Instance segmentation and model reconstruction method and system based on deep learning and synthetic point cloud

Abstract

The invention discloses an instance segmentation and model reconstruction method and system based on deep learning and synthetic point clouds, and the method comprises the steps: firstly, carrying out the sampling of a BIM model, and generating an instance-level synthetic point cloud; comprising the following steps: uniformly sampling to generate a non-occlusion point cloud, generating an occlusion point cloud by applying a hidden point removal algorithm, and selecting candidate viewpoints to simulate the occlusion point cloud in a real scene. And then, the generated synthetic point cloud is used for training a deep learning instance segmentation model, and the trained model is used for segmenting an actual point cloud scene, so that component-level accurate identification is realized. And finally, parametric modeling is carried out on the segmented point cloud through Dynamo software, a building component is reconstructed according to an instance segmentation result, and three-dimensional model reconstruction in a complex scene is completed. The method can effectively improve the precision and modeling efficiency of point cloud segmentation in a complex environment, and is widely applied to the fields of building intelligent operation and maintenance, electromechanical equipment management and the like.

Description

Technical Field

[0001] The present invention belongs to the fields of civil engineering and artificial intelligence, and particularly relates to a method and system for instance segmentation and model reconstruction based on deep learning and synthetic point clouds. Background Art

[0002] The 3D reconstruction technology based on point clouds can generate building information models, thus effectively promoting the operation and maintenance of building components. Point cloud data only contains geometric information and lacks semantic annotations. Therefore, further object recognition and classification processing are required to achieve model reconstruction based on point clouds. However, traditional machine learning-based semantic segmentation methods face problems such as low efficiency and difficult parameter tuning when dealing with complex scenarios. Summary of the Invention

[0003] Aiming at the problems existing in the prior art, the present invention provides a method and system for instance segmentation and model reconstruction based on deep learning and synthetic point clouds, which can support the model reconstruction and operation and maintenance management of the building information model (BIM) system and improve the efficiency of model reconstruction in building scenarios.

[0004] To solve the above technical problems, the present invention provides the following technical solution: A method for instance segmentation and model reconstruction based on deep learning and synthetic point clouds, comprising the following steps:

[0005] S1. Generate instance-level synthetic point clouds by sampling each component in the target scene with a BIM model, including first uniformly sampling the BIM model to generate unoccluded point clouds, then removing hidden points to generate occluded synthetic points, and finally selecting key points to synthesize the point cloud scene under real scene occlusion;

[0006] S2. Construct a synthetic point cloud dataset, including the synthetic point clouds obtained in step S1 and the corresponding labels, construct a deep learning network for point cloud component segmentation, use the synthetic point clouds as input and the corresponding class labels as output, train the deep learning network for point cloud component segmentation to obtain a point cloud component segmentation model, and then use the trained point cloud component segmentation model to segment the actual point cloud scene to obtain the point cloud segmentation result of the target scene;

[0007] S3. Based on the point cloud segmentation result, use Dynamo to reconstruct the model of the segmented point cloud and output the constructed position and size information.

[0008] Further, in the aforementioned step S1, the unoccluded point clouds generated by uniformly sampling the BIM model are obtained according to the following steps: SA-1.1. Divide the surface of each component into several triangular faces and calculate the area A of each triangular face i The following formula:

[0009]

[0010] Among them, v 1 , v 2 , v 3 are the vertices of the triangular faces,

[0011] SA - 1.2. Add up the areas of each triangular face to calculate the total surface area A total ,

[0012]

[0013] where N is the total number of triangular faces in the surface mesh;

[0014] SA - 1.3. Calculate the number of sampling points n for each triangular face i . The number of sampling points n is proportional to the proportion of its area in the total surface area, and is calculated as follows

[0015]

[0016] where N total is the total number of sampling points;

[0017] SA - 1.4. The sampling points of each triangular face are generated by barycentric coordinates. By varying the barycentric coordinates r 1 and r 2 , ensure that the sampling points are evenly distributed within the triangular face to achieve uniform sampling on the surface of the 3D model. The sampling point p is calculated by the following formula:

[0018] p = (1 - r 1 - r 2 )v 1 + r 1 v 2 + r 2 v 3 , where r 1 , r 2 ∈[0, 1), r 1 + r 2 < 1 (4)

[0019] where v 1 , v 2 , v 3 are the vertices of the triangle.

