DESIGN DEVICE AND METHOD FOR PROVIDING A 3D MODEL FOR THE DESIGN DEVELOPMENT OF AN OBJECT

Description

[0001] The invention relates to a method for providing a 3D model for the design development of an object. Furthermore, the invention relates to a design device configured to carry out the method.

[0002] Creating 3D CAD geometry is a crucial milestone in the design process, providing a clear understanding of proportions, technical constraints, and the overall product / object. Currently, a designer uses CAD (computer-aided design) software to generate 3D geometry based on 2D sketches or photographs. This 3D geometry is refined through multiple design reviews with engineering and design teams until a certain level of technical and aesthetic maturity is reached. Typically, 2D sketches are created, often requiring several adjustments to each sketch. A 3D model is then created using CAD based on these 2D sketches. Renderings are then generated from this 3D model for decision-makers, a process that often involves several iterations.

[0003] This approach has several disadvantages. Particularly with 2D sketches, multiple sketches per perspective are necessary for a single 3D object, and any design change requires individual adjustments to each perspective. This means that every incremental design modification necessitates numerous adjustments to the sketches. Furthermore, a media break occurs when transitioning to the 3D model based on the 2D sketches. Creating the 3D models typically requires specialized expertise. Again, incremental design changes result in time-consuming manual adjustments to the 3D model, and extensive communication between designers and CAD professionals is essential. This is because a 2D sketch often lacks sufficient information to accurately describe and, consequently, create a 3D model.

[0004] Furthermore, the publication by S. Radhakrishnan et al, "Creative Intelligence - Automating Car Design Studio with Generative Adversarial Networks (GAN)", 16th European Conference - Computer Vision - ECCV 2020, Cornell University Library, 201, Ithaca, NY 14853, pp. 160 - 175 (24.08.2018), describes a system based on generative adverse networks (GANS) to create novel car designs from a minimal design studio sketch.

[0005] Therefore, the purpose of the invention is to simplify and / or accelerate the design development of an object.

[0006] This problem is solved by the independent patent claims. Advantageous embodiments of the invention are disclosed in the dependent patent claims, the following description, and the figures.

[0007] The invention provides a method for generating a 3D model for the design development of an object by performing the method on a computing device. The method comprises the following steps, which are performed by a computing device, in particular a processor and / or microchip. The method includes specifying an object type to at least one trained artificial neural network, wherein the at least one artificial neural network is trained with a plurality of basic image data of objects of this object type. This means that an object type, which defines the type of object to be designed, can first be specified. For example, the object type can be specified to a computing device or a computer program configured to perform the method.The object type can be used to define, for example, whether the program should create a vehicle component as an object, such as a body, rims, an interior component, and / or components from other areas outside of automotive technology. The object type can be specified to the computer program or the trained artificial neural network, for example, using a graphical user interface (GUI).

[0008] At least one artificial neural network can, for example, include generative adversarial networks (GANs). An artificial neural network, often referred to simply as an artificial network, is a network of artificial neurons used for machine learning and artificial intelligence. Various computer-based problems can be solved using an artificial neural network. Ultimately, an artificial neural network is an algorithm that can interpret different data sources, such as image data, and extract information or patterns from them. These extracted information or patterns can then be applied to previously unknown data. This ultimately allows for data-driven predictions, enabling the generation of artificial data from the artificial neural network's training data.The artificial neural network can be trained using image data, especially basic image data of objects of the specified object type.

[0009] In other words, the base image data can be training data for the artificial neural network, derived from multiple objects of the same type. For example, the object for which a new design is to be developed could be a motor vehicle, and the design process could specify an object type to be designed by the method, such as a vehicle body or the vehicle's rims. Consequently, the artificial neural network can then preferably be trained using base image data from different vehicle bodies or rims, particularly from a plurality of different motor vehicles. This base image data can therefore comprise a set of multiple images, each representing different configurations of the object type.For example, a training process can be implemented prior to the described procedure step, in which the GAN is trained using a dataset of base image data from real or artificial objects of the object type. By applying the GAN, the artificial intelligence can then generate new images based on the image properties of the base image data.

[0010] As a further process step, an image data set of the specified object type is generated by the trained artificial neural network, whereby the artificial neural network provides image data of the object type from several different perspectives as an image data set, which are determined from the basic image data with which the artificial neural network was trained.

