "Yes" in step S18, and the process returns to step S1 in FIG.

In this way, the control unit 11 determines whether the number of updates is equal to or greater than a predetermined number and changes the processing accordingly, thereby enabling machine learning using a generative adversarial network to be performed consecutively exactly the same number of times as the predetermined number.

By the control unit 11 executing the processes of steps S11 to S18, the machine learning unit 117 can perform machine learning using a generative adversarial network, with the second neural network N2 as the generator of the generative adversarial network and the first neural network N1 as the classifier of the generative adversarial network. In machine learning using a generative adversarial network, the generator generates data and the classifier classifies the generated data. The generator then performs machine learning with the goal of having the classifier classify the generated data as correct data.

In machine learning using a generative adversarial network, where the generator is a second neural network N2 related to generating a generated 3D human model and the classifier is a first neural network N1 that has learned evaluation criteria for 3D human models through machine learning, when the classifier identifies generated data as correct data, this is equivalent to the first neural network N1 that learned the evaluation criteria through machine learning evaluating the generated 3D human model generated by the second neural network N2 as a 3D human model that has the attributes desired by the user (orderer). In other words, the second neural network N2 performs machine learning with the aim of generating a generated 3D human model that will be evaluated by the first neural network N1 as having the attributes desired by the user (orderer). This enables the generation unit 112 using the second neural network N2 to generate a desired 3D human model with greater detail.

The machine learning using the above-mentioned generative adversarial network is unsupervised learning, so it does not increase the designer's workload. Therefore, it is possible to provide a 3D human model generation device 1 that allows computer graphics designers to generate the 3D human model desired by the user (client) with minimal effort.

The control unit 11 causes the second neural network N2 stored in the neural network table 122 to perform unsupervised machine learning to generate a three-dimensional human model, and after updating the second neural network N2 stored in the neural network table 122, returns the process to step S1 in Figure 4.

<Example of use of the 3D human model generating device 1>
Next, an example of how to use the three-dimensional human model generating device 1 according to this embodiment will be described with reference to FIGS.

[Learn the assessment criteria]
A designer who has received an order from a user (orderer) to generate a three-dimensional human model with desired attributes causes the generation device 1 to execute a three-dimensional human model generation process. The generation device 1 executes the evaluation learning unit 111, and uses the human model configuration information, evaluation information, and attribute information stored in the human model information table 121 to cause the first neural network N1 to perform machine learning of evaluation criteria.

An example of evaluation criteria for machine learning by the first neural network N1 will be described using Figure 2. The human model information table 121 shown in Figure 2 stores, with regard to attribute information related to the shape of a three-dimensional human model, human model configuration information M1 and M3 having the attribute information "long and thin face," and human model configuration information M2 and M4 having the attribute information "wide face."

The evaluation learning unit 111 trains the first neural network N1 to use the characteristics of the person model configuration information common to the person model configuration information M1 and M3, such as "long and thin face" and "light skin," as evaluation criteria for the attribute information "long and thin face," and trains the first neural network N1 to use the characteristics of the person model configuration information common to the person model configuration information M2 and M4, such as "wide face," "dark skin," "no beard," and "dark hair," as evaluation criteria for the attribute information "wide face."

[Receive attribute information]
FIG. 6 is an explanatory diagram showing an example of generating a three-dimensional human model using attribute information. Using the example shown in FIG. 6, three-dimensional human model generation using attribute information will be described. The designer transmits (inputs) attributes desired by a user (orderer) to the generation device 1 as attribute information. In the example shown in FIG. 6, the description will be given assuming that the attribute desired by the user (orderer) is the attribute information "wide face." The generation device 1 executes the generation unit 112 in step S2 of FIG. 4 and receives the attribute information "wide face" transmitted by the designer.

[Output the output 3D human model]
The generation unit 112 generates random three-dimensional models G1 to G4 to be generated using the second neural network N2 and the attribute information "wide face."

The evaluation unit 113 evaluates each of the generated 3D models G1 to G4 using the evaluation criteria learned by machine learning by the first neural network N1 and the attribute information "wide face."