[0020] Furthermore, in the aforementioned step S1, a hidden point removal algorithm is used to generate an occlusion point removal synthetic point cloud. The specific steps are as follows:

[0021] SB - 1.1. Define the observer position c ∈ R 3 , and set it as the center of the sphere enclosing the point cloud. For each point p i ∈ P, project the point to the outside of the view point through the following formula, that is, the transformed point:

[0022] q i = p i + λ(p i - c) (5)

[0023] where λ > 0 is a scaling factor to ensure that q i is located outside c;

[0024] SB - 1.2. Calculate the convex hull H of the transformed points, as follows. The vertices of the convex hull correspond to the visible points,

[0025]

[0026] In the formula, ConvexHull represents performing a convex hull operation on the set of transformed points q i that is, calculating the smallest convex polyhedron or convex polygon of a set of points that contains all the transformed points. The vertices of the convex hull are the visible points;

[0027] SB - 1.3. Mark as visible points: The points p i in the original point cloud associated with these vertices are marked as visible points, as follows:

[0028] P HPR = {p i | q i ∈ H} (7)

[0029] Furthermore, in the aforementioned step S1, selecting key points to simulate point clouds under real - scene occlusion includes the following sub - steps: SC - 1.1. Represent the room feasible region map as a grid map M ∈ R H×W , where H and W respectively represent the height and width of the room. If the grid cell M(i, j) is a feasible region, it is assigned a value of 1; otherwise, it is assigned a value of 0. The grid resolution is defined by s, where s represents the actual distance corresponding to each grid cell;

[0030] SC - 1.2. Determine the candidate key points P cand as the center points of all feasible grid cells, that is, the grid cells that satisfy M(i, j) = 1; SC - 1.3. Set a minimum distance constraint d min to ensure that the key points are not too densely clustered, thus achieving a more uniform room coverage. Specifically: For any two selected key points p i and p j , the distance D(p i , p j ) needs to satisfy D(p i , p j ) ≥ d min ;

[0031] SC-1.4. The final selection step aims to ensure full coverage of the feasible region. Each selected key point represents a specific coverage area. When the coverage area exceeds the preset threshold, the viewpoint selection stops. If the required coverage rate is not reached, a greedy algorithm is introduced to iteratively select points that can maximize the uncovered area while keeping the minimum distance constraint unchanged.

[0032] Furthermore, in the aforementioned step S1, when synthesizing the point cloud scene under real - scene occlusion, noise is simulated. The noise is simulated in the synthesized point cloud through a Gaussian function. The probability density function of the Gaussian random variable z is defined as follows:

[0033]

[0034] where z is the Gaussian random variable representing the value of the noise, μ is the mean of the Gaussian distribution, controlling the central position of the noise, σ is the standard deviation of the Gaussian distribution, controlling the distribution width or spread of the noise, and PG(z) is the probability density function of the Gaussian noise, representing the probability of a specific noise value z occurring.

[0035] Furthermore, in the aforementioned step S2, the class labels include walls, tables, columns, air conditioners, and electrical boxes.

[0036] Furthermore, in the aforementioned step S2, the point cloud component segmentation deep - learning network is constructed based on oneformer3D.

[0037] Furthermore, in the aforementioned step S3, Dynamo is used to perform model reconstruction on the segmented point cloud. First, model reconstruction is performed on the independent components, and then on the dependent components.

[0038] On the other hand, the present invention provides an electronic device, including a memory, a processor, and a computer program stored on the memory and executable on the processor. When the processor executes the computer program, the steps of any method described in the present invention are implemented.

[0039] The present invention also provides a computer - readable storage medium, on which a computer program is stored. When the computer program is executed by a processor, the steps of any method described in the present invention are implemented.

[0040] Compared with the prior art, the beneficial technical effects of the present invention adopting the above - mentioned technical solutions are as follows: The previous model reconstruction methods relied on semantic segmentation of deep learning, but a large number of machine - learning parameter adjustments were still required later, which was time - consuming and laborious. The model reconstruction method of this patent directly uses instance segmentation of deep learning to accurately segment each component in the point cloud, thus more effectively promoting the model reconstruction. In addition, to alleviate the problem of insufficient training data, this patent proposes a method for synthesizing point cloud enhancement to improve the effects of instance segmentation and model reconstruction by synthesizing point clouds. Description of the Drawings

[0041] Figure 1 is the flowchart of the method of the present invention.

[0042] Figure 2 is a schematic diagram of the algorithm flow for generating instance-level synthetic point clouds from BIM model sampling.

[0043] Figure 3 is an example diagram of an indoor walkable map and key point selection.

