Indoor three-dimensional scene layout optimization method based on wall features

Abstract

The invention discloses an indoor three-dimensional scene layout optimization method based on wall features, and the method comprises the steps: firstly obtaining furniture attributes of a scene from a data set, and carrying out the structural representation of wall information; secondly, on the basis of furniture attributes and wall information, the indoor three-dimensional scene is divided into a plurality of local scene units with conventional furniture as the center; and optimizing the layout attribute of furniture in the local scene unit by using a diffusion model. And finally, converting the optimized local three-dimensional scene information into a global three-dimensional scene, and performing confrontation optimization to complete indoor three-dimensional scene layout optimization. According to the wall structure extraction module based on the two-dimensional top view, the wall structure is reserved as a space boundary constraint, the perception ability of the model to the space structure is enhanced, the fine-grained space relationship of furniture layout can be better captured, and thus the accuracy of layout optimization is improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of indoor three-dimensional scene layout optimization, and in particular to a method for optimizing the layout of an indoor three-dimensional scene based on wall features. BACKGROUND

[0002] In the fields of game development, film production, and interior design, indoor three-dimensional scene layout generation technology exhibits important application value. This technology automatically generates high-quality virtual indoor three-dimensional scenes, significantly reducing the cost of scene layout construction in the fields of film and television, games, etc., and providing strong support for interior design, promoting the intelligentization and efficiency of the design process.

[0003] However, existing methods face two major challenges: existing generation models do not adequately consider room layout, making it difficult to accurately model the spatial relationship between furniture and between furniture and walls; and the number of existing indoor three-dimensional scene datasets is limited, often facing the situation of model overfitting due to the scarcity of training data, which has become a key bottleneck restricting the improvement of model performance.

[0004] Among them, the 3D-FRONT dataset, as a large indoor three-dimensional scene dataset, is widely used due to its rich indoor three-dimensional scene models and professional furniture layout schemes. However, such high-quality datasets still have some error conditions, which can cause bias in models trained directly based on this dataset, thereby reducing the spatial rationality and visual realism of generated indoor three-dimensional scenes. SUMMARY

[0005] To address the above problems, the present application proposes a method for optimizing the layout of an indoor three-dimensional scene based on wall features (WOSG), which generates a scene with good spatial consistency and overall expressiveness. Users only need to input an original scene to be optimized (including room structure and furniture layout) to obtain an optimized scene. Experimental results show that the present application maintains the global style properties of the source scene, significantly improves the modeling ability of fine-grained spatial relationships between furniture, and generates indoor three-dimensional scenes that exhibit obvious advantages in layout rationality and visual quality.

[0006] To accurately model the spatial relationship between furniture and between furniture and walls, the present application considers the strong constraint of walls in the scene and designs a wall extraction module to construct wall objects in the input scene data and optimize the placement of furniture. To break the size and diversity limitations of the dataset, the present application focuses on local scene details optimization and proposes a local-global optimization strategy to localize the complete scene, enriching the training data while improving the model's ability to perceive details.

[0007] Firstly, the wall structure extraction method based on two-dimensional top view is used to constrain the scene wall with two-dimensional corner point representation. The obtained wall constraint is used in combination with furniture information to construct a local scene, and the data set is expanded. The expanded data set is used to train the U-Net diffusion model, and finally the obtained local optimized scene is globally integrated.

[0008] The local-global indoor three-dimensional scene layout optimization method based on wall features comprises the following steps:

[0009] Step 1, obtain the furniture attributes of the scene from the data set, and structure the wall information;

[0010] Step 2, based on the furniture attributes and wall information, the indoor three-dimensional scene is divided into a plurality of local scene units centered on the regular furniture;

[0011] Step 3, the layout properties of the furniture in the local scene unit are optimized using a diffusion model;

[0012] Step 4, the local three-dimensional scene information obtained after optimization is converted into a global three-dimensional scene, and an adversarial optimization is performed to complete the indoor three-dimensional scene layout optimization.

