Vectorization rough planar graph segmentation method based on double-flow graph neural network
Through the construction of dual-stream graph neural network and dual graph, combined with vertex and edge properties, the problems of boundary aliasing and outlier fragments in plan diagram segmentation are solved, and the regular segmentation on vector graphics is realized, which improves segmentation performance and is suitable for industrial design .
Patent Information
- Application Number
- CN202411176201.7
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-08-26
- Publication Date
- 2024-11-29
AI Technical Summary
The prior art is susceptible to pixel noise and drawing style disturbances in plan segmentation, resulting in problems of room boundary aliasing prediction and outlier segments, and traditional methods rely on text and furniture to deal with rough floor plans.
Using the dual-stream graph neural network, by constructing the original graph and dual graph, using vertex and edge properties as input features, learning advanced semantic features, realizing semantic segmentation of vector plan diagrams, transforming them into polygonal region classification tasks, and designing a new modulation GAT layer to promote two Interaction between streams.
It realizes more regular and complete plan segmentation on vector graphics, improves segmentation performance, is suitable for intelligent industrial design and analysis, and avoids the defects of pixel-level processing.
Smart Images

Figure CN119048756A_ABST
Abstract
Description
Technical Field
[0001] The present invention belongs to the field of image segmentation, and in particular relates to a vectorized rough plane graph segmentation method based on a dual-flow graph neural network, which uses two flow graph neural networks to perform semantic segmentation on the vectorized rough plane graph. Background Art
[0002] In the field of image segmentation, semantic segmentation of rough floor plans into rooms with labeled types (e.g., bedroom, kitchen, etc.) is a fundamental task for downstream applications. Interior designers usually start by drawing a rough floor plan, including basic elements such as wall blocks and pipe barrels for real estate development. Afterwards, interior decoration, furniture layout, and 3D room space can be built and customized. In this process, it is important to obtain semantic segmentation of room space to meet the above requirements. While it is simple for humans to identify room layout from wall structure, automatic recognition with accurate semantics and clean boundaries is challenging. Typically, segmentation mainly deals with well-decorated floor plans in raster images, and is likely to produce confusing boundaries and outlier fragments in the segmented rooms, due to the fact that pixel-level segmentation ignores regular elements (e.g., line segments) in vector floor plans.
[0003] Floor plan understanding plays a vital role in various indoor applications, such as interior design, room space modeling, and real estate. Traditional methods treat floor plans as images and rely on low-level image processing and predefined rules to detect walls, doors, and windows. However, they are limited by hand-crafted features and cannot handle floor plan elements with different drawing styles, such as different shapes, lengths, or thicknesses. Recently, deep learning-based floor plan segmentation has dominated this field and can achieve superior performance. In recent years, some studies have adopted a fully convolutional network (FCN) to identify room types, boundaries, and indoor furniture; some have designed a deep multi-task network to learn room types and boundaries separately and model their spatial interactions through boundary-guided attention. The latest research adopts a more powerful semantic segmentation network and uses object detection to detect furniture and text as semantic priors to further enhance the segmentation performance. However, image-based room segmentation is still easily disturbed by pixel noise and drawing styles, resulting in aliased predictions near room boundaries and outlier fragments in segmented rooms.
[0004] The shortcomings of the existing technology are as follows: Traditional methods use powerful image-based segmentation networks on rasterized floor plans to predict room segmentation in a pixel-level manner. Since the pixel-by-pixel processing ignores the integrity of structural elements, their results often have jigsaw boundaries and fragmented semantic regions. In addition, these methods usually rely on text and furniture to determine semantic labels, which are not available in rough floor plans.
[0005] In recent years, vector graphics recognition has been widely used in 2D CAD design, urban design, graphic design, and circuit design to promote high-precision geometric modeling without resolution. It avoids the inaccuracy in rasterization-based methods, thereby achieving higher accuracy, faster reasoning time, and fewer training parameters, and has broad application prospects. Summary of the invention
[0006] In view of the deficiencies in the prior art, the present invention provides a vectorized rough plane graph segmentation method based on a dual-flow graph neural network.
[0007] The present invention adopts a graph neural network model to better consider the structural information in the graph data, and performs semantic segmentation on the 2D vector rough plane map when only rough plane map data is provided as input.
