Method and system for generating 3D building models based on integration of voxel GAN and conditional GAN

Abstract

The present invention provides a building model generation method comprising receiving layout data configured in two dimensions for a specific building, generating a plurality of first voxel models of a three-dimensional shape for a specific building using a pre-trained voxel generation model, generating a plurality of second voxel models of a three-dimensional shape satisfying the conditions of the specific building according to the layout data using a pre-trained conditional generation model, and optimizing the plurality of first voxel models and the plurality of second voxel models based on a plurality of predetermined optimization items for the specific building to generate a building model of a three-dimensional shape corresponding to the layout data.

Description

Technology Field

[0001] The present invention relates to a method and system for generating a 3D building model based on the integration of a voxel GAN ​​and a conditional GAN. Background Technology

[0002] Generative Adversarial Networks (GANs) can be trained to synthesize high-quality data, such as images or videos, and are utilized in various fields. For example, GANs can be applied to a wide range of tasks, from realistic image synthesis to data augmentation and image restoration. In these GANs, two neural networks—a generator and a discriminator—achieve a specific goal through learning based on an adversarial process.

[0003] In particular, in the field of 3D modeling, generative adversarial networks can be trained to generate 3D models based on various input forms such as 2D images and voxel grids. Therefore, these generative adversarial networks are being applied in various fields such as urban planning, architecture, games, and virtual reality.

[0004] In this regard, Conditional Generative Adversarial Networks (cGANs) are implemented based on the framework of Generative Adversarial Networks to condition the generation process according to additional information, such as class labels or images. The problem to be solved

[0005] The present invention relates to a method and system for generating a 3D building model based on the integration of a voxel GAN ​​and a conditional GAN, which can effectively improve the structural accuracy and quality of the 3D building model. means of solving the problem

[0006] To solve the problem described above, a building model generation method according to the present invention may include: receiving layout data configured in two dimensions for a specific building; generating a plurality of first voxel models of a three-dimensional shape for a specific building using a pre-trained voxel generation model; generating a plurality of second voxel models of a three-dimensional shape satisfying the conditions of the specific building according to the layout data using a pre-trained conditional generation model; and optimizing the plurality of first voxel models and the plurality of second voxel models based on a plurality of predetermined optimization items for the specific building to generate a building model of a three-dimensional shape corresponding to the layout data.

[0007] In addition, the building model generation system according to the present invention includes a storage unit in which layout data configured in two dimensions for a specific building is stored; and a control unit that generates a building model of a three-dimensional shape based on the layout data. The control unit generates a plurality of first voxel models of a three-dimensional shape for a specific building using a pre-learned voxel generation model, generates a plurality of second voxel models of a three-dimensional shape that satisfy the conditions of the specific building according to the layout data using a pre-learned conditional generation model, and optimizes the plurality of first voxel models and the plurality of second voxel models based on a plurality of optimization items predetermined for the specific building to generate the building model of a three-dimensional shape corresponding to the layout data.

[0008] Additionally, a program stored on a computer-readable recording medium according to the present invention is executed by one or more processes in an electronic device and is a program stored on a computer-readable recording medium, wherein the program may include instructions for performing the steps of: receiving layout data configured in two dimensions for a specific building; generating a plurality of first voxel models of a three-dimensional shape for a specific building using a pre-learned voxel generation model; generating a plurality of second voxel models of a three-dimensional shape satisfying the conditions of the specific building according to the layout data using a pre-learned conditional generation model; and optimizing the plurality of first voxel models and the plurality of second voxel models based on a plurality of predetermined optimization items for the specific building to generate a building model of a three-dimensional shape corresponding to the layout data. Effects of the invention

[0009] According to various embodiments of the present invention, a method and system for generating a 3D building model based on the integration of a voxel GAN ​​and a conditional GAN ​​can effectively improve the structural accuracy and quality of a 3D building model by generating a plurality of voxel models through different methods based on a layout for a specific building and generating a 3D building model for the specific building through optimization of the generated plurality of voxel models. Brief explanation of the drawing

[0010] FIG. 1 illustrates an embodiment of generating a building model according to the present invention. FIG. 2 illustrates a building model generation system according to the present invention. FIG. 3 is a flowchart illustrating a method for generating a building model according to the present invention. FIG. 4 illustrates an example of generating a first voxel model. FIG. 5 illustrates an example of training a voxel generation model. FIG. 6 illustrates an example of generating a second voxel model. FIG. 7 illustrates an example of training a conditional voxel model. Specific details for implementing the invention

[0011] Hereinafter, embodiments disclosed in this specification will be described in detail with reference to the attached drawings. Identical or similar components are assigned the same reference number regardless of the drawing symbols, and redundant descriptions thereof will be omitted. The suffixes "module" and "part" used for components in the following description are assigned or used interchangeably solely for the ease of drafting the specification and do not have distinct meanings or roles in themselves. Furthermore, in describing the embodiments disclosed in this specification, if it is determined that a detailed description of related prior art could obscure the essence of the embodiments disclosed in this specification, such detailed description will be omitted. Additionally, the attached drawings are intended only to facilitate understanding of the embodiments disclosed in this specification; the technical concept disclosed in this specification is not limited by the attached drawings, and it should be understood that they include all modifications, equivalents, and substitutions that fall within the spirit and technical scope of the present invention.