[0011] In other words, the trained artificial neural network can create new image data based on the trained base image data, particularly for multiple different perspectives or views of the object of the specified object type. The artificial neural network that was trained with base image data of the same object type can be selected to generate the artificial image data in the respective different perspectives of the object, where "image data" refers to 2D views from the different perspectives.

[0012] In the next step, the generated image dataset is automatically checked for the presence of a design selection criterion by another neural network. If the design selection criterion is not met, image data from one or more perspectives is regenerated by modifying an image property criterion. The design selection criterion could, for example, include whether image properties in the respective perspectives of the image data match or are consistent with each other. For instance, if there were a color difference between two perspectives, this would constitute an inconsistency, and in this case, the design selection criterion would not be met.In the example of a vehicle body, different model types might be depicted, for instance, a sedan in one perspective and a station wagon in another. This would constitute an inconsistency and thus fail to meet the design selection criterion. Alternatively or additionally, the design selection criterion can also depend on a designer's design sensibility, i.e., it can be an aesthetic criterion. In this case, the designer can, for example, manually approve or reject the image data.

[0013] If the design selection criterion is not available, the image data of one or more perspectives can be regenerated by the artificial neural network by modifying an image property criterion. This image property criterion can be a rule, an algorithm, data, and / or information that can be used to modify the image data. In particular, several of the generated image data from a single perspective can be regenerated by modifying the image property criterion through interpolation and / or recombination of image properties or partial image data from the generated image data. For example, when using a GAN, one or more latent vectors can be modified to change the image data of at least one perspective.

[0014] Interpolation preferably involves interpolation between at least two provided image data points from the image dataset. A particular configuration of the GAN is preferably used, namely a style-based GAN. Mathematically speaking, the first artificial neural network of the style-based GAN serves to map random vectors into an intermediate latent space. For this purpose, linear interpolation can be performed, for example, between two or more latent vectors of the image dataset, each of which is assigned to one of the at least two provided image data points. In this process, a further latent vector is calculated, which is assigned to additional image data points that differ from the at least two provided image data points. This additional latent vector is assigned to image data points that lie between the two provided latent vectors of the image dataset.Mathematically speaking, interpolation generates a straight line through the image dataset, which can be understood as a kind of image data space or latent image space, connecting the two provided image data points. Each location on this line represents further image data points located at that position within the image dataset. Interpolation thus provides additional image data that can be extracted from the image dataset.

[0015] Interpolation can be used to generate a weighted average between at least two provided image data sets. For example, information, particularly design properties, from one image data set can be adopted at 70 percent, while information from the other image data set can be adopted at only 30 percent. The combination of this respective information then results in additional image data that differs from the provided image data. Ultimately, interpolation can generate any combination of the provided image data, including combinations selected by a user through appropriate modification of the latent vectors.

[0016] Recombination is based on the understanding that image data describes various image properties, each of which is described by sub-image data within the image data. For example, the provided image data can have up to 18 image properties. These image properties can be understood as different levels of information or layers within the image data. Image properties can be described as abstract features. One or more image properties can, for example, at least partially describe and / or determine the number of spokes on a rim, the color scheme of the rim, and / or the shape of the spokes.

[0017] During recombination, it is possible, for example, to select that certain primary image properties should be adopted from the image data, while differing secondary image properties should be adopted from other image data. In this way, for instance, the color of a spoke can be extracted from the first set of image data, and the number of individual spokes of the rim from the second set of image data, for the generation of new image data. Thus, during recombination, the selected image properties of the respective provided image data can be combined.

[0018] In other words, recombination is based on the fact that the image properties of the image data are generated at different levels within the artificial neural network. In the case of a GAN, these properties are generated by the GAN's so-called generator. The generator is the GAN's artificial neural network, which, after a training process with the base image data, generates artificial image data, for example, of rims, thus creating the image dataset. Preferably, during the generator's training process, multiple latent vectors are always used for a given object type or object style, for example, a given rim style, mathematically speaking, preferably at all levels of the generator.In an exemplary case of the training process, where for the sake of clarity only two latent vectors are described here, a first latent vector can be assigned to levels 1 to k, and a second vector to levels k+1 to n, where n represents the maximum number of levels. When applying the trained artificial neural network, a different latent vector can be used at each level. With preferably up to 18 such levels in the generator, a multitude of possible, each distinct, combinations of image properties result. Preferably, several provided image data sets are combined during recombination, and thus the trained generator has several different latent vectors, in particular up to n different levels, where n represents the number of levels in the trained generator.Ultimately, this can, for example, generate one or more new image data sets with combinations of image properties that best correspond to a designer's personal preferences.

