An end-to-end house type drawing design method, system, computer device and storage medium
By employing an end-to-end floor plan design method, which utilizes a large language model and graph convolutional networks to generate pixel-level details in floor plans, this approach solves the problems of user customization and multimodal transfer in existing technologies, and achieves efficient generation of floor plans and furniture arrangement.
Patent Information
- Application Number
- CN202411820339.6
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-12-11
- Publication Date
- 2026-06-12
AI Technical Summary
Existing floor plan design models cannot meet user customization needs and multimodal migration, and cannot generate elements connecting rooms or describe the spatial relationships between them.
An end-to-end floor plan design method is adopted, which uses a large language model and graph convolutional network to generate the geometric relationships of each element in the floor plan. The room division and furniture arrangement are realized through encoder and decoder, generating pixel-level details.
It enables users to design customized floor plans, generates logically consistent partial detail diagrams, provides multimodal transfer capabilities, and supports image and text retrieval and pixelated generation.
Smart Images
Figure CN122197116A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of floor plan design technology, specifically an end-to-end floor plan design method, system, computer equipment, and storage medium. Background Technology
[0002] Existing floor plan design model schemes can be divided into two categories: (1) Based on bubblediagram, the attribute description of room nodes is completed, and the corresponding room bounding box is generated; (2) Most of the schemes that directly generate to the pixel level adopt Conditional GAN, where Conditional only has the limitation of floor plan, without the limitation of the spatial position of indoor room nodes and their attribute description.
[0003] Bubblediagram-based floor plan design models cannot generate elements connecting rooms, such as pixel-level details like doors, windows, and walls. Conditional GANs based on simple original floor plans cannot describe the spatial relationships between rooms and their related attributes. Therefore, the two existing floor plan design models cannot meet users' customization needs and multimodal transfer requirements. Summary of the Invention
[0004] Purpose of the invention: To address the problem that existing floor plan design models cannot meet user customization needs and multimodal migration, and to address the problem that existing floor plan design models cannot simultaneously generate elements connecting rooms, describe the spatial relationships between rooms and their related attributes, this invention proposes an end-to-end floor plan design method, system, computer equipment, and storage medium.
[0005] Technical solution: An end-to-end floor plan design method, comprising:
[0006] The geometric relationships of each element in the floor plan and the image features of the floor plan are input into the room division model, and the floor plan with the rooms divided is output.
[0007] Furthermore, the geometric relationships of the elements in the floor plan are obtained according to the following steps:
[0008] Using a large language model and prompt words, generate a bubble diagram or relationship abstraction diagram that describes the geometric relationships of each element in the floor plan;
[0009] Each element abstracts the attributes of the element, including size, shape, and position; the geometric relationships include the relative spatial positional relationships and connection relationships between the elements;
[0010] The prompts describe the geometric relationships between the elements in the floor plan.
[0011] Furthermore, the room division includes an encoder and a decoder;
[0012] In the encoder, the geometric relationships between the elements in the floor plan and the image features of the floor plan are learned through a graph convolutional network to generate a Canvas Latent.
[0013] The decoder converts the Canvas Latent into pixelated room partitions.
[0014] Furthermore, the decoder employs a generator from a GAN.
[0015] Furthermore, the graph convolutional network includes: a first GS network, a first GP network, a first GO network, a second GS network, a second GP network, a second GO network, a first H network, a second H network, and a third H network; nodes v1, v2, and edge vr1 are respectively input into the first GS network, the first GP network, and the first GO network; nodes v2, v3, and edge vr2 are respectively input into the second GS network, the second GP network, and the second GO network; the output of the first GS network is input into the first H network, outputting v1'; the first GP network outputs vr1'; the outputs of the first GO network and the second and second GS networks are both input into the second H network, outputting v2'; the second GP network outputs vr2'; the output of the second GO network is input into the third H network, outputting v3';
[0016] Here, edge vr1 represents the edge connecting nodes v1 and v2, and edge vr2 represents the edge connecting nodes v2 and v3; each GS network, each GP network, and each GO network are networks with the same structure, defined as g networks, and each h network has the same structure. v1' represents the feature corresponding to node v1, v2' represents the feature corresponding to node v2, v3' represents the feature corresponding to node v3, and vr1' and vr2' represent the features corresponding to edges vr1 and vr2, respectively.
[0017] For each h network, its input passes through the first hidden layer, symmetric pooling, the second hidden layer, the ReLU activation function, and the first output layer in sequence to obtain the output;
[0018] For each g network, there are three inputs. Each input passes through its own Din layer and then enters the 3Din layer. The output of the 3Din layer is input to the third hidden layer. After the third hidden layer, the ReLU activation function is applied, and the output is sent to the corresponding hidden layer or Dout.