[0012] Terms including ordinal numbers, such as first, second, etc., may be used to describe various components, but said components are not limited by said terms. These terms are used solely for the purpose of distinguishing one component from another.

[0013] When it is stated that one component is "connected" or "connected" to another component, it should be understood that while it may be directly connected or connected to that other component, there may also be other components in between. On the other hand, when it is stated that one component is "directly connected" or "directly connected" to another component, it should be understood that there are no other components in between.

[0014] Singular expressions include plural expressions unless the context clearly indicates otherwise.

[0015] In this application, terms such as “comprising” or “having” are intended to specify the existence of the features, numbers, steps, actions, components, parts, or combinations thereof described in the specification, and should be understood as not precluding the existence or addition of one or more other features, numbers, steps, actions, components, parts, or combinations thereof.

[0016] FIG. 1 illustrates an embodiment of generating a building model according to the present invention. FIG. 2 illustrates a building model generation system according to the present invention.

[0017] Referring to FIG. 1, a building model generation system (100) according to the present invention can generate a plurality of first voxel models according to a voxel generation model and a plurality of second voxel models according to a conditional generation model based on two-dimensional layout data for a specific building, and generate a building model of a three-dimensional shape through optimization of the plurality of first voxel models and the plurality of second voxel models.

[0018] Here, layout data may include information regarding one or more spatial structures and architectural components constituting one or more spatial structures for a specific building. For example, layout data may include spatial structure data representing one or more spatial structures included in a specific building and component data representing the shape, location, material, etc., of the architectural components constituting each spatial structure. In this case, the spatial structure data may include information related to spatial structures such as bedrooms, kitchens, living rooms, and bathrooms, and the component data may include information related to architectural components such as walls, roofs, doors, windows, columns, and furniture.

[0019] A voxel generation model is trained to generate a voxel model of a three-dimensional shape when a predetermined noise is input. In one embodiment, the voxel generation model may be implemented as a Generative Adversarial Network (GAN). That is, the voxel generation model may be trained based on a first voxel model generated from random noise (or latent vector) through a generator, and whether the first voxel model is real or fake, as output by a discriminator.

[0020] The first voxel model may represent a three-dimensional space containing one or more three-dimensional spaces of a predetermined size, as a three-dimensional binary tensor or array generated from noise through a voxel generation model. This first voxel model can be understood as simulating the spatial structure of a given building. Accordingly, the first voxel model may be generated to indicate whether each location in the three-dimensional space is occupied by material or is empty.

[0021] A conditional generative model is trained to generate a voxel model of a 3D shape corresponding to the input layout data when layout data is input along with a predetermined noise. In one embodiment, the conditional generative model may be implemented as a cGAN (Conditional Generative Adversarial Networks). That is, the conditional generative model may be trained based on a second voxel model generated from arbitrary noise (or latent vector) to correspond to the layout data through a generator, and whether the second voxel model is real or fake (or whether it is applied to the layout data) output from a discriminator.

[0022] The second voxel model may represent a three-dimensional space containing one or more three-dimensional spaces of a predetermined size, as a three-dimensional binary tensor or array generated from noise and layout data through a conditional generation model. This second voxel model can be understood as simulating the spatial structure of a given building. The second voxel model may be generated to indicate whether each location in the three-dimensional space is occupied by a material or is empty.

[0023] In this regard, the voxel generation model and the conditional generation model may be trained based on various loss functions such as Equation 1 below, depending on the embodiment.

[0024]

[0025] Here, L MSE represents the loss due to mean squared error, and L L1 represents the loss due to the mean absolute error, and L perceptual represents the loss according to Perceptual, and L IoU represents the loss due to Intersection over Union, and L TVcan represent the loss due to Total Variation. Additionally, G can represent the generators for the voxel generation model and the conditional generation model, and y can represent the ground truth voxel model, which is the labeling data for the training layout data prepared to train the voxel generation model and the conditional generation model, respectively. Furthermore, N can represent the number of datasets prepared to train the voxel generation model and the conditional generation model, respectively, and V i,j,k can represent a voxel model generated by a voxel generation model and a conditional generation model, where i, j, and k each represent a value according to the 3-dimensional space of the voxel model.

[0026] Meanwhile, the building model is a voxel model generated according to the optimization of a plurality of first voxel models and a plurality of second voxel models, and can be understood as a combination of at least one of the plurality of different first voxel models and a plurality of second voxel models.