[0013] The present application has the following advantages:

[0014] 1. The wall structure extraction module based on two-dimensional top view retains the wall structure as a spatial boundary constraint, enhancing the model's perception of spatial structure;

[0015] 2. The data set expansion method based on scene localization significantly expands the training data size, effectively improving the model's generalization ability, enabling the model to better capture the fine-grained spatial relationships of furniture layout, thereby improving the accuracy of layout optimization, while preventing overfitting;

[0016] 3. The local scene optimization module based on the U-Net diffusion model, the feature encoder abstracts the input features through multiple convolution and residual blocks, mapping them to a high-dimensional semantic space; the corresponding decoder reconstructs the furniture attribute information through upsampling operation and jump connection. The cross-attention mechanism significantly improves the accuracy of feature extraction and reconstruction by modeling the relevance between different features. The Time Embedding module provides time information for network dynamics, enabling the model to adaptively adjust the feature representation according to different time steps. BRIEF DESCRIPTION OF DRAWINGS

[0017] Figure 1 is the algorithm structure diagram;

[0018] Figure 2 is the scene top-down mask diagram;

[0019] Figure 3 To localize the scene schematic diagram;

[0020] Figure 4 To network structure schematic diagram. DETAILED DESCRIPTION

[0021] As Figure 1 shown, the indoor three-dimensional scene layout optimization method based on wall features includes the following steps: 1. Obtain the furniture and scene wall attributes of the scene from the data set; 2. Extract and structure the wall information from the input scene; 3. Divide the complete indoor three-dimensional scene into several local scene units centered on regular furniture; 4. Use the diffusion model to optimize the layout attributes of the furniture in the local indoor three-dimensional scene; 5. Convert the optimized local three-dimensional scene information into a global three-dimensional scene; 6. Perform network adversarial optimization.

[0022] Step 1, obtain the furniture attributes of the scene:

[0023] This step requires defining the indoor three-dimensional scene data representation and obtaining the corresponding representation from the data set.

[0024] The position information of the furniture is defined in the world coordinate system, with the geometric center of the furniture model three-dimensional bounding box as the position reference. Therefore, this algorithm focuses on the horizontal level of the coordinates in the modeling of furniture position information. The vertical direction axis height can be accurately calculated by the height of the furniture model size , that is .

[0025] The size information of the furniture is modeled in the form of a three-dimensional bounding box (length x width x height); the orientation of the furniture is parameterized in the world coordinate system, specifically with the positive direction of the axis as the reference, and the rotation angle in the counterclockwise direction to define the orientation of the furniture. The category information of the furniture is represented in a one-hot encoding manner, with different categories of furniture corresponding to discrete vectors in the semantic space. Due to the different functional attributes of different rooms, the number and types of furniture categories also vary. For example, in a living room scene, furniture categories include sofas, TV cabinets, etc., totaling 23 types, while in a bedroom scene, they include beds, wardrobes, etc., totaling 19 types.

[0026] Step 2, wall information extraction:

[0027] After obtaining the information of the furniture, the wall information needs to be extracted.

[0028] In the indoor 3D scene dataset, the wall model is controlled by several corner points. However, these corner points have the following problems in practical applications: (1) the corner point coordinates are incorrect, affecting direct use; (2) redundant corner points do not make a substantial contribution to scene understanding; (3) the distance between adjacent corner points is too small to form an effective spatial constraint.

[0029] However, the two-dimensional top view and corner data are highly correlated, so the three-dimensional problem can be reduced to a two-dimensional plane using top-view projection. First, the two-dimensional outline of the room is extracted, and then the corner points of the walls are precisely matched and their positions are adjusted through an iterative nearest-point algorithm, finally obtaining a wall structure that meets the requirements.