[0019] The image properties that are adjusted by modifying the latent vectors can preferably be known, allowing for the targeted regeneration or modification of image data properties. For example, shapes, proportions, colors, and / or individual details in the respective image data can be specifically altered.

[0020] If the design selection criterion is met, a digital 3D model is then generated from the image dataset using a reconstruction algorithm. The computing device uses this algorithm to calculate the digital 3D model from the image data of several different perspectives. Preferably, the multiple different perspectives can be predefined such that they enable the reconstruction algorithm to calculate the 3D model. The number of perspectives required can depend on the object, with at least two perspectives, preferably 5 to 10, being used from different angles to the object. Known reconstruction techniques can be used to reconstruct the digital 3D model from the image dataset, which comprises 2D views of the object. These techniques can be based, in particular, on neural networks and / or photogrammetry and / or differentiable rendering.

[0021] Finally, the generated digital 3D model is provided as a design tool for the object, whereby the object and / or a physical 3D model is produced from the provided digital 3D model using a manufacturing facility. The digital 3D model can, for example, be provided to a CAD program for further development of the object, such as a vehicle. Alternatively, the digital 3D model can be provided to a manufacturing facility, which creates a physical 3D model from the digital 3D model (in this case, the manufacturing facility could be a 3D printer), and / or the digital 3D model can be provided to a manufacturing facility, such as a production line for vehicle bodies, to produce the object according to the specifications of the digital 3D model.

[0022] The process involves creating the object and / or a physical 3D model from the provided digital 3D model using a manufacturing system. This means that the data from the digital 3D model can control a manufacturing system that produces a physical object. Preferably, the manufacturing system can produce the object automatically using the data or information from the digital 3D model, i.e., without additional manual steps. The manufacturing system can, for example, be based on an additive manufacturing process, in particular comprising a 3D printer or a photolithography system. Alternatively or additionally, the manufacturing system can include or be the injection molding machine. Thus, for example, a prototype can be produced based on the digital 3D model provided by the process.Ultimately, objects of various designs, such as a vehicle body or a rim for a motor vehicle, can be manufactured simply and cost-effectively. Thus, the method according to the invention supports the entire manufacturing process of the object, from design development to the production of the newly designed object.

[0023] The invention offers the advantage of accelerating and streamlining design development. New designs can be derived quickly, thus reducing the time required for design development. Furthermore, rapid iterations and enhancements of existing designs, as well as targeted adaptation to existing branding or the various design languages ​​of a brand, become possible. This also eliminates the need for media breaks, meaning no switch from, for example, paper drawings to digital object representations. Ultimately, it enables abstract yet targeted interpolation and / or recombination of artificial base image data, as specified by the designer, to easily and quickly control designs, 3D geometries, and styles—something that is often very complex or impossible with established CAD / CAS approaches.

[0024] The invention also includes embodiments that offer additional advantages.

[0025] One embodiment provides for the provision of multiple artificial neural networks, with each artificial neural network being trained for a given perspective of the object using base image data of objects of this type from that perspective. Similarly, each artificial neural network can then generate only the image data of the perspective it was trained on to create the image dataset. In other words, multiple trained neural networks can be trained for a given 2D view, with the base image data used for training being provided to each neural network from the same perspective.Each of these trained neural networks generates only one 2D view from that perspective, while all trained neural networks together provide the multiple perspectives that can subsequently be used to reconstruct the 3D model. For example, the artificial neural networks can be generative models, in particular StyleGAN 2 or SWAGAN. This offers the advantage that a preferred embodiment for generating image data from multiple perspectives can be provided.