[0044] Figure 4 is a comparison schematic diagram of the generated point cloud scene and the complete point cloud scene. In the figure, (a) is the real point cloud scene diagram, and (b) is the generated synthetic point cloud scene diagram.

[0045] Figure 5 is the real point cloud scene and instance segmentation diagram of the meeting room in the embodiment. In the figure, (a) is the schematic diagram of the real point cloud scene of the meeting room, and (b) is the schematic diagram of the instance segmentation situation of the meeting room.

[0046] Figure 6 is the model reconstruction diagram of Dynamo and the schematic diagram of the BIM model after model reconstruction is completed. Detailed Embodiments

[0047] For a better understanding of the technical content of the present invention, specific embodiments are hereby given and described in conjunction with the accompanying drawings as follows.

[0048] In the present invention, various aspects of the present invention are described with reference to the accompanying drawings, in which many illustrative embodiments are shown. The embodiments of the present invention are not limited to those described in the drawings. It should be understood that the present invention can be implemented by any one of the various concepts and embodiments introduced above, and those described in detail below, because the concepts and embodiments disclosed in the present invention are not limited to any embodiment. In addition, some aspects disclosed in the present invention can be used alone, or in any suitable combination with other aspects disclosed in the present invention.

[0049] Reference Figure 1 , the present invention provides a method for instance segmentation and model reconstruction based on deep learning and synthetic point clouds, including the following steps:

[0050] S1. Generate instance-level synthetic point clouds by sampling the BIM model for each component in the target scene, including first uniformly sampling the BIM model to generate unoccluded point clouds, then removing hidden points to generate occluded synthetic points, and finally selecting key points to synthesize the point cloud scene under real scene occlusion;

[0051] S2. Construct a synthetic point cloud dataset, including the synthetic point cloud obtained in step S1 and the corresponding labels. Construct a deep learning network for point cloud component segmentation, with the synthetic point cloud as the input and the corresponding class label as the output. Train the deep learning network for point cloud component segmentation to obtain a point cloud component segmentation model, and then use the trained point cloud component segmentation model to segment the actual point cloud scene to obtain the point cloud segmentation result of the target scene;

[0052] S3. Based on the point cloud segmentation result, use Dynamo to reconstruct the model of the segmented point cloud and output the constructed position and size information.

[0053] Reference Figure 2 , as a preferred embodiment of the present invention, in step S1, the BIM model is uniformly sampled to generate an unoccluded point cloud, including exporting Autodesk Revit to the obj format, importing the obj format file using a python library, and developing an instance-level synthetic point cloud algorithm using python. The specific steps are as follows:

[0054] SA-1.1. First, obtain a complete point cloud through uniform sampling. For each component of the obj file, it is composed of the areas of each triangle. Divide the surface of each component into several triangular faces and calculate the area A of each triangular face i The following formula:

[0055]

[0056] where v 1 , v 2 , v 3 are the vertices of the triangular face,

[0057] SA-1.2. Add up the areas of each triangular face to calculate the total surface area A total ,

[0058]

[0059] where N is the total number of triangular faces in the surface mesh.

[0060] SA-1.3. Calculate the number of sampling points n for each triangular face i , and the number of sampling points n is proportional to the proportion of its area in the total surface area, calculated as follows

[0061]

[0062] where N total is the total number of sampling points;

[0063] SA-1.4. For each triangular face in the OBJ model, the sampling points of each triangular face are generated by barycentric coordinates. By varying the barycentric coordinates r 1 and r 2 , ensure that the sampling points are evenly distributed within the triangular face, achieving uniform sampling on the surface of the 3D model. The sampling point p is calculated by the following formula:

[0064] p = (1 - r 1 - r 2 )v 1 + r 1 v 2 + r 2 v 3 , where r 1 , r 2 ∈[0, 1), r 1 + r 2 < 1 (4)

[0065] where v 1 , v 2 , v 3 are the vertices of the triangle.

[0066] This method ensures that the sampling points are evenly distributed within the triangle by varying the barycentric coordinates r 1 and r 2 , thus achieving uniform sampling on the surface of the 3D model.

[0067] To simulate the occlusion effect in the real world, the present invention uses the HPR (Hidden Point Removal) algorithm to generate the synthetic point cloud with occluded points removed (Synthetic_H). This algorithm identifies visible points from a specified viewpoint and removes the occluded points.