[0019] Furthermore, the decoder includes a decoding input layer, a first average pooling layer, a first multi-layer convolution, a second average pooling layer, a second multi-layer convolution, a third average pooling layer, a third multi-layer convolution, and a decoding output layer;
[0020] The input feature information to the decoder is divided into three paths. The first path is as follows: the input feature information to the decoder passes through the first average pooling layer, random noise is added, and then it is input into the first multi-layer convolution. Random noise is added again, and then it enters the second average pooling layer. The second path is as follows: the input feature information to the decoder is input into the second average pooling layer, the output of the second average pooling layer is input into the second multi-layer convolution, random noise is added, and then it enters the third average pooling layer. The third path is as follows: the input feature information to the decoder is input into the third average pooling layer, and the output of the third average pooling layer is input into the third multi-layer convolution to obtain the decoder output.
[0021] This invention also discloses an end-to-end floor plan design method, comprising:
[0022] The geometric relationships of the elements to be arranged in the room diagram and the image features of the room diagram are input into the furniture arrangement model, and the bounding box coordinates of each piece of furniture in the room diagram are output.
[0023] Furthermore, the geometric relationships of the elements to be arranged in the room diagram are obtained according to the following steps:
[0024] Using a large language model and prompt words, generate a bubble diagram or abstract diagram of the geometric relationships of the elements to be arranged in the room diagram;
[0025] Each element abstracts the attributes of the element, including size, shape, and position; the geometric relationships include the relative spatial positional relationships and connection relationships between the elements;
[0026] The prompts describe the geometric relationships between the elements to be arranged in the room diagram.
[0027] Furthermore, the furniture arrangement model includes: an encoder and a decoder;
[0028] In the encoder, the geometric relationships of each element to be placed in the room diagram and the image features of the room diagram are learned through a graph convolutional network to learn the high-dimensional abstract relationships between each element, generating the bounding box latent space.
[0029] The decoder decodes the bounding box latent space into the actual bounding box coordinates.
[0030] Furthermore, the graph convolutional network includes: a first GS network, a first GP network, a first GO network, a second GS network, a second GP network, a second GO network, a first H network, a second H network, and a third H network; nodes v1, v2, and edge vr1 are respectively input into the first GS network, the first GP network, and the first GO network; nodes v2, v3, and edge vr2 are respectively input into the second GS network, the second GP network, and the second GO network; the output of the first GS network is input into the first H network, outputting v1'; the first GP network outputs vr1'; the outputs of the first GO network and the second and second GS networks are both input into the second H network, outputting v2'; the second GP network outputs vr2'; the output of the second GO network is input into the third H network, outputting v3';
[0031] Here, edge vr1 represents the edge connecting nodes v1 and v2, and edge vr2 represents the edge connecting nodes v2 and v3; each GS network, each GP network, and each GO network are networks with the same structure, defined as g networks, and each h network has the same structure. v1' represents the feature corresponding to node v1, v2' represents the feature corresponding to node v2, v3' represents the feature corresponding to node v3, and vr1' and vr2' represent the features corresponding to edges vr1 and vr2, respectively.
[0032] For each h network, its input passes through the first hidden layer, symmetric pooling, the second hidden layer, the ReLU activation function, and the first output layer in sequence to obtain the output;
[0033] For each g network, there are three inputs. Each input passes through its own Din layer and then enters the 3Din layer. The output of the 3Din layer is input to the third hidden layer. After the third hidden layer, the ReLU activation function is applied, and the output is sent to the corresponding hidden layer or Dout.
[0034] Furthermore, the decoder uses graph convolutional networks and multilayer perceptrons to decode the bounding box latent space into bounding box coordinates.
[0035] The present invention also discloses an end-to-end floor plan design system for performing the steps of the aforementioned end-to-end floor plan design method.
[0036] The present invention also discloses a computer device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the steps of the end-to-end floor plan design method disclosed above.
[0037] The present invention also discloses a storage medium storing a floor plan design program, which, when executed by at least one processor, implements the steps of the end-to-end floor plan design method disclosed above.
[0038] Beneficial effects: Compared with the prior art, the present invention has the following advantages:
[0039] (1) The present invention adopts a unified and flexible model framework, which can be combined with a large language model to provide users with customized room descriptions to generate floor plan design results, or generate local detail diagrams that conform to the floor plan design logic according to the characteristics of the data.
[0040] (2) The generation method of Canvas Latent in the model framework of this invention can be used as the basis for downstream applications; specifically, the Latent vector output by the encoder in the model framework integrates the spatial geometric relationship between elements and their respective attributes as well as the outline information of the overall house type, and uses the decoder to generate the downstream room division and furniture division results; the encoder of this invention can also be used as a feature vector extractor for image and text retrieval. For example, if the user inputs the house type and the desired layout information description, the encoder will extract the feature vector and quickly extract the corresponding results in the vector database. At the same time, by utilizing the high-dimensional clustering characteristics of vectors, similar layout information can be extracted better, which is convenient for designers to search and check for duplicates.