[0030] When processing scene corner points, the 3D scene corner points are first projected onto a 2D plane to form a set of points to be registered. ,in Refers to the first One registration point; simultaneously, a mask image is obtained based on the top-down planar view of the 3D scene, such as... Figure 2 As shown, the reference point set is extracted. ,in Refers to the first There are several reference points. In each iteration, the algorithm adjusts the value based on the currently estimated rigid transformation matrix. Pair set Perform the update. The update process includes the following steps:

[0031] First, calculate the least squares error; the goal is to find a rigid transformation matrix. (in It is a rotation matrix. (is a displacement vector), making the point set with point set The distance between them is minimized. The error function can be expressed as:

[0032]

[0033] exist middle, This indicates that the points to be registered are transformed by the transformation matrix. Position after rotation This represents the displacement vector.

[0034] Further optimize the transformation matrix by minimizing the error function. To obtain the optimal rotation matrix and displacement vector Then update the point set by applying the calculated transformation matrix to the point set. Update:

[0035]

[0036] Finally, iterate and repeat the above steps until the error function is reached. Converging to a threshold indicates that the matching between the two point sets has reached its optimal level. Through multiple iterations, the ICP algorithm can continuously adjust the position of the wall corner points to align them with the target wall point set, thereby achieving accurate wall extraction.

[0037] During the wall extraction process, a complete geometric representation system was established, including three core parameters: position, rotation, and size. The corner data obtained through the above steps allows for the creation of an accurate wall description model. For each wall, its shape can be represented by the following parameters:

[0038] First, there are the dimensional parameters; the dimensions of a wall are usually expressed in terms of its length. ,thickness and height To describe. For a wall, its length It can be calculated using the following formula:

[0039]

[0040] in and These are two adjacent corner points in space. The height and thickness of the walls are standardized to the same value in this model to simplify calculations. Secondly, there's the rotation angle; the rotation angle of the walls... This describes the rotation of the wall relative to the reference coordinate system. This angle can be calculated from the relative positions of the two endpoints of the wall.

[0041]

[0042] in and These are the coordinates of the two endpoints of the wall.

[0043] Finally, there are the location parameters. The position of the wall can be determined by comparing the coordinates of adjacent corner points. and OK. First, calculate the horizontal coordinates of the wall's center:

[0044]

[0045] Considering that the wall may rotate in the world coordinate system, the rotation angle of the wall needs to be taken into account in order to transform the center coordinates to the world coordinate system. Assume the bottom of the wall is in contact with the ground, and the height is... along If distributed along the axis, its position parameters in three-dimensional space can be expressed as:

[0046]

[0047] in The coordinates are taken as This indicates the center height of the wall. By determining the aforementioned dimensions, rotation angles, and positional parameters, the geometry of the wall is fully described, providing accurate data support for subsequent spatial analysis and model construction.

[0048] Step 3, Local Scene Construction:

[0049] Only after obtaining the two-dimensional wall information can the scene data be complete. This step requires localized processing of the scene data.

[0050] This step divides the complete 3D indoor scene into several local scene units centered around regular furniture, enabling localized processing of scene data.

[0051] First, a comprehensive analysis of the standard furniture in the interior scene is performed, establishing spatial relationships by calculating the two-dimensional Euclidean distance between the furniture. Specifically, assuming the currently selected standard furniture... The position is Then furniture and furniture Two-dimensional Euclidean distance between It can be represented as:

[0052]

[0053] Based on a preset distance threshold (Unit distance), the criteria for determining the local scene: if Then determine the furniture Belongs to furniture A localized scene centered on the subject. For example... Figure 3 As shown, the first column represents a two-dimensional global schematic diagram, while the second and third columns represent corresponding two-dimensional local schematic diagrams. This distance measurement method can effectively capture spatial combinations with functional relationships (such as the spatial configuration relationship between a bed and a bedside table), providing basic spatial units for subsequent layout optimization.

[0054] To improve the model's training performance and generalization ability, additional filtering conditions were added to the obtained local scenes. The total number of regular furniture and walls in each local scene was set. Must meet The conditional constraints of this threshold range optimize the model training effect by balancing the richness of scene information and computational complexity. This ensures that enough furniture interaction relationships are retained for feature learning, while avoiding feature dilution caused by overly complex scenes, thereby improving the model's generalization ability in real-world environments.