[0008] After obtaining the original graph and the dual graph, this method can embed vertex and edge attributes as input features into a two-stream graph neural network (two-stream GNN). The original stream predicts whether the line segments separate different room areas, while the two-stream classifies the segmented polygonal areas into different room types, thereby realizing that the semantic segmentation of the vector floor plan is reformulated as polygonal area classification.
[0009] A vectorized rough plane graph segmentation method based on a dual-flow graph neural network comprises the following steps:
[0010] Step 1: Based on the known rough plan, and considering the segmentation based on the spatial relationship between the wall structure and the regional layout, two vector plan graphs, namely the original graph and the dual graph, are constructed.
[0011] Step 2: Using the vertex and edge attributes of the original graph and the dual graph as input features, calculate the embedded features of the vertices and edges of the dual flow graph.
[0012] Step 3: After obtaining the embedded features of the two graphs, feed them into the two-stream graph neural network to learn high-level semantic features. The two-stream graph neural network (two-stream GNN) is used to achieve boundary line classification and polygon area classification.
[0013] Furthermore, the specific method of step one is as follows:
[0014] In order to learn segmentation based on the spatial relationship between wall structure and area layout, the rough plan is constructed into two vector plan graphs, the primal graph and the dual graph, and encoded. The primal graph is used to represent the input that captures the line segment structure:
[0015]
[0016] Where ν represents the set of line segment endpoints and ε represents the set of line segments. k∈ν has two-dimensional coordinates And connect the vertices ν i and ν j The edge is the original edge e ij ∈ε consists of a tuple Encoding, where Yes ij The angle between the x-axis and Indicates e ij Is it an extended line segment or the original line segment in the vector plane.
[0017] The partitioned polygonal room regions are captured using a dual graph, where regions are represented by vertices and their adjacencies are defined by edges. The dual graph is defined as Every vertex ν in the dual graph * ∈ν * Represents a polygonal area, each polygonal area corresponds to a subgraph Each subgraph comes from the original graph on the boundary of the region. Take the subgraph of the dual vertex v* The average value of the original vertex attributes (i.e., two-dimensional coordinates) is used as the attribute (i.e., two-dimensional coordinates) of the dual vertex. Each edge in the dual graph is the dual edge e ij *∈ε * Represents the adjacency of two polygonal regions through their two endpoints, i.e., the dual vertex ν i * and ν j * The connection of the attributes to obtain its attributes is (ν i *,ν j *).
[0018] There is also a correspondence between the primal edge and the dual edge. If ν i *,ν j *∈ν * There is a dual edge e ij *, then they correspond to the subgraphs of the polygonal regions and There are shared primitive segments On the other hand, without loss of generality, assuming that the original edge , and at the same time with the area ν i *,ν j * adjacent, then it has a corresponding dual edge, i.e. e ij *. In summary, the dualization operation *(·) on the edge is defined as:
[0019]
[0020] *(e)=e ij *,e∈*(e ij *);
[0021] in, Contains at least one original edge.
[0022] Furthermore, the specific method of step 2 is as follows:
[0023] Vertex and edge attributes are used as input features to calculate vertex and edge embedding features. At the same time, image features of raster images and geometric features of vector plane graphs are used for vertex embedding.
[0024] Embedded feature f of vertex ν of the original graph v The formula is as follows:
[0025] f v =x v +pe v ,v∈ν;
[0026] in (i.e., C-dimensional real vector space) represents the feature vector indexed by the coordinates of the feature vertices ν of the raster image obtained by rasterizing the rough plane image, It is the sinusoidal position encoding of the vertex position after rasterization.
[0027] Dual graph vertex ν * The embedding feature f v * It is used to capture the features of the polygonal area. Therefore, it is calculated as The average characteristics of the points sampled in the enclosed polygonal area. First, select To maintain the corner information. In addition, use The center of the triangle inside is opposite to the internal point Sampling is performed. The dual vertex embedding feature f v * Calculated as:
[0028]
[0029] Edge feature embedding is achieved by mapping the attributes of the original edge and the dual edge to the input features. Since the original edge e∈ε contains attributes other than the endpoint positions, a separate learnable projection matrix W e ,W θ , map the coordinates and additional attributes of the original edge to the C-dim feature, and get the embedded feature of the original edge :
[0030]
[0031] where || represents feature concatenation. For the dual edge ε * , calculate its embedding features based on the dual vertex coordinates:
[0032]
[0033] in is a learnable projection matrix.