[0041] (3) The present invention provides a higher (non-low-level pixel level) abstract interpretation of the direct operation of bubblediagram, which is closer to human description and provides a foundation for future multimodal operations, including generating bubblediagram based on text or language.
[0042] (4) The present invention achieves pixel-level generation of results, and can learn more pixel-level details from the data using the model itself, such as doors, walls and windows. Attached Figure Description
[0043] Figure 1 This is a schematic diagram of the framework of an end-to-end floor plan design method for realizing image pixel-level room division proposed in Example 1;
[0044] Figure 2 This is a schematic diagram of a decoder for Canvas Latent;
[0045] Figure 3 This is a schematic diagram of the structure of a graph convolutional network;
[0046] Figure 4 This is a schematic diagram of the decoder structure;
[0047] Figure 5This is a schematic diagram of the framework of an end-to-end floor plan design method for realizing home layout proposed in Example 2;
[0048] Figure 6 This is the original room layout diagram for the furniture arrangement in Example 3;
[0049] Figure 7 This is the furniture arrangement result diagram automatically generated based on the user's description in Example 3;
[0050] Figure 8 This is the original room layout diagram for the furniture arrangement in Example 4;
[0051] Figure 9 This is the furniture arrangement result diagram automatically generated based on the user's description and the original room layout diagram in Example 4; Figure 10 A schematic diagram showing that each dimension of the Latent vector generated for the encoder contains decomposable feature information;
[0052] Figure 11 This is a schematic diagram of the master bedroom furniture scheme generated by Prompt1 in Example 4;
[0053] Figure 12 This is a schematic diagram of the master bedroom furniture scheme generated by Prompt2 in Example 4;
[0054] Figure 13 This is a schematic diagram of the master bedroom furniture scheme generated by Prompt3 in Example 4;
[0055] Figure 14 This is a schematic diagram of the master bedroom furniture scheme generated by Prompt4 in Example 4;
[0056] Figure 15 This is a schematic diagram of the master bedroom furniture scheme generated by Prompt5 in Example 4. Detailed Implementation
[0057] To facilitate understanding of the technical solution of this invention, the relevant technical terms are now explained.
[0058] Bubble diagram, or bubble chart, can be used to illustrate the relationship between multiple variables.
[0059] The bounding box represents the bounding box.
[0060] GAN, or Generative Adversarial Network, is a deep learning model consisting of two parts: a generator and a discriminator. The generator aims to generate data that is as realistic as possible, while the discriminator aims to distinguish real data from fake data generated by the generator. The generator and discriminator compete with each other during training to improve their respective performance.
[0061] Conditional GAN, or Conditional Generative Adversarial Networks (CGAN), is a model developed based on Generative Adversarial Networks (GANs) that controls the process of generating data by introducing additional conditional information.
[0062] Canvas Latent represents the canvas latent space, which is the space where the high-dimensional vectors corresponding to the canvas exist; Latent represents the space where high-dimensional vectors exist.
[0063] Bounding Box Latent represents the bounding box latent space, the space in which the high-dimensional vectors corresponding to the bounding box exist.
[0064] Graph Convolutional Network (GCN) is a type of neural network specifically designed for processing graph-structured data.
[0065] 2dpixel represents a two-dimensional pixel, referring to a pixel in two-dimensional space.
[0066] To make the objectives, technical solutions, and advantages of this invention clearer, the following description, in conjunction with the accompanying drawings, will further illustrate the end-to-end floor plan design method, system, computer equipment, and storage medium proposed in this invention.
[0067] Example 1:
[0068] like Figure 1 As shown in the figure, this embodiment discloses an end-to-end floor plan design method for realizing room division at the image pixel level, which mainly includes the following steps:
[0069] Step 1: Use a bubblediagram or abstract relationship graph to represent the geometric relationships of each element in the original floor plan to be divided. In this embodiment, the element is a room node. The nodes in the bubblediagram or abstract relationship graph abstract the attributes of the room nodes, including but not limited to: room size, shape, and location. The edges in the bubblediagram or abstract relationship graph abstract the relationships between nodes, including but not limited to: relative spatial position relationships and connection relationships.
[0070] In this step, the bubblediagram can be generated directly from the user's interaction of marking points on the floor plan, or it can be automatically generated using a large language model based on the user's prompt description.
[0071] Step 2: For each specific floor plan, abstract the image features as physical vectors, and use them as conditions in subsequent modeling.