[0027] Performing optimization on a plurality of first voxel models and a plurality of second voxel models can be understood as generating an optimized building model by extracting and combining elements satisfying optimization items from a plurality of first voxel models and a plurality of second voxel models based on a plurality of predetermined optimization items. In one embodiment, a building model generation system (100) can perform optimization based on a plurality of first voxel models and a plurality of second voxel models based on NSGA (Non-Dominated Sorting Genetic Algorithm).

[0028] In this case, the predetermined multiple optimization items may include items considered in building design, for example, structural soundness, space utilization (e.g., spatial arrangement and floor area distribution), material efficiency (e.g., satisfaction of requirements for construction materials), aesthetic appeal, and cost efficiency.

[0029] Accordingly, the building model generation system (100) evaluates multiple voxel models (e.g., multiple first voxel models and multiple second voxel models) generated using a voxel generation model and a conditional generation model according to multiple optimization items, classifies multiple voxels into multiple layers through Pareto optimization based on the evaluation results, and can form a Pareto Front by extracting multiple voxel models based on the density distance between multiple voxel models belonging to each of the multiple layers.

[0030] Additionally, the building model generation system (100) can repeat the process of evaluating multiple voxel models belonging to a Pareto front according to multiple optimization items, classifying multiple voxels into multiple layers through Pareto optimization based on the evaluation results, and extracting multiple voxel models based on the density distance between multiple voxel models belonging to each of the multiple layers to form a Pareto front.

[0031] At this time, the building model generation system (100) can terminate the repetition of the optimization process based on a predetermined number of repetitions, an evaluation result according to an optimization item, or the density distance between multiple voxel models belonging to the Pareto front, and accordingly, each of the multiple voxel models belonging to the Pareto front finally configured can be specified as a building model.

[0032] Meanwhile, referring to FIG. 2, the building model generation system (100) may include an input unit (110), a storage unit (120), a control unit (130), and an output unit (140). These input unit (110), storage unit (120), control unit (130), and output unit (140) may be connected to each other through a system bus, a network, or other connection mechanism.

[0033] The input unit (110) may receive information necessary for the operation of the building model generation system (100) according to the present invention. To this end, the input unit (110) may be connected to a separate input device, server, or external storage device via a wireless or wired network. For example, the input unit (110) may receive certain information based on a touch screen, computer mouse, keyboard, keypad, touchpad, trackball, joystick, voice recognition module, or other similar device.

[0034] Accordingly, the input unit (110) may receive layout data (10) from a separate input device, server, or external storage device, etc. Additionally, depending on the embodiment, the input unit (110) may receive training layout data and a correct answer voxel model so as to train at least one of a voxel generation model (20) and a conditional generation model (30).

[0035] Additionally, the storage unit (120) may store instructions and information necessary for the operation of the building model generation system (100). As a data storage unit (120), this storage unit (120) may include one or more non-transient computer-readable storage media that can be read and / or accessed by a control unit (130) (e.g., at least one of one or more processors).

[0036] One or more non-transient computer-readable storage media may include volatile and / or non-volatile storage components, such as optical, magnetic, organic, or other memory or disk storage devices. In some examples, the storage unit (120) may be implemented using a single physical device (e.g., one optical, magnetic, organic, or other memory or disk storage device), whereas in other examples, the storage unit (120) may be implemented using two or more physical devices.

[0037] Additionally, the storage unit (120) may include computer-readable instructions and additional data. The storage unit (120) may include a storage necessary to perform at least some of the methods, scenarios, and techniques described herein and / or at least some of the functions of the device and network.

[0038] Accordingly, the storage unit (120) may store layout data (10), a voxel generation model (20), and a conditional generation model (30), and additionally, the storage unit (120) may store a plurality of first voxel models, a plurality of second voxel models, and a building model (40). In this way, the storage unit (120) may store various information generated during the process of generating the building model (40).

[0039] A control unit (130) can control the overall operation of a building model generation system (100). This control unit (130) may be one or more processors and may include one or more general-purpose processors and / or one or more special-purpose processors (e.g., digital signal processor, tensor processing unit (TPU), graphics processing unit (GPU), neural network processing unit (NPU), application integrated circuit, application semiconductor (ASIC), etc.). The control unit (130) may be configured to execute computer-readable instructions contained in a storage unit (120) and / or other instructions described herein.

[0040] Accordingly, the control unit (130) can generate a plurality of first voxel models according to a voxel generation model (20) and a plurality of second voxel models according to a conditional generation model (30) based on two-dimensional layout data (10) for a specific building, and generate a three-dimensional building model (30) through optimization of the plurality of first voxel models and the plurality of second voxel models.

[0041] Specifically, the control unit (130) can receive layout data (10) configured in two dimensions for a specific building. That is, the control unit (130) can receive layout data (10) that includes information about one or more spatial structures and architectural components forming one or more spatial structures for a specific building.