[0055] For the local scene that meets the restriction condition, coordinate transformation and rotation operation are performed. First, for each furniture in the local scene, the plane coordinate difference of the furniture relative to the center furniture in the local scene is calculated. Assuming that the position of the furniture is and the position of the center furniture is , the plane coordinate of the furniture relative to the center furniture is:

[0056]

[0057] At the same time, in the present algorithm, the coordinate rotation of all furniture is based on the world coordinate origin of the room. In order to unify the orientation of the furniture, the orientation of the center furniture is fixed as up, and then the coordinates are rotated through a rotation matrix, specifically, if the angle of the furniture is , the coordinates of the furniture can be transformed into new coordinates through a rotation matrix :

[0058]

[0059] The rotation matrix is defined as:

[0060]

[0061] This transformation converts the coordinates of the furniture to a unified direction, so that in the subsequent calculation, the relative position of the furniture is more clear.

[0062] Step 4, local diffusion optimization:

[0063] After the local scene is constructed, a model is used to diffuse and optimize the layout.

[0064] This step proposes a diffusion model of U-Net architecture, which aims to optimize the layout attributes of furniture in the local indoor three-dimensional scene.

[0065] As a breakthrough technology in the field of generative artificial intelligence, the diffusion model can effectively capture the characteristics of data distribution through its iterative denoising mechanism, and generate visual content with high fidelity and rich details. In the task of two-dimensional image generation, the diffusion model exhibits excellent generalization ability and can generate diverse image content. When migrating to the task of indoor three-dimensional scene layout generation, the diffusion model can generate indoor three-dimensional scenes with diverse styles and reasonable layouts based on the given room structure. However, the diffusion model has limitations in regulating the pose parameters of furniture. The layout of the indoor three-dimensional scene is highly dependent on the position, orientation, and spatial relationship with other furniture of the furniture, so compared to two-dimensional image generation, the indoor three-dimensional scene layout generation task has more stringent requirements for the spatial layout of furniture.

[0066] The scene attributes optimized by the model of the application mainly include position , size , orientation and category . In order to effectively process the input furniture features, as shown in Figure 4 , the diffusion model adopts multiple key modules: feature encoder, feature decoder, cross-attention mechanism, self-attention mechanism and residual block. The feature encoder gradually abstracts the input scene attributes through multiple convolutional layers and residual blocks, and maps them to a high-dimensional semantic space; the corresponding decoder gradually reconstructs the attribute information of the furniture through upsampling operation and jump connection. The cross-attention mechanism significantly improves the accuracy of feature extraction and reconstruction by modeling the relevance between different features. The time embedding module provides time information for network dynamics, so that the diffusion model can adaptively adjust the feature representation according to different time steps.

[0067] In summary, this module takes a 15-dimensional attribute matrix of furniture in the local scene as input (including regular furniture and walls as special furniture), and after mapping by the encoder, it applies conditional control and loss constraints during the denoising process to optimize the layout. The wall attributes remain unchanged as a fixed context, which maintains the integrity of the scene structure and avoids redundant parameter updates, ensuring training stability.

[0068] Step 5: Global scene integration:

[0069] After obtaining the optimized local scene, the local scene needs to be integrated.

[0070] This step will focus on the presentation of the final result of the algorithm, converting the optimized local three-dimensional scene information into a global three-dimensional scene.

[0071] Let the position of the central furniture in the local coordinate system be , and the rotation angle of the central furniture in the local coordinate system be to a position and an angle in the global coordinate system.

[0072] Firstly, the transformation of the position, the position information of the central furniture in the local coordinate system needs to be converted to the global coordinate system by a rotation matrix and a displacement vector . Assuming that the position of the furniture in the global coordinate system is , the transformation formula is:

[0073]

[0074] wherein is a rotation matrix describing the current local coordinate system to the global coordinate system, is the rotation angle of the local coordinate system relative to the global coordinate system.