[0072] Step 3: Input the physical vectors and the bubble diagram or relational abstract graph into the encoder. In the encoder, the physical vectors and nodes in the bubble diagram or relational abstract graph are used to learn high-dimensional abstract relationships between nodes through a Graph Convolutional Network (GCN), generating the CanvasLatent. In the CanvasLatent, the Latent vector of each node is transformed into a canvas and associated with 2dpixels through differentiable interpolation. This abstraction facilitates the subsequent decoder to directly generate the pixel-level room shape, size, and position based on this. The CanvasLatent abstraction generated by the encoder contains the relative position, size, and shape of each room, as well as pixel-level details of the connections between rooms (doors, walls, windows). In this step, the spatial relationships and attributes of the rooms are abstracted using the bubble diagram, and the Graph Convolutional Network (GCN) is used to extract node and edge information, thus allowing for the reasonable modeling of arbitrary layout geometric relationships.
[0073] For users, given the same empty floor plan, they can freely describe the spatial relationships between various elements to generate satisfactory results from the model. This descriptive approach allows users to directly mark points on the floor plan or describe it using a rule-based language, enabling multimodal migration. This solves the problem that Conditional GANs based on simple original floor plans cannot describe the spatial relationships between rooms and their related attributes, thus making it difficult to meet user customization needs and achieve multimodal migration.
[0074] In this embodiment, the information of spatial nodes / houses is fused into a CanvasLatent using GCN. Each node can incorporate the size and shape information of the room, providing a basis for diverse generation cases.
[0075] Figure 3 The structure of the graph convolutional network used in this embodiment is shown. Figure 3In the diagram, v1, v2, and v3 each represent a node, and vr1 and vr2 represent edges connecting the nodes. Specifically, vr1 represents the edge connecting nodes v1 and v2, and vr2 represents the edge connecting nodes v2 and v3. gs, gp, and go are all networks with the same structure. Figure 3 In the network, network g is defined as the network, and network h is defined as another network. v1' represents the feature corresponding to node v1, v2' represents the feature corresponding to node v2, v3' represents the feature corresponding to node v3, and vr1' and vr2' represent the features corresponding to edges vr1 and vr2. vs (vertex of subject) and vo (vertex of object) specify the meaning of each node. A directed edge vr is formed from vs to vo. In the g and h networks, Din (input dimensions of vector) represents the input, Dout (output dimensions of vector) represents the output, and Hidden (hidden dimension of vector) represents the hidden layer, abbreviated as H. ReLU represents the activation function. Set Union represents merging, and Mean represents taking the mean. Set Union and Mean form a symmetric pooling.
[0076] Now combined Figure 3 The structure of the graph convolutional network used in this embodiment is further described. This graph convolutional network includes: a first GS network, a first GP network, a first GO network, a second GS network, a second GP network, a second GO network, a first H network, a second H network, and a third H network. Nodes v1, v2, and edge vr1 are input into the first GS network, the first GP network, and the first GO network, respectively. Nodes v2, v3, and edge vr2 are input into the second GS network, the second GP network, and the second GO network, respectively. The output of the first GS network is input into the first H network, outputting v1'. The first GP network outputs vr1'. The outputs of the first GO network and the second GS network are both input into the second H network, outputting v2'. The second GP network outputs vr2'. The output of the second GO network is input into the third H network, outputting v3'.
[0077] For each h network, its input passes through the first hidden layer, symmetric pooling, the second hidden layer, the ReLU activation function, and the h network output layer in sequence to obtain the output.
[0078] For each g network, there are three inputs. Each input enters its own Din layer, then enters the 3Din layer. The output of the 3Din layer is input to the third hidden layer. After the third hidden layer, the ReLU activation function is applied, and the output is sent to the corresponding hidden layer or Dout.
[0079] Step 4: Each dimension of the Latent vector generated by the encoder contains decomposable feature information. This feature information integrates the spatial geometric relationships between elements, their respective attributes, and the overall floor plan outline. By interpolating or calculating this feature information, various cases can be generated at the decoder terminal, achieving seamless and smooth deformation between generated samples. For example... Figure 10 As shown, the 11th dimension manages the length of the orange boxes, and the 62nd dimension manages the proportion of the blue boxes. Therefore, the decoder uses the latent vector from the encoder output to generate the downstream room partitioning results.
[0080] Figure 4 The diagram shows the structure of the decoder used in this step, which mainly includes: a decoding input layer, a first average pooling layer, multiple convolutional layers, a second average pooling layer, a third average pooling layer, and a decoding output layer. The feature information input to the decoder passes through the first average pooling layer, adds random noise, goes through multiple convolutional layers, adds random noise again, and then passes through the second average pooling layer. Simultaneously, the feature information input to the decoder is fed into the second average pooling layer. The output of the second average pooling layer passes through multiple convolutional layers, adds random noise, and then enters the third average pooling layer. At the same time, the feature information input to the decoder is also fed into the third average pooling layer. The output of the third average pooling layer passes through multiple convolutional layers to obtain the decoder output.
[0081] like Figure 2 As shown, for the Canvas Latent, the decoder (GAN Decoder) uses a random noise vector containing high-frequency information in combination with the generator in GAN to generate relative pixels. At the same time, the discriminator controls the model to generate richer and more realistic details. This solves the problem that the floor plan design model based on bubblediagram cannot generate elements that connect rooms, such as pixel-level details, such as doors, walls and windows.