[0042] Accordingly, the control unit (130) can generate a plurality of first voxel models of a three-dimensional shape for a predetermined building using a pre-learned voxel generation model (20). To this end, the control unit (130) can generate a plurality of first voxel models corresponding to each of the plurality of noises by inputting different noises into the pre-learned voxel generation model (20) to generate a voxel model of a three-dimensional shape when a predetermined noise is input.

[0043] Additionally, the control unit (130) can generate a plurality of second voxel models of a three-dimensional shape that satisfy the conditions of a specific building according to the layout data (10) using a pre-learned conditional generation model (30). To this end, when the layout data (10) is input along with a predetermined noise, the control unit (130) can generate a plurality of second voxel models corresponding to each of the multiple noises and the layout data (10) by inputting different multiple noises and the layout data (10) into a pre-learned conditional generation model (30) to generate a voxel model of a three-dimensional shape corresponding to the input layout data (10).

[0044] Furthermore, the control unit (130) can optimize a plurality of first voxel models and a plurality of second voxel models based on a plurality of predetermined optimization items for a specific building to generate a three-dimensional building model (40) corresponding to the layout data (10).

[0045] At this time, the control unit (130) can generate a building model (40) corresponding to the layout data (10) by performing optimization for each of the plurality of first voxel models and the plurality of second voxel models based on NSGA.

[0046] The output unit (140) can output information generated by the operation of the building model generation system (100) according to the present invention. To this end, the output unit (140) may be connected to a separate visual output device, server, or external storage device via a wireless or wired network.

[0047] For example, the output unit (140) may be configured to provide output to one or more user display devices, such as cathode ray tubes (CRTs), liquid crystal displays, light-emitting diodes (LEDs), displays using digital light processing (DLP) technology, and printers. Additionally, the output unit (140) may be configured to produce audible output using devices such as speakers, speaker jacks, audio output ports, audio output devices, earphones and / or other similar devices, and one or more tactile devices may be additionally configured to produce tactile outputs, such as vibrations and / or other outputs detectable by touch and / or physical contact.

[0048] Accordingly, the output unit (140) can output layout data (10), a plurality of first voxel models, a plurality of second voxel models, and a building model (40), etc., so that a user can visually check them through a separate output device, server, or external storage device, etc., and according to an embodiment, the layout data (10), a plurality of first voxel models, a plurality of second voxel models, and a building model (40), etc. may also be transmitted to another device.

[0049] Meanwhile, the building model generation system (100) may be a computing device comprising one or more sensors and / or a power system. Here, one or more sensors may be configured to measure conditions within the computing device and / or conditions of the computing device environment and to provide data regarding these conditions. The power system may include one or more batteries and / or one or more external power interfaces for providing power to the computing device.

[0050] Based on the configuration of the building model generation system (100) described above, the method for generating a building model will be explained in more detail below.

[0051] FIG. 3 is a flowchart illustrating a method for generating a building model according to the present invention. FIG. 4 illustrates an embodiment for generating a first voxel model. FIG. 5 illustrates an embodiment for training a voxel generation model. FIG. 6 illustrates an embodiment for generating a second voxel model. FIG. 7 illustrates an embodiment for training a conditional voxel model.

[0052] Referring to FIG. 3, a building model generation system (100) according to the present invention can receive layout data configured in two dimensions for a specific building (S100).

[0053] Specifically, a building model generation system (100) can receive layout data that includes information about one or more spatial structures and architectural components forming one or more spatial structures for a specific building.

[0054] For example, layout data may include spatial structure data representing one or more spatial structures included in a specific building, and component data representing the shape, location, material, etc. of architectural components forming each spatial structure.

[0055] In this case, spatial structure data may include information such as the types of spatial structures included in a specific building and the arrangement of different spatial structures, and here, the spatial structure may represent the structure of spaces such as bedrooms, kitchens, living rooms, and bathrooms.

[0056] Additionally, component data may include information such as the type, shape, location, and material of architectural components forming each spatial structure, wherein architectural components may include walls, roofs, doors, windows, columns, furniture, etc. Accordingly, component data may include information regarding the type (e.g., walls, roofs, doors, windows, columns, furniture, etc.), shape (e.g., dimensions, thickness, angle, style, etc.), and material (e.g., steel, plastic, wood, glass, etc.) of each architectural component, as well as information regarding the location where each architectural component is placed (e.g., coordinates, angle, and direction, etc.).

[0057] The building model generation system (100) according to the present invention can generate a plurality of first voxel models of three-dimensional shapes for a predetermined building using a pre-learned voxel generation model (S200).

[0058] Specifically, the building model generation system (100) can generate multiple first voxel models corresponding to each of the multiple noises by inputting multiple different noises into a voxel generation model that has been pre-trained to generate a voxel model of a three-dimensional shape when a predetermined noise is input.

[0059] Referring to FIG. 4, for example, a building model generation system (100) may input noise (51) (or latent vector) sampled based on a random distribution into a voxel generation model (20). Accordingly, the building model generation system (100) may generate a three-dimensional voxel grid by performing a transpose convolution or upscaling on the noise (51) through the voxel generation model (20), and generate a first voxel model (21) by specifying each of the values ​​of the previously generated voxel grid according to a pre-learned criterion.