[0075] Subsequently, the angle transformation of other furniture in the local scene is performed. In the local coordinate system, the rotation angle of the furniture is usually relative to the local coordinate system, while in the global coordinate system, it needs to be adjusted appropriately to be consistent. Assuming that the rotation angle of the furniture in the global coordinate system is , the transformation formula is:

[0076]

[0077] wherein is the rotation angle of the local coordinate system relative to the global coordinate system.

[0078] In order to restore the information from the local coordinate system to the global coordinate system, the algorithm first restores the local data to the range of the global coordinate system through the inverse normalization operation. Specifically, by restoring the position, angle and other furniture attributes in the local data, these local information can be mapped back to the global coordinate system. The restoration process includes two main steps of displacement recovery and angle recovery:

[0079] Firstly, the displacement recovery, the position in the local coordinate system is adjusted by the offset relative to the origin of the local coordinate system, and further compensated by using the offset of the global coordinate system.

[0080] Secondly, the angle recovery, for the rotation angle, the recovery process adjusts the angle by using the inverse rotation matrix. Assuming that the rotation angle of the furniture in the local coordinate system is , the angle in the global coordinate system is .

[0081] In actual implementation, the recovery of the rotation angle is closely related to the process of position recovery, and usually needs to be adjusted by using the inverse rotation matrix.

[0082] To finally integrate all local results into a global scene, the algorithm first merges all local scene

[0083] The position information, angle information, size information, and category labels of each piece of furniture are merged. Since each piece of furniture can generate multiple optimization results after local restoration, the algorithm obtains the final position, angle, and size information by averaging these merged results. By integrating these information, a complete global scene containing all furniture information is finally obtained. The core formula of this process is:

[0084]

[0085] where, for each piece of furniture its global information is calculated by:

[0086]

[0087] where represents the global scene, containing the information of all furniture (including position, angle, size, and category label), represents the number of results of each piece of furniture after local scene restoration, represents the information of the th piece of furniture in the th local restoration.

[0088] Through this processing method, local scene information is effectively integrated, generating a global scene containing complete furniture information.

[0089] Step 6, adversarial optimization:

[0090] To effectively realize the joint optimization of multiple furniture attributes, this paper combines multiple loss functions, and the specific loss function settings are as follows:

[0091] (1) Mean square error loss:

[0092] The mean square error loss is applied to the noise recovery stage of the model. This loss minimizes the difference between the predicted noise and the true noise, ensuring that the network can accurately restore the original layout during the diffusion process. Specifically, the calculation method of the MSE loss is:

[0093]

[0094] where represents the predicted noise of the model at time step ; is the true noise added during the diffusion process; is the scene data after adding noise at time step ; is the scene data after adding noise at time step ; is the scene data after adding noise at time step representing conditional information, such as furniture category, etc.

[0095] (2) Displacement and orientation loss:

[0096] The displacement loss of furniture is measured by L2 norm of the predicted position to the ground truth position in Euclidean distance:

[0097]

[0098] Considering the periodic nature of the angle, the cosine similarity loss is used to measure the angle loss of furniture to avoid the 2π periodicity problem:

[0099]

[0100] (3) Regular furniture IoU loss:

[0101] To improve the accuracy of the prediction of the bounding box of regular furniture, the regular furniture IoU loss is designed to optimize the overlapping area of the bounding boxes of regular furniture. The specific formula is as follows:

[0102]

[0103] where and are the predicted bounding boxes of different regular furniture, and are their intersection and union areas, respectively.

[0104] Wall IoU:

[0105] To effectively model the spatial constraint relationship between regular furniture and walls in indoor three-dimensional scenes, this paper designs a wall IoU loss. This loss function quantifies the degree of spatial overlap between the predicted bounding box of regular furniture and the wall bounding box, guiding the model to learn furniture layout patterns that conform to architectural specifications. The specific formula is as follows:

[0106]

[0107] where represents the predicted bounding box of the th regular furniture, represents the corresponding wall ground truth bounding box.