[0068] As a preferred embodiment of the present invention, the hidden point removal algorithm is used to generate the synthetic point cloud with occluded points removed. The specific steps are as follows:

[0069] SB-1.1. Define the observer position c ∈ R 3 , set as the center of the sphere enclosing the point cloud. For each point p i ∈ P, project the point to the outside of the viewpoint through the following formula, i.e., the transformed point:

[0070] q i = p i + λ(p i - c) (5)

[0071] where λ > 0 is the scaling factor to ensure that q i is located further outside c.

[0072] SB-1.2. Calculate the convex hull H of the transformed points as follows. The vertices of the convex hull correspond to the visible points.

[0073]

[0074] Wherein, ConvexHull represents the convex hull operation on the transformed point set q i That is, calculate the minimum convex polyhedron or convex polygon of a set of points, which contains all the transformed points. The vertices of the convex hull are the visible points.

[0075] SB-1.3. Mark the visible points: The points p i associated with these vertices in the original point cloud are marked as visible points as follows:

[0076] P HPR ={p i 」q i ∈H} (7)

[0077] After generating a single occluded point through the HPR algorithm, a view point is generated by generating a map of the feasible region of the room. For each view point, the HPR algorithm is used to generate an occluded visible point cloud from the entire point cloud, and the intersection is taken to obtain the point cloud occluded by the true situation of the entire scene. This method first initializes a two-dimensional map of the feasible region of the room, where the infeasible regions (such as desks and chairs) are marked accordingly. To avoid the view point being too close to the infeasible region, a buffer zone (such as 0.2 meters) is set around the infeasible region. The regions located outside the infeasible region and the buffer zone are defined as the feasible regions. Subsequently, view points are selected within the feasible regions and the HPR operation is performed.

[0078] Select key points to simulate the point cloud under real scene occlusion: After generating a single occluded point in the scene, a view point is generated by generating a map of the feasible region of the room. For each candidate key point, the hidden point removal algorithm is used to generate an occluded visible point cloud from the entire point cloud, and the intersection is taken to obtain the point cloud occluded by the true situation of the entire scene. It includes the following sub-steps:

[0079] SC-1.1. Grid map representation: Represent the map of the feasible region of the room as a grid map M∈R H×W , where H and W respectively represent the length and width of the room. If the grid cell M(i,j) is a feasible region, it is assigned a value of 1, otherwise it is assigned a value of 0. The grid resolution is defined by s, and s represents the actual distance corresponding to each grid cell;

[0080] SC-1.2. Candidate key point selection: Determine the candidate key points P cand as the center points of all feasible grid cells, that is, the grid cells that satisfy M(i,j)=1;

[0081] SC-1.3. Distance constraint screening: To ensure that there is enough spacing between the selected key points, set a minimum distance constraint d min, ensure that key points are not too densely clustered, thus achieving more uniform room coverage. Specifically, for any two selected key points p i and p j , the distance D(p i ,p j ) needs to satisfy D(p i ,p j )≥d min .

[0082] SC-1.4, Full-coverage Key Point Selection: The final selection step aims to ensure the complete coverage of the feasible region. Each selected key point represents a specific coverage area. When the coverage area exceeds a preset threshold (such as 70% of the room), the viewpoint selection stops. If the required coverage rate is not reached, a greedy algorithm is introduced to iteratively select points that can maximize the uncovered area while keeping the minimum distance constraint unchanged.

[0083] In step S1, when synthesizing the point cloud scene under the occlusion of the real scene, noise is simulated. The noise is simulated in the synthesized point cloud through a Gaussian function. The probability density function of the Gaussian random variable z is defined as follows:

[0084]

[0085] where z is the Gaussian random variable representing the value of the noise, μ is the mean of the Gaussian distribution, controlling the central position of the noise, σ is the standard deviation of the Gaussian distribution, controlling the distribution width or spread of the noise, and PG(z) is the probability density function of the Gaussian noise, representing the probability of a specific noise value z appearing.

[0086] Subsequently, according to the existing BIM model, such as the BIM model of the existing S3DIS dataset, this method is used to generate the synthesized point cloud. An example of a real point cloud scene and the synthesized point cloud generated by the present invention is Figure 4 shown in the figure. In the figure, (a) is the real point cloud scene diagram, and (b) is the generated synthesized point cloud scene diagram. It can be seen that the synthesized point cloud scene can accurately reflect the occlusion situation of the real scene point cloud.