[0060] At this time, the building model generation system (100) can generate a plurality of noises (51) based on a random distribution and input each of the plurality of noises (51) into a voxel generation model (20) to generate a plurality of first voxel models (21) corresponding to each of the plurality of noises (51).

[0061] To this end, the voxel generation model (20) may include an input layer, a hidden layer, and an output layer, the input layer is implemented to receive noise (51), the hidden layer is implemented to generate a three-dimensional voxel grid by performing a transpose convolution or upsampling on the noise (51), and the output layer is implemented to generate a first voxel model (21) by specifying each of the values ​​of the voxel grid according to a pre-learned criterion. In particular, the hidden layer may be implemented to generate a voxel grid from the noise (51) based on at least one of multiple transpose convolutions and an activation function (e.g., a Rectified Linear Unit (ReLU) function), and the output layer may be implemented to generate a first voxel model (21) based on a pre-learned criterion in the voxel grid based on a sigmoid function.

[0062] In this regard, referring together with FIG. 5, a building model generation system (100) (or a separate learning system) can train a voxel generation model (20) using a correct voxel model (13), which is labeling data for the learning layout data (11). That is, the voxel generation model (20) may be trained using a correct voxel model (13), which is labeling data for the learning layout data (11), and the correct voxel model (13) may include a predetermined building model implemented in three dimensions corresponding to the learning layout data (11).

[0063] At this time, the voxel generation model (20) may include a generator (23) and a discriminator (25). Accordingly, the building model generation system (100) can generate a first voxel model (27) for training by inputting noise (52) (or latent vector) sampled based on a random distribution into the voxel generation model (20) through the generator (23), and train the generator (23) based on the comparison result by comparing the first voxel model (27) for training with the correct voxel model (13) through the discriminator (25).

[0064] That is, the building model generation system (100) can train the generator (23) to simulate the correct voxel model (13) from arbitrary noise (52) using the discriminator (25), and through this, the voxel generation model (20) can simulate the correct voxel model (13) from the noise (51) to generate the first voxel model (21).

[0065] Here, each of the generator (23) and the discriminator (25) may include, depending on the embodiment, an input layer, a hidden layer, and an output layer, wherein the input layer of the generator (23) is implemented to receive noise (51, 52), the hidden layer of the generator (23) is implemented to generate a three-dimensional voxel grid by performing a transpose convolution or upsampling on the noise (51, 52), and the output layer of the generator (23) may be implemented to generate a first voxel model (21) (or a first voxel model for learning (27)) by specifying each of the values ​​of the voxel grid according to a pre-learned criterion.

[0066] Additionally, the input layer of the discriminator (25) is implemented to receive the first voxel model (27) for training and the correct voxel model (13), the hidden layer of the discriminator (25) is implemented to extract a feature vector by performing downsampling on the first voxel model (27) for training, and the output layer of the discriminator (25) can be implemented to compare the first voxel model (27) for training and the correct voxel model (13) to output a probability of whether the first voxel model (27) for training is a real voxel model or a fake voxel model.

[0067] In one embodiment, the objective function of the generator (23) can be expressed as Equation 2 below, and the objective function of the discriminator (25) can be expressed as Equation 3 below.

[0068]

[0069]

[0070] Here, G represents the generator (23), D represents the discriminator (25), z represents the noise (51, 52) (or latent vector), and x may represent the first voxel model (21) (or the first voxel model for training (27)). Also, p z (z) represents a random distribution (e.g., a Gaussian distribution), and p data(x) can represent the distribution for the voxel model.

[0071] Referring again to FIG. 3, the building model generation system (100) according to the present invention can generate a plurality of second voxel models of three-dimensional shapes satisfying the conditions of a specific building according to layout data using a pre-learned conditional generation model (S300).

[0072] Specifically, when layout data is input along with a predetermined noise, the building model generation system (100) can generate multiple second voxel models corresponding to each of the multiple noises and layout data by inputting different multiple noises and layout data into a conditional generation model that has been pre-trained to generate a voxel model of a three-dimensional shape corresponding to the input layout data.

[0073] For example, with reference to FIG. 6, a building model generation system (100) may input layout data (10) along with noise (53) (or latent vector) sampled based on a random distribution into a conditional generation model (30). Accordingly, the building model generation system (100) may generate a three-dimensional voxel grid by performing a transpose convolution or upscaling on the noise (53) based on the layout data (10) through the conditional generation model (30), and generate a second voxel model (32) by specifying each of the values ​​of the previously generated voxel grid according to a pre-learned criterion.

[0074] At this time, the building model generation system (100) can generate a plurality of noises (53) based on a random distribution, and input layout data (10) along with each of the plurality of noises (53) into a conditional generation model (30) to generate a plurality of second voxel models (32) corresponding to each of the plurality of noises (53).