[0026] Another embodiment provides that at least one artificial neural network is trained using 2D views of 3D objects of the object type, with the 2D views being provided from several predefined perspectives of the respective 3D object. In other words, the artificial neural network is trained using 3D objects, whereby the respective 3D object can be rotated into several predefined perspectives, and the resulting 2D view is then used for training. For example, real or digital 3D objects or models can be used, which are digitally rotated into the different perspectives. For instance, the 3D object for training the artificial neural network can be stored in a CAD tool, which then rotates the 3D object into predefined perspectives to provide 2D views for training.Thus, the artificial neural network can learn the 2D representations of 3D objects and subsequently generate new image data from these learned 2D views when creating the image dataset. Alternatively, a heterogeneous dataset of base image data of the 3D object, with respect to perspective, can be used to train at least one artificial neural network. Style-NERF, for example, can be used as the artificial neural network in this context. This implementation offers the advantage that the neural network can be trained using consistent base image data of a 3D object, thereby reducing inconsistencies when generating new image data.

[0027] Preferably, the design selection criterion is met at least if an image property is consistent across multiple perspectives of the image data. In particular, it can be checked whether the image property of the image data is the same across multiple perspectives or whether, for example, there is a stylistic inconsistency, especially inconsistent shapes, between the different perspectives. For example, in the case of a vehicle body, it can be checked whether a station wagon is depicted in one perspective and a sedan in another, in which case the design selection criterion would not be met. The design selection criterion can preferably be checked automatically, in particular by another neural network that examines the generated image data accordingly. Alternatively or additionally, the design selection criterion can be checked by a designer who can then approve the image data set.This offers the advantage that the design development of an object can be further simplified.

[0028] Another embodiment uses a generative adversarial network (GAN) as the artificial neural network. The generating generic network can alternatively be described as a type of generative deep neural network. The GAN is an algorithmic architecture that employs two artificial neural networks. These two artificial neural networks ultimately serve to generate synthetic, i.e., artificial, new datasets, particularly image datasets.

[0029] One artificial neural network in the GAN is called the generator and creates the image dataset. The other artificial neural network in the GAN is called the discriminator and evaluates the generated image data from the dataset. Mathematically speaking, the generator typically maps a vector of latent variables to a desired result space. During training, the generator learns to produce the image data according to a predefined distribution. The discriminator is trained to distinguish these generator outputs from data within the same predefined distribution. The predefined distribution is the base image data. At the end of the training, the generator is configured to produce image data that the discriminator cannot distinguish from the base image data.The aim is for the generated distribution, i.e., the generated image data set, to gradually approximate a real distribution, meaning, for example, realistic images of objects such as a vehicle body.

[0030] A preferred configuration of the GAN is a style-based GAN, such as StyleGAN 1, StyleGAN 2, and / or SWAGAN (Style and Wavelet Based GAN). Mathematically speaking, the first artificial neural network of the style-based GAN, for example, StyleGAN 1, maps random vectors to an intermediate latent space. The output of the first artificial neural network is fed to the second artificial neural network via various adaptive instance normalization layers (AdaIN) to ensure that the image data generated by the second neural network is realistic representation of the object. Alternatively or additionally, the output of the first artificial neural network can be fed to the second artificial neural network via similarly functioning layers.The similarly functioning layers, for example in the case of StyleGAN 2, are so-called demodulations and / or modulations. The second neural network can be described as a synthesis network. The generation of image data by the synthesis network occurs in stages, starting with an image resolution of 4x4 pixels, which is gradually refined by the artificial neural network. The latent vectors are used as style vectors at different levels. The number of levels can be up to 18, depending on the desired size of the generated image data. The final output is, for example, a portable network graphic (PNG image file). The influence of a specific style vector or latent vector on the generated image data depends, among other things, on the chosen depth of the respective AdaIN layer.

[0031] The image data generated by the generator possesses image properties, which, when using style-based GANs, can be referred to as styles. Image properties assigned to the first levels primarily govern global properties of the imaged object. Image properties assigned to the lower levels primarily determine local properties of the object, such as color. Depending on the chosen resolution, the synthesis network can process image data with up to 18 style vectors, i.e., 18 image properties. These concatenated style vectors can also be referred to as DNA vectors.

[0032] A prerequisite for successful training of the style-based GAN is the availability of suitable base image data. This data should ideally contain a wide variety of object configurations. Ideally, the base image data should also be evaluated using principal component analysis (PCA) in Fourier space. The results of this evaluation are then transformed into a two-dimensional image space using T-distributed stochastic neighbor embedding (T-SNE). This allows for the early identification and, for example, the elimination of particularly similar image data.