[0034] Furthermore, the specific method of step three is as follows:
[0035] After obtaining the embedded features of the two graphs, they are fed into a two-stream graph neural network (two-stream GNN) to learn high-level semantic features.
[0036] The dual-stream graph neural network (dual-stream GNN) takes the embedded features of the original graph and the dual graph as input, and uses two parallel graph neural networks (GNNs) for processing. After 6 GAT modulation layers continuously update the vertex features and edge features, they are sent to the MLP layer to complete the classification task and probability prediction task. Each network has L = 6 modulated GAT layers.
[0037] For the lth, l∈Lth level modulated GAT layer, the dual-stream graph vertex embedding features From its adjacent vertices Accumulate features to get the next level vertex features The formula is as follows:
[0038]
[0039] where θ γ To update the MLP layer of vertex features, is the learnable projection matrix, α ij An adaptive weight to balance neighborhood features.
[0040] α ij The specific calculation formula is as follows:
[0041]
[0042] Where W q , is a learnable parameter, The dual edge features of the other flow are mapped to a weight matrix to adjust the similarity between two vertices. There is a learnable tensor Moreover, for e ii The degradation of where I is the identity matrix.
[0043] After the feature aggregation on the vertex, that is, updating the vertex features in the l-th level modulation GAT layer of the two streams, the edge features in the l-th level modulation GAT layer of the two streams are updated in the following way:
[0044]
[0045] where θ e Represents the MLP layer that updates edge features.
[0046] The output features of the two parallel graph neural networks (GNNs), namely the final edge features and vertex features, are used for the prediction task. The final edge features in the original graph g produced by the original stream GNN are fed into the binary classification head of the output MLP layer to predict whether the edges in the original graph g separate different room areas. * The final vertex features in are projected into the probability of vertices belonging to different room types in the output MLP layer, that is, the output of the MLP layer can be represented as the probability distribution of each vertex belonging to different room types.
[0047] The beneficial effects of the present invention are as follows:
[0048] In the present invention, a new semantic segmentation method on typical vector graphics is proposed, that is, directly processing vector rough floor plans and producing compact and regular segmentations. The present invention proposes a two-stream graph neural network that directly processes vector graphics and converts the problem into a dual task of room boundary classification and partition area classification. At the same time, a new modulated GAT layer is designed to achieve effective interaction between the two streams, thereby enhancing their mutual enhancement. The method provided by the present invention has excellent performance and can be used to produce more regular and complete floor plan segmentations. Considering the importance of vector graphics to industrial design, the present invention is expected to inspire future deep learning research based on vector graphics for intelligent industrial design and analysis. BRIEF DESCRIPTION OF THE DRAWINGS
[0049] Figure 1 The present invention is a flowchart of a method for segmenting a vectorized rough plane graph according to an embodiment of the present invention.
[0050] Figure 2 Qualitative comparison results with image-based planar segmentation methods on the dataset.
[0051] Figure 3 Visualization of the attention weights generated by the modulated GAT, where from left to right are the segmentation results, attention visualization and its zoom-in. DETAILED DESCRIPTION
[0052] The present invention will be described in detail below with reference to the accompanying drawings, and the purpose and effects of the present invention will become more apparent.
[0053] like Figure 1 As shown in FIG, a vectorized rough plane graph segmentation method based on a dual-flow graph neural network is shown in FIG. The steps are as follows:
[0054] Step 1: Based on the known rough plan, and considering the segmentation based on the spatial relationship between the wall structure and the regional layout, two vector plan graphs, namely the original graph and the dual graph, are constructed.
[0055] In order to learn segmentation based on the spatial relationship between wall structure and area layout, the rough plan is constructed into two vector plan graphs, the primal graph and the dual graph, and encoded. The primal graph is used to represent the input that captures the line segment structure:
[0056]
[0057] Where ν represents the set of line segment endpoints and ε represents the set of line segments. k ∈ν has two-dimensional coordinates And connect the vertices ν i and ν j The edge is the original edge e ij ∈ε consists of a tuple Encoding, where Yes ij The angle between the x-axis and Indicates e ij Is it an extended line segment or an original line segment in a vector plane? e is significant because the original line segments from the input are very likely to represent walls and room boundaries, while the expanded line segments are only one possible room boundary.