[0075] To this end, the conditional generation model (30) may include an input layer, a hidden layer, and an output layer, wherein the input layer is implemented to receive noise (53) and layout data (10), the hidden layer is implemented to generate a 3-dimensional voxel grid by performing a transpose convolution or upsampling on the noise (53) based on the layout data (10), and the output layer is implemented to generate a second voxel model (32) by specifying each of the values ​​of the voxel grid according to a pre-learned criterion. In particular, the hidden layer may be implemented to generate a voxel grid from the noise (53) based on at least one of multiple transpose convolutions and an activation function (e.g., a Rectified Linear Unit (ReLU) function), and the output layer may be implemented to generate a second voxel model (32) based on a pre-learned criterion in the voxel grid based on a sigmoid function.

[0076] In this regard, referring together with FIG. 7, a building model generation system (100) (or a separate learning system) can train a conditional generation model (30) using learning layout data (11) (and / or a correct voxel model). That is, the conditional generation model (30) may be trained using learning layout data (11) (and / or a correct voxel model).

[0077] At this time, the conditional generation model (30) may include a generator (33) and a discriminator (35). Accordingly, the building model generation system (100) may generate a second voxel model (37) for learning by inputting noise (54) (or latent vector) sampled based on a random distribution and learning layout data (11) into the conditional generation model (30) through the generator (33), determine whether information regarding the spatial structure and architectural components according to the learning layout data (11) has been applied to the second voxel model (37) for learning through the discriminator (35), and train the generator (33) based on the determination result. Here, according to an embodiment, the building model generation system (100) may compare the second voxel model (37) for learning and the correct voxel model, which is labeling data for the learning layout data (11), through the discriminator (35), and train the generator (33) based on the comparison result.

[0078] That is, the building model generation system (100) can train the generator (33) to generate a second voxel model with layout data (10) applied from arbitrary noise (53) using a discriminator (35), and through this, the conditional generation model (30) can generate a second voxel model (32) from the noise (53) and layout data (10).

[0079] Here, each of the generator (33) and the discriminator (35) may include, depending on the embodiment, an input layer, a hidden layer, and an output layer, wherein the input layer of the generator (33) is implemented to receive noise (53, 54) and layout data (10) (or learning layout data (11)), the hidden layer of the generator (33) is implemented to generate a three-dimensional voxel grid by performing a transpose convolution or upsampling on the noise (53, 54) based on the layout data (10) (or learning layout data (11)), and the output layer of the generator (33) may be implemented to generate a second voxel model (32) (or a second learning voxel model (37)) by specifying each of the values ​​of the voxel grid according to a pre-learned criterion.

[0080] Additionally, the input layer of the discriminator (35) is implemented to receive a second voxel model (37) for training and a layout data (11) (and / or a correct answer voxel model), the hidden layer of the discriminator (35) is implemented to extract a feature vector by performing downsampling on the second voxel model (37) for training, and the output layer of the discriminator (35) can be implemented to determine whether the second voxel model (37) for training contains information corresponding to the layout data (11) based on the feature vector extracted from the hidden layer.

[0081] Referring again to FIG. 3, the building model generation system (100) according to the present invention can generate a building model of a three-dimensional shape corresponding to layout data by optimizing a plurality of first voxel models and a plurality of second voxel models based on a plurality of optimization items predetermined for a specific building (S400).

[0082] Specifically, the building model generation system (100) can generate a building model corresponding to layout data by performing optimization for each of a plurality of first voxel models and a plurality of second voxel models based on NSGA.

[0083] For example, a building model generation system (100) can specify a score corresponding to each of a plurality of optimization items for each of a plurality of first voxel models and a plurality of second voxel models, and extract a plurality of voxel models corresponding to Pareto optimization based on the specified score.

[0084] At this time, the building model generation system (100) can classify the plurality of first voxel models and the plurality of second voxel models into a plurality of layers by repeatedly performing Pareto optimization on the plurality of first voxel models and the plurality of second voxel models.

[0085] That is, the building model generation system (100) can extract one or more voxel models corresponding to the first layer through Pareto optimization of a plurality of first voxel models and a plurality of second voxel models.

[0086] Next, the building model generation system (100) can perform Pareto optimization on a plurality of other voxel models, excluding one or more voxel models extracted from the first layer, among a plurality of first voxel models and a plurality of second voxel models. That is, the building model generation system (100) can extract one or more voxel models corresponding to the second layer through Pareto optimization on a plurality of other voxel models excluding the first layer.

[0087] Additionally, the building model generation system (100) can perform Pareto optimization on a plurality of other voxel models among a plurality of first voxel models and a plurality of second voxel models, excluding the plurality of voxel models extracted from the first layer and the second layer. That is, the building model generation system (100) can extract one or more voxel models corresponding to the third layer through Pareto optimization on the plurality of other voxel models excluding the first layer and the second layer.