Claims

1. A method for layout optimization of an indoor three-dimensional scene based on wall features, characterized in that, The method comprises the following steps: Step 1, obtaining furniture attributes of a scene from a data set, and structurally representing wall information; Step 2, based on the furniture attributes and the wall information, segmenting the indoor three-dimensional scene into a plurality of local scene units centered on regular furniture; Step 3, optimizing the layout attributes of the furniture in the local scene units using a diffusion model; Step 4, converting the obtained local three-dimensional scene information after optimization into a global three-dimensional scene, and performing adversarial optimization to complete the indoor three-dimensional scene layout optimization.

2. The wall feature based indoor three-dimensional scene layout optimization method according to claim 1, wherein, The specific implementation process of obtaining the furniture attributes of the scene is as follows: The position information of the furniture is defined in a world coordinate system, with the geometric center of a three-dimensional bounding box of a furniture model as a position reference. In modeling the position information of the furniture, attention is paid to the horizontal coordinates in a horizontal plane, and the vertical height is calculated according to the size of the furniture model. The vertical height is calculated according to the size of the furniture model.​ The size information of the furniture is modeled in the form of a three-dimensional bounding box; the orientation of the furniture is represented by an angle parameterization in a world coordinate system, specifically in the form of The positive direction of the axis is the reference, and the rotation angle in the counterclockwise direction The orientation of the furniture is defined; the category information of the furniture is represented by one-hot encoding, and different categories of furniture correspond to discrete vectors in the semantic space, and the number and types of furniture categories are different.

3. The wall feature based indoor three-dimensional scene layout optimization method according to claim 2, characterized in that, The specific implementation process of structurally representing the wall information is as follows: After obtaining the information of the furniture, in the indoor three-dimensional scene data set, the wall model is controlled by a plurality of corner points; Firstly, the two-dimensional contour of the room is extracted, and then the wall corner points are matched and position-adjusted through an iterative closest point algorithm, so that a wall structure meeting the requirements is finally obtained; In processing the scene corner points, first project the three-dimensional scene corner points to a two-dimensional plane to form a set of points to be registered , wherein the element refers to the th point to be registered; meanwhile, based on a three-dimensional scene plan view, a mask map is obtained, and a set of reference points is extracted , wherein the element refers to the th reference point; in each iteration, based on the current estimated rigid transformation matrix , the point set is updated; For each wall, its shape is represented by the following parameters: The dimensions of the wall are usually described using its length , thickness and height ; for a wall, its length is calculated by the two adjacent corner points in the space; Secondly, the rotation angle, the rotation angle of the wall The rotation of the wall relative to the reference coordinate system is described, and the angle is calculated by the relative positions of the two end points of the wall; the position of the wall is determined by the coordinate pair of the adjacent corner points; the central horizontal coordinate of the wall is calculated by the coordinate pair of the adjacent corner points, considering the rotation angle of the wall , assuming that the bottom surface of the wall is in contact with the ground, and the height is distributed along the axis direction, according to the central horizontal coordinate , combined with the rotation angle , the position parameter representation of the wall in three-dimensional space is obtained , wherein The coordinates are taken as to represent the central height of the wall.

4. The wall feature based indoor three-dimensional scene layout optimization method according to claim 3, characterized in that, The pair of point sets The process of updating is as follows: First, the least square error is calculated to find a rigid transformation matrix where is a rotation matrix, is a displacement vector, such that the set of points has the minimum distance to the set of points The error function is represented as ; Re-optimizing the transformation matrix by minimizing an error function Obtaining an optimal rotation matrix and a displacement vector ; then updating the point set using the to-be-registered points by the transformation matrix the position after rotation plus the displacement vector , the point set is updated ; The final iteration is performed, repeating the above steps until the error function converges to a threshold value.

