Point cloud to parametric model generation method based on joint optimization of hop flow distillation and confidence

Abstract

The application provides a point cloud to parameterized model generation method based on joint optimization of hop flow distillation and confidence, relates to the field of three-dimensional point cloud processing and computer graphics three-dimensional modeling technology, obtains three-dimensional point cloud and component space adjacency relationship, divides three-dimensional point cloud to generate component point cloud unit, determines point cloud coverage confidence based on component point cloud unit and divides geometric fitting point cloud unit and generates inference point cloud unit; generate fitting three-dimensional shape parameters and generative three-dimensional shape parameters, jointly check two types of parameters to adjust the processing path of the component point cloud unit that fails to check, and update the three-dimensional shape generation network according to the component type; use the updated three-dimensional shape generation network to generate updated three-dimensional shape parameters, and generate a point cloud parameterized three-dimensional model through iterative checking and spatial connection. The method can take into account the differential modeling of observed sufficient components and insufficient components, and improve the efficiency and shape connection consistency of large-scale point cloud parameterized modeling.

Description

Technical Field

[0001] This application relates to the fields of 3D point cloud processing and 3D modeling technology in computer graphics, and more specifically, to a method for generating parameterized models from point clouds based on jump flow distillation and confidence optimization. Background Technology

[0002] With the application of laser scanning, image reconstruction, and 3D sensing equipment in existing building surveying, digital renovation, and spatial information reconstruction, the number of components contained in 3D point clouds continues to increase, and the differences between the completeness of point cloud observation, point cloud distribution, and component shape complexity are becoming more pronounced. How to form parametric 3D models that can characterize the position, size, and orientation of each component from large-scale 3D point clouds has become an important technical requirement in 3D point cloud modeling.

[0003] Existing parametric point cloud modeling schemes typically begin by identifying point sets corresponding to different components from a 3D point cloud. Then, pre-configured geometric templates are called according to component category to perform parameter fitting on each component's point set. For components with missing point clouds, neighbor interpolation, template replacement for similar components, or multi-round generation inference can be used to supplement the component's shape. After the parameter results for each component are generated, they are usually arranged according to a unified coordinate position, and a parametric 3D model is directly output.

[0004] However, the aforementioned technical solutions employ a relatively fixed modeling approach for the point cloud observation states of different components. When the component point cloud coverage is sufficient, the complex shape supplementation process increases unnecessary computation; when the component point cloud coverage is insufficient, relying solely on geometric template fitting is prone to significant shape deviations due to insufficient observation constraints. Using multi-round generation and inference to supplement component shapes also accumulates significant inference time as the number of components increases, making it difficult to adapt to batch parametric modeling of large-scale point clouds. Furthermore, the fitting or generation results for each component are usually generated separately, lacking a process for joint verification and feedback adjustment based on component spatial relationships. Even if a local component shape can match the point cloud individually, it may still experience connection misalignment or mismatched inclusion relationships with adjacent components. When a specific component type consistently exhibits generation deviations, existing generation models lack a technical path for adaptive updates specific to that component type. Therefore, existing technologies struggle to simultaneously address modeling quality, generation and inference efficiency, and shape continuity consistency between components under different point cloud coverage states. Summary of the Invention

[0005] This application provides a point cloud to parameterized model generation method based on jump flow distillation and confidence optimization, so as to at least alleviate the above-mentioned technical problems.

[0006] A method for generating parameterized models from point clouds based on joint optimization of skip flow distillation and confidence levels includes: Obtain the spatial adjacency relationship between the 3D point cloud and the component, segment the 3D point cloud to generate component point cloud units, determine the point cloud coverage confidence based on the component point cloud units, and divide the component point cloud units into geometric fitting point cloud units and generation inference point cloud units according to the point cloud coverage confidence. The fitting geometric fitting point cloud unit generates the fitting three-dimensional shape parameters; the teacher multi-step flow model generates the three-dimensional shape state transformation trajectory for the generated inference point cloud unit, and the jump flow distillation is performed based on the three-dimensional shape state transformation trajectory to generate the three-dimensional shape generation network, and the generative three-dimensional shape parameters are determined by the three-dimensional shape generation network. The shape continuity verification results are generated by jointly verifying the fitted 3D shape parameters and generative 3D shape parameters according to the spatial adjacency relationship of the components. The processing path of the point cloud unit of the failed component is adjusted according to the shape continuity verification results. The component type failure status is generated according to the component type, and the 3D shape generation network is updated according to the component type failure status. The updated 3D shape parameters of the point cloud units of the failed component are generated by the updated 3D shape generation network. The updated 3D shape parameters are iteratively verified to generate a set of verified 3D shape parameters. The set of verified 3D shape parameters is then continuously verified according to the spatial adjacency relationship of the components to generate a parameterized 3D model of the point cloud.

[0007] Optionally, the segmented 3D point cloud generating component point cloud unit includes: The semantic features of each point cloud point are extracted from the 3D point cloud using a component semantic segmentation network, and the spatial adjacency relationship of the point cloud is determined based on the spatial position of each point cloud point in the 3D point cloud. Point cloud points with the same semantic features and spatial adjacency are aggregated according to the spatial adjacency relationship of the point cloud to generate component point cloud units; The component type confidence score is generated from the semantic features of the point cloud contained in the component point cloud unit, and the component type corresponding to the component point cloud unit is determined based on the component type confidence score; When the component type confidence score cannot uniquely determine the component type corresponding to the component point cloud unit, a set of candidate component types is generated based on the component type confidence score. The set of candidate component types is then associated with the corresponding component point cloud unit to determine the component type corresponding to the component point cloud unit.