[0088] In this way, the building model generation system (100) can repeatedly perform Pareto optimization until each of the plurality of first voxel models and the plurality of second voxel models is classified into a plurality of layers.

[0089] Accordingly, the building model generation system (100) can extract multiple voxel models corresponding to a predetermined maximum number based on the density distance between multiple voxel models classified into multiple layers. At this time, the building model generation system (100) can extract multiple voxel models step-by-step from each of the multiple layers.

[0090] In one embodiment, the building model generation system (100) can calculate the density distance according to the following mathematical formula 4.

[0091]

[0092] Here, CD i can represent the density distance for a specific voxel model i, M can represent the number of multiple optimization items, and m can represent any one of the multiple optimization items. Also, d i,m can represent the distance between a specific voxel model i and an adjacent voxel model for any one of multiple optimization items, and R m can represent the difference between the minimum and maximum values ​​of any one of the optimization items as a range of any one of the multiple optimization items, among the multiple first voxel models and the multiple second voxel models (or multiple voxel models belonging to the Pareto front).

[0093] Accordingly, the building model generation system (100) can extract one or more voxel models with a relatively long distance between multiple voxel models belonging to a first layer, and then, if the number of voxel models extracted from the first layer is less than a predetermined maximum number, extract one or more voxel models with a relatively long distance between multiple voxel models belonging to a second layer. In this way, the building model generation system (100) can sequentially extract voxel models from each of the multiple layers until the number of voxel models extracted from the multiple layers reaches a predetermined maximum number.

[0094] At this time, the building model generation system (100) may extract one or more voxel models in each of the multiple layers whose density distance is farther than a predetermined standard distance, or extract a predetermined number of voxel models in each layer according to the ratio of the number of voxel models belonging to each of the multiple layers.

[0095] That is, the building model generation system (100) can sequentially extract one or more voxel models whose density distance in a specific layer is farther than a predetermined reference distance, and if there are no voxel models whose density distance in a specific layer is farther than a predetermined reference distance, it can extract a voxel model in the next layer according to the reference distance. At this time, the building model generation system (100) can extract one or more voxel models among multiple voxel models belonging to a specific layer that have the highest score corresponding to at least one of multiple optimization items, and extract other voxel models according to the density distance with the extracted one or more voxel models. In addition, the building model generation system (100) may vary the reference distance in each layer according to the number of voxel models belonging to each layer, and in this case, as the number of voxel models belonging to each layer increases, the reference distance may become closer, and as the number of voxel models belonging to each layer decreases, the reference distance may become farther.

[0096] Alternatively, the building model generation system (100) may determine the number of voxel models extracted from each layer based on the ratio of the number of voxel models belonging to each of the multiple layers. For example, the building model generation system (100) may extract more voxel models from a specific layer as the number of voxel models belonging to a specific layer increases, and extract fewer voxel models from a specific layer as the number of voxel models belonging to a specific layer decreases.

[0097] In this way, the building model generation system (100) can extract multiple voxel models corresponding to a predetermined maximum number from multiple layers, and the multiple voxel models extracted accordingly can be understood as corresponding to Pareto fronts.

[0098] Accordingly, the building model generation system (100) can generate multiple different voxel models by crossing multiple voxel models belonging to the Pareto front. To this end, the building model generation system (100) can generate multiple different voxel models by randomly decomposing and combining multiple voxel models, or by decomposing and combining them according to data corresponding to optimization items.

[0099] Next, the building model generation system (100) may repeat the previously described Pareto optimization process, Pareto front extraction process, and cross-generation process for multiple voxel models belonging to the Pareto front for multiple voxel models generated based on the Pareto front. At this time, the building model generation system (100) may repeat the Pareto optimization process, Pareto front extraction process, and cross-generation process for multiple voxel models belonging to the Pareto front a predetermined number of times, or repeat the Pareto optimization process, Pareto front extraction process, and cross-generation process for multiple voxel models belonging to the Pareto front until the average density distance between multiple voxel models belonging to the Pareto front or the total sum of scores for multiple optimization items satisfies a predetermined criterion.

[0100] Through this, the building model generation system (100) can identify multiple voxel models that are finally extracted as Pareto fronts as building models. That is, the building model generation system (100) can identify each of the multiple voxel models extracted as Pareto fronts as building models at the point when the repetition of the Pareto optimization process, the Pareto front extraction process, and the cross-generation process for multiple voxel models belonging to the Pareto fronts ends.

[0101] Through the above configurations, the building model generation system (100) according to the present invention generates a plurality of voxel models through different methods based on a layout for a specific building, and generates a three-dimensional building model for the specific building through optimization of the generated plurality of voxel models, thereby effectively improving the structural accuracy and quality of the three-dimensional building model.

[0102] Furthermore, the present invention described above can be implemented as a program that is executed by one or more processes in an electronic device and stored in a computer-readable recording medium.

[0103] Accordingly, the present invention can be implemented as computer-readable code or instructions on a medium on which a program is recorded. That is, various control methods according to the present invention can be provided in the form of integrated or individual programs.