The generated 3D model G1 satisfies four of the four evaluation criteria for a "wide face" - "wide face," "dark skin," "no beard," and "dark hair" - and therefore receives a high rating of "100" out of 100. Furthermore, the generated 3D model G3 satisfies two of the four evaluation criteria for a "wide face" - "wide face" and "dark hair" - and therefore receives a high rating of "50."

On the other hand, the generated 3D model G2 only satisfies "no beard" out of the four evaluation criteria for a "wide face," and therefore receives a low score of "25." Furthermore, the generated 3D model G4 only satisfies "dark skin" out of the four evaluation criteria for a "wide face," and therefore receives a low score of "25."

The output unit 114 outputs the generated 3D models G1 and G3 that have received a predetermined evaluation as output 3D person models P1 and P2. The generation unit 112 generates a generated 3D model with the attribute information "wide face", which is evaluated by the evaluation unit 113, and the generated 3D models G1 and G3 that have received a predetermined evaluation are output, so that output 3D person models P1 and P2, which are desired 3D person models with wide facial features, are output, as shown in FIG. 6.

In this way, the entire process of outputting the output 3D human model is carried out automatically, allowing the designer to output a 3D human model with the "wide face" attribute information desired by the user (client) with less effort.

[Receive user evaluation information]
FIG. 7 is an explanatory diagram showing an example of machine learning using user evaluation information. Machine learning using user evaluation information will be described using the example shown in FIG. 7 . A user (orderer) transmits (inputs) evaluations of the output 3D human models P1 and P2 to the generating device 1 as user evaluation information. In the example shown in FIG. 7 , the user (orderer) gave the output 3D human model P1 a rating of "80" points and the output 3D human model P2 a rating of "90" points. In other words, the user (orderer) gave a higher rating to the output 3D human model P2, which has a wider face, making an evaluation different from the evaluation using the evaluation criteria. The user (orderer) then transmitted these evaluations to the generating device 1 as user evaluation information S1 and S2. The generating device 1 executes the receiving unit 115 and the updating unit 116 to receive user evaluation information and update the human model information table 121.

[Perform unsupervised machine learning]
The generating device 1 executes the machine learning unit 117, which performs unsupervised machine learning to generate a 3D human model for the second neural network N2. The machine learning unit 117 then updates the second neural network N2 stored in the neural network table 122.

As described above, the evaluation unit 113 gave a higher evaluation to the output three-dimensional human model P1 than to the output three-dimensional human model P2, based on the information stored in the initial human model information table 121 shown in FIG. 2.

However, as shown in FIG. 7(a), what the generating device 1 received was user evaluation information S1 that gave a rating of "80" points to the output three-dimensional human model P1, and user evaluation information S2 that gave a rating of "90" points to the output three-dimensional human model P1. In other words, the user (orderer) gave a higher rating to the output three-dimensional human model P2, which has a wider face, which is a different evaluation from the evaluation using the evaluation criteria. This is because the evaluation criteria based on machine learning using the initial human model information table 121 shown in FIG. 2 included "dark skin," "no beard," and "dark hair," which are characteristics of the three-dimensional human model configuration information that are not directly related to a "wide face," in the evaluation criteria for the attribute information "wide face."

The update unit 116 updates the character model information table 121 using the output three-dimensional character models P1 and P2 and the user evaluation information S1 and S2. The machine learning unit 117 then uses the updated character model information table 121 to perform unsupervised machine learning on the second neural network N2. As a result, the second neural network N2 performs machine learning on the characteristics of the three-dimensional character model related to the attribute information "wide face" based on the above-mentioned evaluation. The second neural network N2 then becomes a neural network that has more deeply learned the characteristics of the three-dimensional character model configuration information related to "wide face."

As shown in FIG. 7(b), after machine learning, the generation unit 112 generates three-dimensional human models G11, G12, G13, and G14 to be generated, each having a variety of skin colors, beards, and hair colors, as well as wide faces. In other words, the unsupervised machine learning performed by the machine learning unit 117 makes it possible to generate a more detailed three-dimensional human model desired by the user (the person ordering the designer). Because unsupervised machine learning can be performed without externally provided model training data, the generation device 1 can generate a more detailed three-dimensional human model desired by the user (the person ordering the designer) without increasing the designer's workload.