[0087] After generating the synthesized point cloud, the present invention uses deep learning to perform instance segmentation on the point cloud components in combination with the generated synthesized point cloud. For example, the generated synthesized point cloud is brought into a deep learning model (such as oneformer3D) for training and inference on the real scene to obtain a point cloud scene with example annotations.

[0088] Specifically, the present invention uses a FARO FOCUS PREMIUM 150 three-dimensional laser scanner to scan four meeting rooms (Meeting Room 301, Meeting Room 302, Meeting Room 304, and Meeting Room 319) in the School of Civil Engineering of Southeast University. After the scanning is completed, the synthetic point cloud method of the present invention is used to segment components in combination with deep learning instance segmentation. Taking Meeting Room 301 as an example, its real point cloud and segmentation results are as Figure 5 shown. In the figure, (a) shows a schematic diagram of the real point cloud scene of Meeting Room 301, and (b) shows a schematic diagram of the instance segmentation of Meeting Room 301. This method can successfully segment semantic categories such as walls, tables, and columns. In addition, miscellaneous elements such as air conditioners and electrical boxes are also correctly identified.

[0089] The present invention uses the instance segmentation results to implement parametric modeling of building components in Dynamo software. This method includes determining the position and size of the components and classifying them into two categories according to the complexity of the components:

[0090] Components that only require position information, such as chairs; components that require determining the position and detailed size, such as beams, tables, walls, and windows. The sizes of some components are determined based on empirical values. For example, the floor slab thickness is set to 150 mm, and the wall thickness is set to 200 mm.

[0091] In terms of the modeling sequence, the present invention gives priority to constructing independent components and then modeling dependent components. Independent components (such as chairs and tables) are independent of each other and can be directly generated in Dynamo. Dependent components (such as windows and doors) need to rely on supporting elements. The wall model needs to be established first, and then the positions of windows and doors are calculated according to the wall structure.

[0092] For the point cloud components classified as cluttered, the present invention initially models all the segmented miscellaneous instances into a Revit general BIM model. Subsequently, these general models are manually corrected according to the actual situation. For example, the general family is corrected into an air conditioner family, etc. As Figure 6 shown, the parametric modeling process realizes the automated modeling of most components, demonstrating the high efficiency of the Dynamo modeling method.

[0093] The laboratory conditions required for training include: a high-performance GPU (such as NVIDIA RTX 4090), running a Linux system, using the Python programming language, and the Pytorch deep learning framework.

[0094] On the other hand, the present invention provides an electronic device, including a memory, a processor, and a computer program stored on the memory and executable on the processor. When the processor executes the computer program, the steps of any method described in the present invention are implemented.

[0095] The present invention also provides a computer-readable storage medium, on which a computer program is stored, and when the computer program is executed by a processor, the steps of any one of the methods described in the present invention are implemented.

[0096] Although the present invention has been described above with preferred embodiments, it is not intended to limit the present invention. Those with ordinary knowledge in the technical field to which the present invention pertains can make various modifications and refinements without departing from the spirit and scope of the present invention. Therefore, the protection scope of the present invention shall be subject to that defined by the claims.

Claims

1. A method for instance segmentation and model reconstruction based on deep learning and synthetic point cloud, characterized in that: The following steps are involved: S1. Generate instance-level synthetic point clouds using BIM model sampling for each component in the target scene, including first generating unobstructed point clouds based on uniform sampling of the BIM model, then removing hidden points to generate occluded synthetic points, and finally selecting key points to synthesize point cloud scenes under occlusion of the real scene; S2, constructing a synthetic point cloud data set, including the synthetic point cloud obtained in step S1 and the corresponding labels, constructing a point cloud component segmentation deep learning network, taking the synthetic point cloud as input and the corresponding category labels as output, training the point cloud component segmentation deep learning network to obtain a point cloud component segmentation model, and then using the trained point cloud component segmentation model to segment the actual point cloud scene to obtain the point cloud segmentation result of the target scene; S3. Based on the point cloud segmentation results, Dynamo is used to reconstruct the point cloud model after segmentation, and the constructed position and size information is output.

2. The method for instance segmentation and model reconstruction based on deep learning and synthetic point cloud according to claim 1, characterized in that: In step S1, the BIM model is uniformly sampled to generate an unobstructed point cloud according to the following steps: SA-1.

1. Divide the surface of each component into several triangular faces and calculate the area A of each triangular face i The following formula: Among them, v1, v2, v3 are the vertices of the triangle surface. SA-1.

2. Add the areas of each triangle to calculate the total surface area A. total , Where N is the total number of triangles in the surface mesh; SA-1.