[0104] Meanwhile, computer-readable media include all types of recording devices in which data that can be read by a computer system is stored. Examples of computer-readable media include HDD (Hard Disk Drive), SSD (Solid State Disk), SSD (Silicon Disk Drive), ROM, RAM, CD-ROM, magnetic tape, floppy disk, optical data storage device, etc.

[0105] Furthermore, the computer-readable medium may be a server or cloud storage that includes a storage and is accessible to an electronic device via communication. In this case, the computer may download the program according to the present invention from the server or cloud storage via wired or wireless communication.

[0106] Furthermore, in the present invention, the computer described above is an electronic device equipped with a processor, namely a CPU (Central Processing Unit), and no special limitations are placed on its type.

[0107] Meanwhile, the above detailed description should not be interpreted restrictively in all respects but should be considered exemplary. The scope of the invention shall be determined by a reasonable interpretation of the appended claims, and all modifications within the equivalent scope of the invention are included within the scope of the invention.

Claims

Claim 1 A method for generating a building model, comprising: receiving layout data configured in two dimensions for a specific building; generating a plurality of first voxel models of a three-dimensional shape for a specific building using a pre-trained voxel generation model; generating a plurality of second voxel models of a three-dimensional shape satisfying the conditions of the specific building according to the layout data using a pre-trained conditional generation model; and optimizing the plurality of first voxel models and the plurality of second voxel models based on a plurality of predetermined optimization items for the specific building to generate a building model of a three-dimensional shape corresponding to the layout data. Claim 2 A building model generation method according to claim 1, wherein the step of generating the building model comprises generating the building model corresponding to the layout data by performing optimization for each of the plurality of first voxel models and the plurality of second voxel models based on NSGA (Non-Dominated Sorting Genetic Algorithm). Claim 3 A method for generating a building model according to claim 2, wherein the step of generating the building model comprises: a step of evaluating the plurality of first voxel models and the plurality of second voxel models according to a plurality of predetermined optimization items; a step of classifying the plurality of first voxel models and the plurality of second voxel models into a plurality of layers through Pareto optimization based on the evaluation results; and a step of extracting the plurality of voxel models based on the density distance between the plurality of voxel models belonging to each of the plurality of layers to form a Pareto Front. Claim 4 A building model generation method according to claim 1, wherein the step of generating the plurality of first voxel models comprises inputting different plurality of noises into the voxel generation model, which is pre-trained to generate a voxel model of a three-dimensional shape when a predetermined noise is input, thereby generating the plurality of first voxel models corresponding to each of the plurality of noises. Claim 5 A building model generation method according to claim 4, wherein the voxel generation model is trained using a correct voxel model, which is labeling data for pre-prepared training layout data, and includes a generator and a discriminator, wherein the generator is trained to mimic the correct voxel model from arbitrary noise using the discriminator. Claim 6 A building model generation method according to claim 1, wherein the step of generating the plurality of second voxel models comprises, when the layout data is input along with a predetermined noise, inputting different plurality of noises and the layout data into the conditional generation model that is pre-trained to generate a voxel model of a three-dimensional shape corresponding to the layout data, thereby generating the plurality of second voxel models corresponding to each of the plurality of noises and the layout data. Claim 7 A building model generation method according to claim 6, wherein the conditional generation model is trained using pre-prepared training layout data, includes a generator and a discriminator, and the generator is trained to generate a second voxel model to which the layout data is applied from arbitrary noise using the discriminator. Claim 8 A method for generating a building model according to claim 1, wherein the layout data comprises: spatial structure data representing one or more spatial structures included in the specific building; and component data representing at least one of the shape, location, and material of the architectural components forming the one or more spatial structures. Claim 9 A building model generation system comprising: a storage unit for storing layout data configured in two dimensions for a specific building; and a control unit for generating a building model of a three-dimensional shape based on the layout data, wherein the control unit generates a plurality of first voxel models of a three-dimensional shape for a predetermined building using a pre-learned voxel generation model, generates a plurality of second voxel models of a three-dimensional shape satisfying the conditions of the specific building according to the layout data using a pre-learned conditional generation model, and optimizes the plurality of first voxel models and the plurality of second voxel models based on a plurality of predetermined optimization items for the specific building to generate the building model of a three-dimensional shape corresponding to the layout data. Claim 10 A program stored on a computer-readable recording medium that is executed by one or more processes in an electronic device, the program comprising instructions for performing the steps of: receiving layout data configured in two dimensions for a specific building; generating a plurality of first voxel models of a three-dimensional shape for a specific building using a pre-learned voxel generation model; generating a plurality of second voxel models of a three-dimensional shape satisfying the conditions of the specific building according to the layout data using a pre-learned conditional generation model; and optimizing the plurality of first voxel models and the plurality of second voxel models based on a plurality of predetermined optimization items for the specific building to generate a building model of a three-dimensional shape corresponding to the layout data.

Images