[0058] The partitioned polygonal room regions are captured using a dual graph, where regions are represented by vertices and their adjacencies are defined by edges. The dual graph is defined as Every vertex ν in the dual graph * ∈ν * Represents a polygonal area, each polygonal area corresponds to a subgraph Each subgraph comes from the original graph at the boundary of the region. For simplicity, we take the subgraph of the dual vertex v* The average value of the original vertex attributes (i.e., two-dimensional coordinates) is used as the attribute (i.e., two-dimensional coordinates) of the dual vertex. Each edge in the dual graph is the dual edge e ij *∈ε * Represents the adjacency of two polygonal regions through their two endpoints, i.e., the dual vertex ν i * and ν j * The connection of the attributes to obtain its attributes is (ν i *,ν j *).
[0059] There is also a correspondence between the primal edge and the dual edge. If ν i *,ν j *∈ν * There is a dual edge eij *, then they correspond to the subgraphs of the polygonal regions and There are shared primitive segments On the other hand, without loss of generality, assuming that the original edge , and at the same time with the area ν i *,ν j * adjacent, then it has a corresponding dual edge, i.e. e ij *. In summary, the dualization operation *(·) on the edge is defined as:
[0060]
[0061] *(e)=e ij *,e∈*(e ij *);
[0062] in, Contains at least one original edge. The set expression is retained because there may be extra vertices that split a straight wall into multiple original edges.
[0063] Step 2: Using the vertex and edge attributes of the original graph and the dual graph as input features, calculate the embedded features of the vertices and edges of the dual flow graph.
[0064] Vertex and edge attributes are used as input features to calculate vertex and edge embedding features. Although raster images (rasterized rough floor plans) and vector floor plans are equivalent when describing room layouts, it is found that using features obtained from raster images as additional input helps perceive a larger spatial receptive field of view and improves performance; therefore, both image features of raster images and geometric features of vector floor plans are used for vertex embedding.
[0065] Embedded feature f of vertex ν of the original graph v The formula is as follows:
[0066]
[0067] in (i.e., C-dimensional real vector space) represents the feature vector indexed by the coordinates of the feature vertices ν of the raster image obtained by rasterizing the rough plane image, It is the sinusoidal position encoding of the vertex position after rasterization.
[0068] Dual graph vertex ν * The embedding feature f v * It is used to capture the features of the polygonal area. Therefore, it is calculated as The average characteristics of the points sampled in the enclosed polygonal area. First, select To maintain the corner information. In addition, use The center of the triangle inside is opposite to the internal point Sampling is performed. The dual vertex embedding feature f v * Calculated as:
[0069]
[0070] Edge feature embedding is achieved by mapping the attributes of the original edge and the dual edge to the input features. Since the original edge e∈ε contains attributes other than the endpoint positions, a separate learnable projection matrix W e ,W θ , map the coordinates and additional attributes of the original edge to the C-dim feature, and get the embedded feature of the original edge
[0071]
[0072] where || represents feature concatenation. For the dual edge ε that mainly describes the adjacency of two vertices and lacks additional attributes * , simply calculate its embedding features based on the dual vertex coordinates:
[0073]
[0074] in is a learnable projection matrix.
[0075] Step 3: After obtaining the embedded features of the two graphs, feed them into the two-stream graph neural network to learn high-level semantic features. The two-stream graph neural network (two-stream GNN) is used to achieve boundary line classification and polygon area classification.
[0076] After obtaining the embedded features of the two graphs, they are fed into a two-stream graph neural network (two-stream GNN) to learn high-level semantic features. The two-stream GNN solves complementary tasks, namely boundary line classification and polygon region classification, solving the segmentation problem from two aspects, thereby enhancing each other.
[0077] For example, if the dual graph If the two vertices in have different signed labels, then they are in The shared line segments of the corresponding polygonal areas in the image should be the room boundary lines. If an edge in is predicted as a room boundary, the areas separated by this edge are likely to have different room types. To facilitate this edge-based modulation, we design a new modulation GAT layer that learns vertex features in one stream by considering the features of the edges in another stream.
[0078] The dual-stream graph neural network (dual-stream GNN) takes the embedded features of the original graph and the dual graph as input, and uses two parallel graph neural networks (GNN) for processing. After 6 GAT modulation layers continuously update the vertex features and edge features, they are sent to the MLP layer to complete the classification task and probability prediction task. Each network has L = 6 modulation GAT layers, such as Figure 1 As shown, since the two GNN streams have exactly the same architecture and operations, only one path shows the complete architecture and layer operations. For the lth, l∈Lth level modulation GAT layer, the dual-stream graph vertex embedding features From its adjacent vertices Accumulate features to get the next level vertex features The formula is as follows:
[0079]
[0080] where θ γ To update the MLP layer of vertex features, is the learnable projection matrix, α ij An adaptive weight to balance neighborhood features.