[0082] Example 2:
[0083] like Figure 5 As shown in the figure, this embodiment discloses an end-to-end floor plan design method for realizing home layout, which mainly includes the following steps:
[0084] Step 1: Represent the geometric relationships of the elements in the room where furniture will be arranged using a bubblediagram or a relational abstract graph. In this embodiment, the room where furniture will be arranged is actually the outline of the room. The element is the furniture node; the nodes in the bubblediagram or relational abstract graph abstract the attributes of the nodes, including but not limited to: the size, shape, and position of the furniture; the edges in the bubblediagram or relational abstract graph abstract the relationships between the nodes, including but not limited to: relative spatial position relationships and connection relationships.
[0085] In this step, the bubble diagram can be automatically generated directly from the user's prompt description.
[0086] Step 2: For each specific room map, abstract the image features as physical vectors, and use them as conditions in subsequent modeling.
[0087] Step 3: Input the physical vectors and the bubble diagram or relational abstract graph into the encoder. In the encoder, the physical vectors and nodes in the bubble diagram or relational abstract graph are used to learn high-dimensional abstract relationships between nodes through a graph convolutional network (GCN), generating a bounding box latent space associated with the bounding box. The bounding box latent space maps the latent vector of each node directly to the bounding box latent space. The decoder then restores the latent vector to the bounding box of the furniture arrangement. The graph convolutional network used in this embodiment is the same as that used in Embodiment 1.
[0088] Step 4: The Latent vector extraction from the encoder output integrates the spatial geometric relationships between elements, their individual attributes, and the overall room contour information. Therefore, the decoder is then used to generate the downstream furniture segmentation results. Specifically, for the Bounding Box Latent, the Graph Convolutional Decoder uses a Graph Convolutional Network (GCN) and a Multilayer Perceptron (MLP) to decode the Bounding Box Latent into the actual bounding box coordinates.
[0089] The room to be furnished, as input in this embodiment, can be obtained by dividing the original floor plan using the method disclosed in Embodiment 1.
[0090] Example 3:
[0091] The end-to-end floor plan design method proposed in Example 1 is used to generate room division layouts based on Prompt. The specific operations include:
[0092] The given original floor plan (including the outer outline and internal load-bearing walls) and prompt are input into the encoder to generate a partitioning scheme. The encoder learns the relative spatial positions between rooms, and the connection relationships between rooms are constrained by the prompt. The constraints of the load-bearing walls and the floor plan are constrained by the pixel images. The generated partitioning scheme satisfies the following: the relative positions and types of rooms match the user's description; the partitioning of each room conforms to the constraints of the internal load-bearing walls; and the details of the connections between rooms, such as non-entrance doors and non-load-bearing walls, are accurate.
[0093] Dynamic, real-time online adjustments are achieved by changing the description input to the encoder.
[0094] Figure 6 The original floor plan is shown. Figure 7 The results show the room division pixel map automatically generated based on the original floor plan and the following Prompt.
[0095] The definition of the relative positions between rooms in Prompt is explained below:
[0096] The non-adjacent relationship between rooms can be represented as: left-above, left-below, right-above, right-below, above, below, left-of, right-of.
[0097] The adjacency relationship between rooms can be represented as:
[0098] left_touching, right_touching, above_touching, below_touching, left_above_touching, left_below_touching, right_above_touching, right_below_touching, left-surrounding, right-surrounding, above-surrounding, below-surrounding, left-above-surrounding, left-below-surrounding, right-above-surrounding, right-below-surrounding, left-surrounded, right-surrounded, above-surrounded, below-surrounded, left-above-surrounded, left-below-surrounded, right-above-surrounded, right-below-surrounded。
[0099] The definition of the relative orientation between a room and a housing type in the Prompt can be expressed as Left_against_wall, Right_against_wall, Above_against_wall, Below_against_wall.
[0100] Based on the above definition, the Prompt adopted in this embodiment is:
[0101] A_P5_19_data = ""
[0102] a secondroom1 is left against the wall, and is above against the wall, and is left touching the kitchen;
[0103] a bathroom1 is left against the wall;
[0104] a bathroom2 is left against the wall;
[0105] a kitchen is above against the wall,and is right touching the secondroom1,and is left above surrounded by the livingdiningroom;
[0106] a masterroom is left against the wall,and is below against the wall,and is left below touching the livingdiningroom;
[0107] a secondroom2 is right against the wall, and is below against the wall, and is right below surrounded by the livingdiningroom;
[0108] a livingdiningroom is above against the wall,and is right against the wall,and is below against the wall,and is below touching the secondroom1.""".