[0033] Based on artificial intelligence methods, training the GAN, especially the style-based GAN, with a suitable basic image dataset for the desired object ensures that reliable, realistic new image data of the object is generated, thus providing a suitable image dataset for the process.

[0034] Another embodiment provides that the image data of one or more perspectives are regenerated by changing the image property criterion through interpolation and / or recombination of the image data. In other words, at least two different image data sets can be used to generate the new and modified image data set in order to perform a variation of the respective individual image data sets provided. Preferably, the trained artificial neural network can first generate several image data sets for each perspective, whereby image properties of two or more of the respective image data sets of a given perspective can be regenerated by interpolation and / or recombination through changing the image property criterion, thereby generating entirely new image data. The functionality of the GAN for interpolation and recombination has already been described above.Starting with the image dataset generated by at least one artificial neural network, combination image data can be created by selecting at least two image data points and interpolating and / or recombinating them. The combination image data preferably describes an image of an object newly designed by the artificial neural network, that is, an object previously unknown to the artificial neural network. Alternatively, the latent vectors can be randomly modified in latent space to generate new image data. This embodiment offers the advantage of allowing targeted variation of the image properties, thus enabling fine-tuning of the design development.

[0035] Preferably, the image property criterion is modified by adjusting latent vectors. In other words, the image property criterion is defined by the latent vectors. These latent vectors can access different levels of the trained base image data and thus image properties that the generated image data should exhibit for the image dataset. In particular, this allows, for example, a designer to selectively modify an image property in one or more image data sets from different perspectives without having to create entirely new 2D sketches in a time-consuming manner. Preferably, the modification of the latent vectors can be performed using a graphical user interface that provides the dependencies between the latent vectors.This means that the graphical user interface can indicate which image property is modified by changing the respective latent vector. In particular, the designer can move a point in the latent space described by the latent vectors along a principal component, thereby changing the expression of specific properties of the object. This offers the advantage of easily modifying image data. Preferably, image properties that are modified by changing the image property criterion include the following: a dimension of the object, such as its dimensions or size; the object's proportions; its stylistic direction; its color; and / or the shapes of individual details of the object.These can be adjusted, for example, using the latent vectors described above, by targeting different levels of the base images.

[0036] Another aspect of the invention relates to a design device for providing a 3D model for the design development of an object, wherein the design device comprises a computing device and a manufacturing system, the computing device being configured to perform a method according to one of the preceding embodiments. For example, the design device can comprise a computing device, in particular a computer, on which the at least one trained artificial neural network can be operated. Preferably, this can be controlled via a graphical user interface and have at least one screen on which individual steps of the method can be monitored and / or the provided digital 3D model can be displayed. The design device can also comprise a manufacturing system to which the digital 3D model is transmitted and which can subsequently produce a physical 3D model.This offers the same advantages and variations as the previous method.

[0037] For use cases or application situations that may arise during the procedure and are not explicitly described here, it may be provided that, according to the procedure, an error message and / or a request for user feedback is issued and / or a default setting and / or a predetermined initial state is set.

[0038] The invention also includes a control device for the design device. The control device can comprise a data processing device or a processor unit configured to carry out an embodiment of the method according to the invention. For this purpose, the processor unit can comprise at least one microprocessor and / or at least one microcontroller and / or at least one FPGA (Field Programmable Gate Array) and / or at least one DSP (Digital Signal Processor). Furthermore, the processor unit can comprise program code configured to carry out the embodiment of the method according to the invention when executed by the processor unit. The program code can be stored in a data memory of the processor unit. Alternatively or additionally, the method, which can be in the form of program code, can be provided by cloud computing.

[0039] The invention also includes further developments of the design device according to the invention, which have features already described in connection with the further developments of the method according to the invention. For this reason, the corresponding further developments of the design device according to the invention are not described again here.

[0040] Preferably, the method can be used to produce a model of a motor vehicle component, in particular for a passenger car, truck, passenger bus and / or a motorcycle.

[0041] As a further solution, the invention also includes a computer-readable storage medium comprising instructions that, when executed by a computer or a computer network, cause it to execute an embodiment of the method according to the invention. The storage medium can, for example, be configured at least partially as a non-volatile data storage medium (e.g., as flash memory and / or as an SSD - solid state drive) and / or at least partially as a volatile data storage medium (e.g., as RAM - random access memory). The computer or computer network can provide a processor circuit with at least one microprocessor. The instructions can be provided as binary code or assembly language and / or as source code in a programming language (e.g., C).