5. The wall feature based indoor three-dimensional scene layout optimization method according to claim 4, characterized in that, The specific implementation process of step 2 is as follows: First, the furniture in the indoor scene is traversed and analyzed. Spatial relationships between furniture are established by calculating the two-dimensional Euclidean distance between them. Assuming the currently selected furniture... Location Then furniture and furniture The two-dimensional Euclidean distance between them is expressed as ; based on a preset distance threshold , a decision criterion for constructing a local scene is: if , it is decided that the furniture belongs to a local scene centered on the furniture ; For local scenes that meet the constraints, coordinate transformation and rotation operations are performed: First, for each piece of furniture within the local scene, the difference in its planar coordinates relative to the central furniture within the local scene is calculated; assuming the furniture... The position is And central furniture Location Then furniture The planar coordinates relative to the center furniture are and The corresponding coordinates are subtracted to obtain the result; Meanwhile, the coordinate rotation of all furniture is based on the world coordinate origin of the room, in order to unify the orientations of the furniture, the orientation of the central furniture is fixed as up, and the coordinates of the furniture are transformed into new coordinates by rotating the matrix if the angle of the furniture is . .

6. The wall feature based indoor three-dimensional scene layout optimization method according to claim 5, wherein, The specific implementation process of step 3 is as follows: A diffusion model of U-Net architecture is proposed to optimize the layout attributes of furniture in the local scene unit of the indoor three-dimensional scene; the scene attributes optimized by the diffusion model include position , size , orientation and category ; The diffusion model comprises a feature encoder, a feature decoder, a cross-attention mechanism, a self-attention mechanism and a residual block; the feature encoder maps the input scene attributes to a semantic space through multi-layer convolution and a residual block for step-by-step abstraction; the corresponding decoder gradually reconstructs the attribute information of the furniture through upsampling operation and jump connection; the cross-attention mechanism improves the accuracy of feature extraction and reconstruction by modeling the relevance between different features; and the time embedding module dynamically provides time sequence information, so that the diffusion model can adaptively adjust the feature representation according to different time steps.

7. The wall feature based indoor three-dimensional scene layout optimization method according to claim 6, wherein, The specific implementation process of converting the obtained local three-dimensional scene information after optimization into a global three-dimensional scene is as follows: the position of the central furniture in the local coordinate system the angle of rotation of the central furniture in the local coordinate system converts this into a position and angle in the global coordinate system; First is the position transformation, the central furniture in the local coordinate system position information By multiplying the rotation matrix Plus displacement vector Convert to the global coordinate system, get the position of the furniture in the global coordinate system ; Subsequently, the angle transformation of other furniture in the local scene is performed, in the local coordinate system, the rotation angle of the furniture is relative to the local coordinate system, and in the global coordinate system, the rotation angle is adjusted to be consistent, and the rotation angle of the furniture is obtained by adding the rotation angle of the local coordinate system relative to the global coordinate system , to obtain the rotation angle of the furniture in the global coordinate system ; The information is recovered from the local coordinate system to the global coordinate system, firstly, the local data is restored to the range of the global coordinate system through the inverse normalization operation: the local information is mapped back to the global coordinate system by restoring the furniture attributes in the local data, and the recovery process includes two processes of displacement recovery and angle recovery.

8. The wall feature based indoor three-dimensional scene layout optimization method according to claim 7, wherein, The two processes of displacement recovery and angle recovery specifically include: merging the position information, angle information, size information and category label of the furniture in all local scenes, obtaining the final position, angle and size information by averaging the merging results, and finally obtaining a complete global scene containing all the furniture information.

9. The wall feature based indoor three-dimensional scene layout optimization method according to claim 8, wherein, In the adversarial optimization process, the loss function is set as follows: Mean square error loss: applied to the noise recovery stage of the model, the loss minimizes the difference between the predicted noise and the real noise; Displacement and orientation loss: the L2 norm is used to measure the Euclidean distance between the predicted position and the real position; the angle loss of the furniture uses the cosine similarity loss; Furniture intersection-over-union loss: used to optimize the overlapping area of the bounding boxes of the regular furniture, and a wall intersection-over-union loss is designed to quantify the spatial overlap between the predicted regular furniture bounding box and the wall bounding box, guiding the model to learn the furniture layout mode conforming to the building specifications.

Images