[0109] Example 4:
[0110] This paper presents an end-to-end floor plan design method proposed in Example 2. Based on user-generated requirements, it generates furniture layout diagrams. Given a room outline, it generates diverse furniture arrangement schemes. The furniture in the generated layouts should meet the following requirements: reasonable distribution within the outline area, not exceeding the outline's boundaries; furniture types and relative orientations should conform to the user's description; and furniture size and orientation can be dynamically customized. The encoder learns the probability distribution of high-dimensional spatial vectors for each type of furniture under the premise of a specific room outline and natural language description. These high-dimensional spatial vectors encode the type, model / size, position, and placement orientation of each type of furniture. The decoder samples and decodes based on the learned vector probability distribution, thereby generating diverse furniture layout orientation diagrams that meet the constraints. Users can dynamically adjust the furniture arrangement in real time according to their preferences.
[0111] The specific operations include:
[0112] The description of the relative orientation of furniture to a room can be as follows:
[0113] 'west_touching','east_touching','north_touching','south_touching','1_oclock','2_oclock','3_oclock','4_oclock','5_oclock','6_oclock','7_oclock','8_oclock','9_oclock','10_oclock','11_oclock','12_oclock'.
[0114] The description of the relative positions between pieces of furniture can be as follows:
[0115] 'S_against_wall','N_against_wall','W_against_wall','E_against_wall'.
[0116] The descriptions for defining furniture attributes can be: "Big", "Medium", "Small", "Standard".
[0117] The description for defining the furniture placement direction can be: WE (east-west placement) or NS (north-south placement).
[0118] Figure 8 The original room layout with furniture arrangement is shown. Figure 9 This shows the automatically generated furniture arrangement diagram based on the user's description and the original room layout; the user's description is:
[0119] this a masterroom; a bed is north against the wall;
[0120] a nightstand1 is north against the wall and is west touching the bed;
[0121] a nightstand2 is right east of nightstand1;
[0122] a studydesk is west against the wall, and is 8 o'clock of the bed, and is north south displayed;
[0123] a chair is west touching the studydesk, and is north south displayed;
[0124] a wardrobe is 5 o'clock of the bed, and is south against the wall;
[0125] a singlesofa is east against the wall, and is right east of thewardrobe.
[0126] The following are specific examples of furniture solutions:
[0127] (1) Room type: Master bedroom
[0128] Natural language description prompt 1:
[0129] This is a masterroom;
[0130] a bed is north against the wall.
[0131] The corresponding output is as follows Figure 11 As shown.
[0132] Natural Language Description Prompt 2:
[0133] This is a masterroom;
[0134] a bed is north against the wall;
[0135] a nightstand1 is north against the wall and is west touching the bed;
[0136] a nightstand2 is right east of nightstand1;
[0137] The corresponding output is as follows Figure 12 As shown.
[0138] Natural Language Description Prompt 3:
[0139] This is a master bedroom; a bed is placed against the north wall;
[0140] A nightstand 1 is placed against the north wall and touches the west side of the bed;
[0141] A nightstand 2 is directly east of nightstand 1;
[0142] A study desk is placed against the west wall, at the 8 o'clock position relative to the bed, and is arranged north-south; a chair touches the west side of the study desk and is also arranged north-south;
[0143] The corresponding output is as Figure 13 shown.
[0144] Natural language description Prompt4:
[0145] This is a master bedroom; a bed is placed against the north wall;
[0146] A nightstand 1 is placed against the north wall and touches the west side of the bed;
[0147] A nightstand 2 is directly east of nightstand 1;
[0148] A study desk is placed against the west wall, at the 8 o'clock position relative to the bed, and is arranged north-south; a chair touches the west side of the study desk and is also arranged north-south;
[0149] a wardrobe is 5 o'clock of the bed, and is south against the wall;
[0150] The corresponding output is as follows Figure 14 As shown.
[0151] Natural Language Description (Prompt 5):
[0152] this a masterroom; a bed is north against the wall;
[0153] a nightstand1 is north against the wall and is west touching the bed;
[0154] a nightstand2 is right east of nightstand1;
[0155] a studydesk is west against the wall, and is 8oclock of the bed, and isnorth south displayed; a chair is west touching the studydesk, and is northsouth displayed;
[0156] a wardrobe is 5 o'clock of the bed, and is south against the wall;
[0157] a singlesofa is east against the wall, and is right east of thewardrobe.
[0158] The corresponding output is as follows Figure 15 As shown.
[0159] Example 5:
[0160] This embodiment proposes an end-to-end floor plan design system, which is used to execute the method provided in Embodiment 1 and has the corresponding functions and beneficial effects of the method.
[0161] Example 6:
[0162] This embodiment proposes an end-to-end floor plan design system, which is used to execute the method provided in Embodiment 2 and has the corresponding functions and beneficial effects of the method.
[0163] Example 7
[0164] This embodiment provides a computer device that provides services for implementing the method of Embodiment 1 described above. The computer device may include: a memory storing a computer-executable program, and a processor coupled to the memory. The processor invokes the computer-executable program stored in the memory to execute the steps of the method described in Embodiment 1 or Embodiment 2.