[0042] The invention also includes combinations of the features of the described embodiments. The invention thus also includes realizations that each exhibit a combination of the features of several of the described embodiments, provided that the embodiments have not been described as mutually exclusive. The scope of protection of the invention is defined by the appended claims.

[0043] The following are exemplary embodiments of the invention described. This is illustrated by: Fig. 1 a schematically represented design device according to an exemplary embodiment; Fig. 2 a schematic process diagram according to an exemplary embodiment.

[0044] The exemplary embodiments described below are preferred embodiments of the invention. In these exemplary embodiments, the described components each represent individual features of the invention, which can be considered independently of one another and each further develops the invention independently. Therefore, the disclosure is intended to include combinations of features of the embodiments other than those shown. Furthermore, the described embodiments can also be supplemented by further features of the invention already described.

[0045] In the figures, identical reference symbols denote functionally equivalent elements.

[0046] In Fig. 1 A design device 1 is outlined, which can be configured to perform a method for providing a 3D model for the design development of an object. The design device 1 can comprise a computing device 2, in particular a computer, which has a processor 3 and a storage medium 4. For example, program code for at least one artificial neural network and / or basic image data of objects can be stored on the storage medium 4, which is, for example, a hard drive of the computer. The processor 3 can be configured to execute the program code of the artificial neural network in order to generate a digital 3D model 5 of an object, where the object in this example can be a motor vehicle. Alternatively, the processor 3 can, for example, be provided in a cloud.

[0047] The computing device 2 can be equipped with a display device 6 for displaying the digital 3D model 5, wherein a method for providing the digital 3D model 5 can be monitored and / or adjusted via the display device 6, preferably via a graphical user interface. In particular, the device 1 can include input devices 7 by means of which settings, especially on the graphical user interface, can be changed. The input devices can, for example, include a computer keyboard, a computer mouse, and / or touch-sensitive input fields.

[0048] The design device 1 preferably additionally comprises a manufacturing system 8, which can be configured as an additive manufacturing system 8, for example as a 3D printer or as a photolithography system. In particular, the manufacturing system 8 can be configured to convert the digital 3D model 5 into a real 3D model 9, i.e., in this example, a real model of the motor vehicle.

[0049] The respective data can be transmitted between the individual devices of the design device 1, in particular between the computing device 2, the display device 6 and / or the manufacturing plant 8, by wired and / or wireless means.

[0050] In Fig. 2A method for providing a 3D model 5 of an object according to an exemplary embodiment is described. In a first method step S1, an artificial neural network 10 can be trained, wherein the artificial neural network 10 is preferably a generative adversarial network, a so-called GAN (generative adversarial network). Such a GAN 10 comprises two artificial neural networks 11, 12, of which the first network is referred to as the generator 11 and the second as the discriminator 12. The GAN 10 can be trained either with several base image data 13 of an object, wherein the base image data are preferably provided from different perspectives of the object and a separate GAN 10 is trained for each of these perspectives. For this purpose, a so-called StyleGAN 2, SWAGAN or StyleNERF is suitable, for example.Alternatively, a 3D object can be provided as base image data 13, which can be rotated into several different 2D views, thus providing predefined perspectives of the respective 3D object. In this example, the base image data could be a vehicle body.

[0051] However, other object types can also be trained using appropriate base image data.

[0052] In a process step S2, an object type 18 can be specified to the trained GAN 10, for which the GAN 10 is to generate images, wherein the GAN 10 was preferably previously trained for the specified object type 18 using the basic image data 13 of object type 18. In this example, object type 18 can therefore be specified as a vehicle body to be provided as a design development.

[0053] In step S3, the GAN 10 can generate an image data set 15 of the specified object type 18, whereby the trained GAN 10 interpolates and / or recombines image features learned from the basic image data 13 or training data to generate new image data 14 of object type 18, in particular image data 14 for several different perspectives of the object.