3. Calculate the number of sampling points n for each triangle i , the number of sampling points n is proportional to the proportion of their area in the total surface area, calculated as follows: Among them, N total is the total number of sampling points; SA-1.

4. Each triangular surface sampling point is generated by the barycentric coordinates. By changing the barycentric coordinates r1 and r2, the sampling points are evenly distributed in the triangular surface to achieve uniform sampling of the 3D model surface. The sampling point p is calculated by the following formula: p=(1-r1-r2) v1+r1v2+r2v3, where r1, r2∈[0, 1), r1+r2<1 (4) Among them, v1, v2, and v3 are triangle vertices.

3. The method for instance segmentation and model reconstruction based on deep learning and synthetic point cloud according to claim 1, characterized in that: In step S1, a hidden point removal algorithm is used to generate an occlusion point removal synthetic point cloud. The specific steps are as follows: SB-1.

1. Define the observer position c∈R 3 , set to the center of the sphere surrounding the point cloud, for each point p i ∈P, the point is projected outside the viewpoint, i.e., the transformation point, by the following formula: what i =p i +λ(p i -c) (5) Among them, λ>0 is a scaling factor to ensure that q i A position outside of c; SB-1.

2. Calculate the convex hull H of the transformed point as follows. The vertices of the convex hull correspond to the visible points. In the formula, ConvexHull represents the transformed point set q i Perform convex hull operation, that is, calculate the smallest convex polyhedron or convex polygon of a set of points, which contains all the transformed points. The vertices of the convex hull are the visible points. SB-1.3, Mark as visible points: Points p associated with these vertices in the original point cloud i Marked as visible points, as follows: P HPR ={p i 」q i ∈H} (7)。 4. The method for instance segmentation and model reconstruction based on deep learning and synthetic point cloud according to claim 3, characterized in that: In step S1, selecting key points to simulate point clouds under real scene occlusion includes the following sub-steps: SC-1.

1. Represent the room feasible region map as a grid map M∈RH×W, where H and W represent the room’s height and width, respectively. If the grid cell M(i,j) is in the feasible region, it is assigned a value of 1, otherwise it is assigned a value of 0. The grid resolution is defined by s, which represents the actual distance corresponding to each grid cell. SC-1.2, determine the candidate key point P cand is the center point of all feasible grid cells, that is, the grid cells that satisfy M(i,j)=1; SC-1.3, set the minimum distance constraint d min , to ensure that the key points are not too densely clustered, thereby achieving more uniform room coverage. Specifically: for any two selected key points p i and p j , the distance D(p i ,p j ) must satisfy D(p i ,p j )≥d min ; SC-1.4, the final selection step aims to ensure complete coverage of the feasible area. Each selected key point represents a specific coverage area. When the coverage area exceeds the preset threshold, the viewpoint selection stops. If the required coverage is not achieved, a greedy algorithm is introduced to iteratively select points that can maximize the uncovered area while keeping the minimum distance constraint unchanged.

5. The method for instance segmentation and model reconstruction based on deep learning and synthetic point cloud according to claim 3, characterized in that: In step S1, when synthesizing a point cloud scene under occlusion of a real scene, noise is simulated. The noise is simulated in the synthetic point cloud by a Gaussian function. The probability density function of the Gaussian random variable z is defined as follows: Where z is a Gaussian random variable, which represents the value of the noise; μ is the mean of the Gaussian distribution, which controls the center position of the noise; σ is the standard deviation of the Gaussian distribution, which controls the distribution width or dispersion of the noise; and PG(z) is the probability density function of Gaussian noise, which represents the probability of a specific noise value z occurring.

6. The method for instance segmentation and model reconstruction based on deep learning and synthetic point cloud according to claim 3, characterized in that: In step S2, the category labels include wall, table, pillar, air conditioner and electrical box.

7. The method for instance segmentation and model reconstruction based on deep learning and synthetic point cloud according to claim 3, characterized in that: In step S2, the point cloud component segmentation deep learning network is built based on oneformer3D.

8. The method for instance segmentation and model reconstruction based on deep learning and synthetic point cloud according to claim 1, characterized in that: In step S3, Dynamo is used to reconstruct the model of the point cloud after segmentation. First, the independent construction is reconstructed, and then the dependent construction is reconstructed.

9. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that: When the processor executes the computer program, the steps of the method according to any one of claims 1 to 8 are implemented.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the method according to any one of claims 1 to 8 are implemented.

Images