[0165] The memory may include computer system readable media in the form of volatile memory, such as random access memory (RAM) and / or cache memory. The device may further include other removable / non-removable, volatile / non-volatile computer system storage media. By way of example only, the memory may be used to read and write non-removable, non-volatile magnetic media (commonly referred to as a "hard disk drive"). A program / utility having a set (at least one) of program modules may be stored in, for example, memory. Such program modules include, but are not limited to, an operating system, one or more application programs, other program modules, and program data. Each or some combination of these examples may include an implementation of a network environment. The computer-executable program of the program modules typically performs the functions and / or methods described in the embodiments of the present invention.
[0166] The code for performing the operations of the present invention can be written in one or more programming languages or a combination thereof, including object-oriented programming languages such as Java, Smalltalk, and C++, as well as conventional procedural programming languages such as the "C" language or similar programming languages.
[0167] The processor executes various functional applications and data processing by running programs stored in memory, thereby implementing the method provided in Embodiment 1 or Embodiment 2.
[0168] Example 8:
[0169] This invention provides a storage medium containing a computer-executable program, which, when executed by a computer processor, is used to perform the method provided in Embodiment 1 or Embodiment 2.
[0170] The storage medium of embodiments of the present invention may be any combination of one or more computer-readable media. A computer-readable medium may be a computer-readable signal medium or a computer-readable storage medium. A computer-readable storage medium may be, for example, but not limited to, an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus, or device, or any combination thereof. More specific examples of computer-readable storage media (a non-exhaustive list) include: an electrical connection having one or more wires, a portable computer disk, a hard disk, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fiber, portable compact disk read-only memory (CD-ROM), optical storage device, magnetic storage device, or any suitable combination thereof. In this document, a computer-readable storage medium may be any tangible medium containing or storing a program that can be used by or in conjunction with an instruction execution system, apparatus, or device.
[0171] The code for a computer-executable program that performs the operations of this invention can be written in one or more programming languages or a combination thereof. Programming languages include object-oriented programming languages such as Java, Smalltalk, and C++, as well as conventional procedural programming languages such as C or similar languages. The program code can be executed entirely on the user's computer, partially on the user's computer, as a standalone software package, partially on the user's computer and partially on a remote computer, or entirely on a remote computer or server. In cases involving remote computers, the remote computer can be connected to the user's computer via any type of network, including a local area network (LAN) or a wide area network (WAN), or it can be connected to an external computer (e.g., via the Internet using an Internet service provider).
[0172] Of course, the computer-executable program provided in this embodiment is not limited to the above-described method operation, but can also perform related operations in the method provided in any embodiment of the present invention.
Claims
1. An end-to-end floor plan design method, characterized in that: include: The geometric relationships of each element in the floor plan and the image features of the floor plan are input into the room division model, and the floor plan with the rooms divided is output.
2. The end-to-end floor plan design method according to claim 1, characterized in that: The geometric relationships of the elements in the floor plan are obtained according to the following steps: Using a large language model and prompt words, generate a bubble diagram or relationship abstraction diagram that describes the geometric relationships of each element in the floor plan; Each element abstracts the attributes of the element, including size, shape, and position; the geometric relationships include the relative spatial positional relationships and connection relationships between the elements; The prompts describe the geometric relationships between the elements in the floor plan.
3. The end-to-end floor plan design method according to claim 1, characterized in that: The room division includes an encoder and a decoder; In the encoder, the geometric relationships between the elements in the floor plan and the image features of the floor plan are learned through a graph convolutional network to generate a Canvas Latent. The decoder converts the Canvas Latent into pixelated room partitions.
4. The end-to-end floor plan design method according to claim 3, characterized in that: The decoder uses a generator from GAN.
5. The end-to-end floor plan design method according to claim 3, characterized in that: The graph convolutional network includes: a first GS network, a first GP network, a first GO network, a second GS network, a second GP network, a second GO network, a first H network, a second H network, and a third H network. Nodes v1, v2, and edge vr1 are input into the first GS network, the first GP network, and the first GO network, respectively. Nodes v2, v3, and edge vr2 are input into the second GS network, the second GP network, and the second GO network, respectively. The output of the first GS network is input into the first H network, outputting v1'. The first GP network outputs vr1'. The outputs of the first GO network and the second and second GS networks are both input into the second H network, outputting v2'. The second GP network outputs vr2'. The output of the second GO network is input into the third H network, outputting v3'. Here, edge vr1 represents the edge connecting nodes v1 and v2, and edge vr2 represents the edge connecting nodes v2 and v3; each GS network, each GP network, and each GO network are networks with the same structure, defined as g networks, and each h network has the same structure. v1' represents the feature corresponding to node v1, v2' represents the feature corresponding to node v2, v3' represents the feature corresponding to node v3, and vr1' and vr2' represent the features corresponding to edges vr1 and vr2, respectively. For each h network, its input passes through the first hidden layer, symmetric pooling, the second hidden layer, the ReLU activation function, and the first output layer in sequence to obtain the output; For each g network, there are three inputs. Each input passes through its own Din layer and then enters the 3Din layer. The output of the 3Din layer is input to the third hidden layer. After the third hidden layer, the ReLU activation function is applied, and the output is sent to the corresponding hidden layer or Dout.