Therefore, the 3D human model generation device 1 of this embodiment allows designers working in computer graphics to generate the 3D human model desired by the user (client) with as little effort as possible.

[Another example of information stored in the character model information table 121]
The attribute information stored in the character model information table 121 may include the following information. The character model information table 121 can store, for example, the body shape of a three-dimensional character model desired by a user. Furthermore, when a three-dimensional character model desired by a user is to be displayed on the display unit 13 or the display unit of the external device 120, the character model information table 121 can store background information suitable for the three-dimensional character model desired by the user. The character model information table 121 can store the type, design, color, and pattern of clothing suitable for the three-dimensional character model desired by the user. Furthermore, the character model information table 121 can store the type, design, and pattern of bags suitable for the three-dimensional character model desired by the user. Furthermore, the character model information table 121 can store the type and color of accessories suitable for the three-dimensional character model desired by the user. Furthermore, the character model information table 121 can store hair and makeup suitable for the three-dimensional character model desired by the user. Furthermore, the character model information table 121 can store footwear suitable for the body shape, clothing, and the like of the three-dimensional character model desired by the user. Footwear includes shoes, geta (wooden clogs), sandals, boots, pumps, mules, sandals, etc. Furthermore, the character model information table 121 can store the facial expression, age, gender (male or female), and face and body angles of the 3D character model desired by the user. Gender information can also include masculine or feminine elements. The face angle includes, for example, the angle of the face centerline relative to the torso centerline. The body angle includes, for example, the pose.

The body shape of the three-dimensional human model includes, for example, thin, standard, chubby, etc. The background of the three-dimensional human model includes, for example, mountains, the sea, urban areas, indoors, outdoors, sky, etc. The types of clothing worn by the human model include shirts, blouses, sweaters, knitwear, jackets, skirts, cardigans, vests, pants, etc. Accessories include necklaces, earrings, bracelets, sunglasses, hats, etc. Hair and makeup include short cuts, bobs, long hair, perms, straight hair, etc.

The type of clothing worn by a 3D character model, the design, color, and pattern of the clothing, the type of bag carried by the 3D character model, the design of the bag, the type of accessories worn by the 3D character model, the color of the accessories, the hair and makeup of the 3D character model, the footwear, the facial expression of the 3D character model, the age, gender (male or female), and the angle of the face and body can all be created by a person using Apparel 3D. With Apparel 3D, a person analyzes the ordered 3D character model and generates the 3D character model by switching and modifying various pre-prepared patterns on the base model.

The type of clothing, design, color, pattern, bag type, bag design, accessory type, color, hair and makeup, shoes, facial expression, age, gender (male or female), and face and body angle of the 3D character model desired by the user may be created by an AI system. The AI system analyzes the ordered 3D character model and generates images of clothing and other items with a design that matches the preferences of the orderer. The apparel 3D system and AI system may be constructed by the 3D character model generation device 1, or may be constructed using an external device 120.

[Methods for using other examples of information in the flowcharts of Figures 4, 5, 6, 7(a) and 7(b)]
In the flowcharts of Figures 4, 5, 6, 7(a) and 7(b), it is possible to process and make judgments at each step using information contained in the attribute information, such as the model's body shape, the model's background, the type of clothes worn by the human model, the design of the clothes, the color of the clothes, the pattern of the clothes, the type of bag carried by the human model, the design of the bag, the type of accessories worn by the human model, the color of the accessories, the human model's hair and makeup, footwear, the facial expression of the three-dimensional human model, the age, gender (male or female), and the angle of the face and body.

[supplementary explanation]
Although the embodiments and various modifications of the present disclosure have been described above, the present disclosure is not limited to these above-described embodiments and various modifications. Furthermore, the effects described in the embodiments and various modifications of the present disclosure are merely a list of the most preferable effects resulting from the present disclosure, and the effects of the present disclosure are not limited to those described in the embodiments and various modifications of the present disclosure.