[0054] In step S4, the generated image data set 15 can be checked to see if a design selection criterion is present 16 or not 17. The design selection criterion could be, for example, whether an image property is consistent in the image data 14. For instance, it can be checked whether the color of the car body is consistent across the respective perspectives and / or whether shapes match. A designer can also check whether a style meets their expectations or whether the image data 14 should be modified. If this is the case and the design selection criterion is not present 17, one or more perspectives can be regenerated by the GAN 10 in step S5 by changing an image property criterion.In this process, one or more latent vectors can be adjusted to control different layers of the base image data, thereby modifying dimensions, proportions, stylistic direction, color, and / or the shapes of individual details. Preferably, image properties of the image data, which can exist for each perspective, are interpolated and / or recombined by changing the image property criterion to selectively adjust the image properties. Consequently, for example, a designer can create entirely new image data 14 or modify only individual image properties of the image data 14 in one or more perspectives to arrive at a new image data set 15. The image data 14 of the image data set 15 provides 2D views from the different perspectives, which can be displayed and checked, for example, by the display device 6.

[0055] If the image data is consistent and the designer is satisfied with the representations of the image data 14 in the different perspectives, the design selection criterion 16 can be met, and in step S6 the digital 3D model 5 can then be calculated using a reconstruction algorithm. The reconstruction algorithm can, for example, be another neural network that creates the 3D model 5 from the views of the image data 14 and / or can be used to calculate the 3D model 5 using methods of photogrammetry and / or differentiable rendering.

[0056] Finally, in step S7, the generated digital 3D model 5 can be provided as a design development for the object, in this case the vehicle body, wherein preferably a manufacturing plant 8 can produce a real 3D model 9, for example a plastic model produced by a 3D printer.

[0057] Overall, the examples show how the invention can achieve a GAN-based generation of 3D geometries.

Claims

1. A method of providing a 3D model (5) for design development of an object, by performing the method on a computing device (2), comprising the steps of: - specifying (S2) an object type (18) for at least one trained artificial neural network (10), wherein the at least one artificial neural network (10) is trained with a plurality of basic image data (13) of objects of this object type (18); - generating (S3) an image data set (15) of the predetermined object type by the trained artificial neural network (10), wherein the artificial neural network (10) provides an image data set (15) of image data (14) of the object type (18) from a plurality of different perspectives which are determined from the acquired basic image data (13), with which the artificial neural network (10) was trained; - automatic checking (S4) the generated image data set (15) by a further neural network for the presence of a design selection criterion, wherein if the design selection criterion is not met (17), image data (14) of one or more perspectives are newly generated by means of a change of an image property criterion (S5); - generating (S6) a digital 3D model (5) from the image data set (15) for which the design selection criterion is met (16) by a reconstruction algorithm, wherein the computing device (2) calculates, by means of the reconstruction algorithm, the digital 3D model (5) from the image data (14) of the plurality of different perspectives; - providing (S7) the generated digital 3D model (5) as design development for the object, wherein the object and / or a real 3D model (9) is generated from the provided digital 3D model (5) by means of a production system (8).

2. The method according to claim 1, wherein multiple artificial neural networks (10) are provided, wherein a respective artificial neural network (10) is trained for a predetermined perspective on said object for said perspective by means of basic image data (13) of objects of said object type.

3. The method of claim 1, wherein the at least one artificial neural network (10) is trained by means of respective 2D views of 3D objects (13) of the object type, wherein the 2D views are provided from a plurality of predetermined perspectives on the respective 3D object.

4. The method according to any one of the preceding claims, wherein the design selection criterion is met (16) at least if an image property is consistent in the plurality of perspectives of the image data.

5. The method according to any one of the preceding claims, wherein a generative adversarial network, GAN, is used as the artificial neural network (10).

6. The method according to any one of the preceding claims, wherein the image data (14) of one or more perspectives is regenerated by changing the image property criterion by interpolating and / or recombining the image data (14).

7. The method according to any one of the preceding claims, wherein the image property criterion is varied by means of latent vector matching.

8. The method according to any one of the preceding claims, wherein image properties adapted by changing the image property criterion comprise: - a dimension of the object; - proportions of the object; - a stylistic direction of the object; - a color of the object; - shapes of individual details of the object.

9. A design device (1) for providing a 3D model for a design development of an object, wherein the design device comprises a computing device (2) and a production system (8), wherein the computing device (2) is configured to carry out a method according to any one of the preceding claims.