6. The end-to-end floor plan design method according to claim 3, characterized in that: The decoder includes a decoding input layer, a first average pooling layer, a first multi-layer convolution, a second average pooling layer, a second multi-layer convolution, a third average pooling layer, a third multi-layer convolution, and a decoding output layer; The input feature information to the decoder is divided into three paths. The first path is as follows: the input feature information to the decoder passes through the first average pooling layer, random noise is added, and then it is input into the first multi-layer convolution. Random noise is added again, and then it enters the second average pooling layer. The second path is as follows: the input feature information to the decoder is input into the second average pooling layer, the output of the second average pooling layer is input into the second multi-layer convolution, random noise is added, and then it enters the third average pooling layer. The third path is as follows: the input feature information to the decoder is input into the third average pooling layer, and the output of the third average pooling layer is input into the third multi-layer convolution to obtain the decoder output.
7. An end-to-end floor plan design method, characterized in that: include: The geometric relationships of the elements to be arranged in the room diagram and the image features of the room diagram are input into the furniture arrangement model, and the bounding box coordinates of each piece of furniture in the room diagram are output.
8. The end-to-end floor plan design method according to claim 7, characterized in that: The geometric relationships of the elements to be arranged in the room diagram are obtained according to the following steps: Using a large language model and prompt words, generate a bubble diagram or abstract diagram of the geometric relationships of the elements to be arranged in the room diagram; Each element abstracts the attributes of the element, including size, shape, and position; the geometric relationships include the relative spatial positional relationships and connection relationships between the elements; The prompts describe the geometric relationships between the elements to be arranged in the room diagram.
9. The end-to-end floor plan design method according to claim 7, characterized in that: The furniture arrangement model includes: an encoder and a decoder; In the encoder, the geometric relationships of each element to be placed in the room diagram and the image features of the room diagram are learned through a graph convolutional network to learn the high-dimensional abstract relationships between the elements and generate the bounding box latent space. The decoder decodes the bounding box latent space into the actual bounding box coordinates.
10. The end-to-end floor plan design method according to claim 9, characterized in that: The graph convolutional network includes: a first GS network, a first GP network, a first GO network, a second GS network, a second GP network, a second GO network, a first H network, a second H network, and a third H network. Nodes v1, v2, and edge vr1 are input into the first GS network, the first GP network, and the first GO network, respectively. Nodes v2, v3, and edge vr2 are input into the second GS network, the second GP network, and the second GO network, respectively. The output of the first GS network is input into the first H network, outputting v1'. The first GP network outputs vr1'. The outputs of the first GO network and the second and second GS networks are both input into the second H network, outputting v2'. The second GP network outputs vr2'. The output of the second GO network is input into the third H network, outputting v3'. Here, edge vr1 represents the edge connecting nodes v1 and v2, and edge vr2 represents the edge connecting nodes v2 and v3; each GS network, each GP network, and each GO network are networks with the same structure, defined as g networks, and each h network has the same structure. v1' represents the feature corresponding to node v1, v2' represents the feature corresponding to node v2, v3' represents the feature corresponding to node v3, and vr1' and vr2' represent the features corresponding to edges vr1 and vr2, respectively. For each h network, its input passes through the first hidden layer, symmetric pooling, the second hidden layer, the ReLU activation function, and the first output layer in sequence to obtain the output; For each g network, there are three inputs. Each input passes through its own Din layer and then enters the 3Din layer. The output of the 3Din layer is input to the third hidden layer. After the third hidden layer, the ReLU activation function is applied, and the output is sent to the corresponding hidden layer or Dout.
11. The end-to-end floor plan design method according to claim 10, characterized in that: The decoder uses graph convolutional networks and multilayer perceptrons to decode the bounding box latent space into bounding box coordinates.
12. An end-to-end floor plan design system, characterized in that: The steps are for performing an end-to-end floor plan design method as described in any one of claims 1 to 6 or any one of claims 7 to 11.
13. A computer device, characterized in that, The invention includes a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the steps of an end-to-end floor plan design method as described in any one of claims 1 to 6 or any one of claims 7 to 11.
14. A storage medium, characterized in that, The storage medium stores a floor plan design program, which, when executed by at least one processor, implements the steps of an end-to-end floor plan design method as described in any one of claims 1 to 6 or any one of claims 7 to 11.