The above-described embodiments have been described in detail to clearly explain the present disclosure, and are not necessarily limited to those including all of the described configurations. It is also possible to replace part of the configuration of one embodiment with the configuration of another embodiment, or to add the configuration of another embodiment to the configuration of one embodiment. Furthermore, the control unit 11 can improve the image quality and accuracy of the image of the 3D human model displayed on the display unit 13 or the display unit 127 by increasing the accuracy of image encoding and the number of times and amount of information in machine learning. Step S1 in FIG. 4 is the machine learning step. Step S6 is the update step. Determining Yes in step S7 is the reception step. Step S5 is the output step. Step S3 is the generation step. Step S4 is the evaluation step. Step S1 is the evaluation learning step.

1 Generation device 11 Control unit 111 Evaluation learning unit 112 Generation unit 113 Evaluation unit 114 Output unit 115 Reception unit 116 Update unit 117 Machine learning unit 12 Storage unit 121 Person model information table 122 Neural network table 13 Display unit 14 Input unit G1, G2, G3, G4 Generated three-dimensional person models G11, G12, G13, G14 Generated three-dimensional person models P1, P2 Output three-dimensional person models S1, S2 User evaluation information N1 First neural network N2 Second neural network

Claims (9)

a database that stores attribute information of a three-dimensional human model and evaluation criteria related to the attributes of a generated three-dimensional human model generated using a neural network for each of the attribute information;
an evaluation unit that receives the attribute information desired by a user, and evaluates the plurality of types of generated three-dimensional human models using the evaluation criteria corresponding to the attribute information and the attribute information associated with each of the plurality of types of generated three-dimensional human models;
an output unit that outputs , as an output three-dimensional human model, one or more types of the three-dimensional human model that have received a relatively high predetermined evaluation by the evaluation unit from among the plurality of types of the three-dimensional human model to be generated;
A three-dimensional human model output device comprising: an update unit that registers, for each of the output three-dimensional human models, human model configuration information capable of configuring a three-dimensional human model for the output three-dimensional human model and human model information associated with the attribute information ;
The three-dimensional human model output device according to claim 1 , further comprising : an evaluation learning unit that performs machine learning of the evaluation criteria based on the attribute information linked to each of the plurality of registered human model information . 3. The three-dimensional human model output device according to claim 2, wherein the evaluation learning unit learns one or more attribute information commonly associated with the plurality of registered human model information associated with specific attribute information as an evaluation criterion for the specific attribute information. a generation unit that receives the attribute information desired by a user and generates the three-dimensional human model to which a plurality of pieces of attribute information are linked;
The three-dimensional human model output device according to claim 2 , wherein the update unit links and registers the output three-dimensional human model and the attribute information used in generating the three-dimensional human model to be generated. The three-dimensional human model output device according to claim 2 , further comprising a machine learning unit capable of performing the machine learning by unsupervised learning for both the neural network used to generate the three-dimensional human model and the machine learning of the evaluation criterion. The three-dimensional human model output device according to claim 1 , wherein the attribute information includes information relating to characteristics of the person of the three-dimensional human model. The three-dimensional human model output device according to claim 1 , wherein the attribute information includes information relating to a purpose of use of the three-dimensional human model. 3D human model output device,
a storing step of storing attribute information of the three-dimensional human model and evaluation criteria relating to the attributes of the generated three-dimensional human model generated using the neural network in a database for each of the attribute information;
an evaluation step of receiving the attribute information desired by a user, and evaluating the plurality of types of three-dimensional human models to be generated in accordance with the evaluation criteria corresponding to the attribute information and the attribute information associated with each of the plurality of types of three-dimensional human models to be generated;
an output step of outputting , as an output three-dimensional human model, one or more types of the three-dimensional human model to be generated that have relatively high predetermined evaluations made in the evaluation step , from among the plurality of types of the three-dimensional human model to be generated;
A program that executes the following. The computer
a storage process for storing attribute information of the three-dimensional human model and evaluation criteria related to the attributes of the generated three-dimensional human model generated using the neural network in a database for each of the attribute information;
an evaluation process of receiving the attribute information desired by a user, and evaluating the plurality of types of three-dimensional human models to be generated according to the evaluation criteria corresponding to the attribute information and the attribute information associated with each of the plurality of types of three-dimensional human models to be generated;
an output process of outputting , as an output three-dimensional human model, one or more types of the three-dimensional human model that have received a relatively high predetermined evaluation in the evaluation process from among the plurality of types of the three-dimensional human model to be generated;
A three-dimensional human model output method that executes the above.