[0081] α ij The specific calculation formula is as follows:
[0082]
[0083] Where W q , is a learnable parameter, The dual edge features of the other flow are mapped to a weight matrix to adjust the similarity between two vertices. For simplicity, we implement There is a learnable tensor Moreover, for e ii In the degenerate case, we simply assume where I is the identity matrix.
[0084] After the feature aggregation on the vertex, that is, updating the vertex features in the l-th level modulation GAT layer of the two streams, the edge features in the l-th level modulation GAT layer of the two streams are updated in the following way:
[0085]
[0086] where θ e Represents the MLP layer that updates edge features.
[0087] The output features of the two parallel graph neural networks (GNNs), namely the final edge features and vertex features, are used for the prediction task. The original graph generated by the original stream GNN The final edge features in are fed into the binary classification head of the output MLP layer to predict the original graph The edges in separate different room areas. * The final vertex features in are projected into the probability of vertices belonging to different room types in the output MLP layer, that is, the output of the MLP layer can be represented as the probability distribution of each vertex belonging to different room types.
[0088] In the task of vectorized rough plane map segmentation, Figure 3 The qualitative comparison of image segmentation under different methods is shown. Since the experiment only focuses on the final regional segmentation results of the rough floor plan, the dataset image only retains the wall as input, rescales the long side of the input image to 256 and maintains its aspect ratio, and fills the short side with white to form a 256×256 input image for feature extraction. Although DeepLabV3+ and OCRNet expand the receptive field and integrate contextual features to improve performance, their outputs still have aliasing boundaries and inconsistent segmentation problems. DFPR integrates boundary information to assist room segmentation, but its performance on rough floor plans is greatly degraded because there is no internal texture or text symbols available for quick semantic reasoning. When zoomed in, the predicted boundaries show severe jagged patterns, which will bring additional burdens to post-processing and downstream applications. In contrast, these results further highlight the ability of the method of the present invention to generate smooth boundaries and consistent room semantic segmentation using regular line segments and regions to vectorize rough floor plans.
[0089] The above contents are further detailed descriptions of the present invention in combination with specific / preferred embodiments, and it cannot be determined that the specific implementation of the present invention is limited to these descriptions. For ordinary technicians in the technical field to which the present invention belongs, they can also make several substitutions or modifications to these described embodiments without departing from the concept of the present invention, and these substitutions or modifications should be regarded as belonging to the protection scope of the present invention.
[0090] Parts of the present invention that are not described in detail belong to the well-known technologies of those skilled in the art.
Claims
1. A vectorized rough plane graph segmentation method based on dual-flow graph neural network, characterized in that: The steps include: Step 1: Based on the known rough plan, two vector plan graphs, namely the original graph and the dual graph, are constructed by considering the segmentation based on the spatial relationship between the wall structure and the regional layout; Step 2: Using the vertex and edge attributes of the original graph and the dual graph as input features, calculate the embedded features of the vertices and edges of the dual flow graph; Step 3: After obtaining the embedded features of the two graphs, feed them into the dual-stream graph neural network to learn high-level semantic features; use the dual-stream graph neural network to realize boundary line classification and polygon area classification.
2. According to claim 1, a vectorized rough plane graph segmentation method based on dual-flow graph neural network is characterized in that: Step 1: In order to learn segmentation based on the spatial relationship between wall structure and area layout, the rough plan map is constructed into two vector plan maps, the primal map and the dual map, and encoded; the primal map is used to represent the input that captures the line segment structure: in represents the set of line segment endpoints, ε represents the set of line segments; each vertex in the original graph With two-dimensional coordinates And connect the vertices ν i and ν j The edge is the original edge e ij ∈ε consists of a tuple Encoding, where Yes ij The angle between the x-axis and Indicates e ij Is it an extended line segment or an original line segment in the vector plane? The partitioned polygonal room regions are captured using a dual graph, where regions are represented by vertices and their adjacencies are defined by edges; the dual graph is defined as Every vertex in the dual graph Represents a polygonal area, each polygonal area corresponds to a subgraph Each subgraph comes from the original graph of the region boundary; take the subgraph of the dual vertex v* The average value of the original vertex attributes is used as the attribute of the dual vertex; each edge in the dual graph is the dual edge e ij *∈ε * Represents the adjacency of two polygonal regions through their two endpoints, i.e., the dual vertex ν i * and ν j * The connection of the attributes to obtain its attributes is (ν i *,ν j *); There is also a correspondence between the primal and dual edges; if There is a dual edge e ij *, then they correspond to the subgraphs of the polygonal regions and There are shared primitive segments On the other hand, without loss of generality, assuming that the original edge and at the same time with the region ν i *,ν j * adjacent, then it has a corresponding dual edge, i.e. e ij *; In summary, the dualization operation *(·) on the edge is defined as: *(and)=and ij *,e∈*(e ij *); in, Contains at least one original edge.
3. The vectorized rough plane graph segmentation method based on dual-flow graph neural network according to claim 2 is characterized in that: The specific method of step 2 is as follows: Vertex and edge attributes are used as input features to calculate vertex and edge embedding features; at the same time, image features of raster images and geometric features of vector plane graphs are used for vertex embedding; Original graph vertices The embedding feature f v The formula is as follows: in The feature vector representing the feature vertex ν coordinate index of the raster image obtained by rasterizing the rough plane image, is the sinusoidal position encoding of the vertex position after rasterization; Dual graph vertex The embedding feature f v * It is used to capture the characteristics of the polygonal area; therefore, it is calculated as The average characteristics of the points sampled in the enclosed polygonal area; First, select ν v* To maintain the corner information; in addition, use The center of the triangle inside is opposite to the internal point I v* Sampling; dual vertex embedding feature f v * Calculated as: Edge feature embedding is achieved by mapping the attributes of the original edge and the dual edge to the input features; since the original edge e∈ε contains attributes other than the endpoint positions, a separate learnable projection matrix W e ,W θ , map the coordinates and additional attributes of the original edge to the C-dim feature, and get the embedded feature f of the original edge eij : where || represents feature concatenation; for the dual edge ε * , calculate its embedding features based on the dual vertex coordinates: in is a learnable projection matrix.
4. The vectorized rough plane graph segmentation method based on dual-flow graph neural network according to claim 3 is characterized in that: Step 3: After obtaining the embedded features of the two graphs, they are fed into a two-stream graph neural network to learn high-level semantic features; The dual-stream graph neural network takes the embedded features of the original graph and the dual graph as input, and uses two parallel graph neural networks GNN for processing. After 6 GAT modulation layers continuously update the vertex features and edge features, they are sent to the MLP layer to complete the classification task and probability prediction task. Each network has L = 6 modulation GAT layers; For the lth, l∈Lth level modulated GAT layer, the dual-stream graph vertex embedding features From its adjacent vertices Accumulate features to get the next level vertex features The formula is as follows: where θ γ To update the MLP layer of vertex features, is the learnable projection matrix, α ij Adaptive weights to balance neighborhood features; α ij The specific calculation formula is as follows: in is a learnable parameter, Map the dual edge features of the other flow to a weight matrix to adjust the similarity between two vertices; There is a learnable tensor Moreover, for e ii The degradation of Where I is the identity matrix; After the feature aggregation on the vertex, that is, updating the vertex features in the l-th level modulation GAT layer of the two streams, the edge features in the l-th level modulation GAT layer of the two streams are updated in the following way: where θ e MLP layer representing updated edge features; The output features of the two parallel graph neural networks GNN, namely the final edge features and vertex features, are used for prediction tasks; the original graph generated by the original stream GNN The final edge features in are fed into the binary classification head of the output MLP layer to predict the original graph Do the edges in separate different room areas? The dual graph from the dual flow GNN The final vertex features in are projected into the probability of vertices belonging to different room types in the output MLP layer, that is, the output of the MLP layer can be represented as the probability distribution of each vertex belonging to different room types.
Citation Information
Patent Citations
Semantic segmentation method based on double-flow feature fusion
CN110782462A
Planar graph segmentation method and device and electronic equipment
CN116824132A
Depth planar graph conversion based on multi-task network
CN117593622A
Semiconductor device
KR1020240141977A
Cited By
Building plane element identification method based on graph neural network and convolutional neural network
CN121305602A
A method for recognizing architectural planar elements based on graph neural networks and convolutional neural networks
CN121305602B