[0008] Optionally, determining the point cloud coverage confidence based on component point cloud units includes: Read the component shape prior based on the component type corresponding to the component point cloud unit; Compare the component point cloud units with the component shape priors to generate the point cloud surface coverage state, point cloud distribution state, and shape feature support state; The point cloud surface coverage state, point cloud distribution state, and shape feature support state are integrated to generate a point cloud coverage state, and then the point cloud coverage confidence degree corresponding to the component point cloud unit is determined based on the point cloud coverage state.

[0009] Optionally, dividing the component point cloud units into geometrically fitted point cloud units and generative inference point cloud units according to point cloud coverage confidence includes: Obtain the 3D shape reconstruction accuracy requirements configured for the parameterized 3D model of the point cloud to be generated, and generate the point cloud coverage confidence threshold based on the component type corresponding to the component point cloud unit and the 3D shape reconstruction accuracy requirements; The confidence score of the point cloud coverage corresponding to the point cloud unit of the component is compared with the confidence score threshold of the point cloud coverage to generate a confidence score comparison result. Based on the confidence level comparison results, component point cloud units whose point cloud coverage confidence level reaches the point cloud coverage confidence level threshold are identified as geometrically fitted point cloud units, while component point cloud units whose point cloud coverage confidence level does not reach the point cloud coverage confidence level threshold are identified as generated inference point cloud units.

[0010] Optionally, the fitted geometric fitting point cloud unit generates fitted three-dimensional shape parameters, including: The component shape primitives and shape fitting tolerances are determined based on the component type corresponding to the geometrically fitted point cloud units and the accuracy requirements for 3D shape reconstruction. The component shape primitives are fitted to the geometric fitting point cloud units to generate shape primitive parameters, and then the fitted three-dimensional shape parameters are determined based on the shape primitive parameters. Calculate the shape fitting residual between the fitted 3D shape parameters and the corresponding geometric fitted point cloud units, and compare the shape fitting residual with the shape fitting tolerance to generate the fitting residual comparison result. Based on the fitting residual comparison results, geometrically fitted point cloud units whose shape fitting residuals exceed the shape fitting tolerance are redefined as generative inference point cloud units, so that the redefined generative inference point cloud units can participate in the determination of generative 3D shape parameters.

[0011] Optionally, the step of generating a three-dimensional shape state transformation trajectory for the generated inference point cloud unit by the teacher multi-step flow model includes: Encode and generate inference point cloud units to generate point cloud condition features, and read the component shape prior based on the component type corresponding to the generated inference point cloud units; By fusing point cloud conditional features and component shape priors, three-dimensional shape generation conditions are generated. Generate the initial state of the three-dimensional shape, and input the initial state of the three-dimensional shape and the conditions for generating the three-dimensional shape into the teacher's multi-step flow model so that the teacher's multi-step flow model can gradually update the initial state of the three-dimensional shape and generate multiple three-dimensional shape states. Multiple 3D shape states are arranged according to their update order to generate the 3D shape state transformation trajectory corresponding to the generated inference point cloud unit.

[0012] Optionally, the stepwise updating of the initial state of the 3D shape and generation of multiple 3D shape states by the teacher's multi-step flow model includes: For the same point cloud unit, multiple initial states of three-dimensional shapes are generated. The teacher multi-step flow model is used to perform state transformation according to each initial state of three-dimensional shape and the three-dimensional shape generation conditions to generate candidate three-dimensional shape state transformation trajectories. The termination state of each candidate 3D shape state transformation trajectory is converted into candidate 3D shape parameters, and the candidate 3D surface point cloud is reconstructed based on the candidate 3D shape parameters. The point cloud coordinate domain is determined based on the 3D point cloud, and each candidate 3D surface point cloud is mapped to the point cloud coordinate domain in order to calculate the point cloud reprojection residual between each candidate 3D surface point cloud and the corresponding generated inference point cloud unit. Candidate 3D shape state transformation trajectories are selected based on the point cloud reprojection residuals, and then the candidate 3D shape state transformation trajectory with the smallest point cloud reprojection residual is determined as the 3D shape state transformation trajectory corresponding to the generated inferred point cloud unit.

[0013] Optionally, the step of performing jump flow distillation based on the three-dimensional shape state transformation trajectory to generate a three-dimensional shape generation network includes: Compare adjacent three-dimensional shape states in the three-dimensional shape state transformation trajectory to generate trajectory change states; The shape parameter dimension is determined based on the component type corresponding to the generated inferred point cloud unit, and the component shape change degree is determined based on the component shape prior. The shape parameter dimension and the component shape change degree are fused to generate the component type complexity. The trajectory of three-dimensional shape state transformation is divided according to the trajectory change state and component type complexity to generate jump flow intervals; The jump start state and jump end state are determined based on the jump flow interval. The jump start state and jump end state are associated to generate jump flow distillation supervision data, which is then used to participate in the jump flow distillation training of the 3D shape generation network.

[0014] Optionally, the jump flow distillation supervision data participates in the jump flow distillation training of the 3D shape generation network, including: An initial 3D shape generation network is constructed, which combines the jump start state, point cloud condition features, and the component type corresponding to the generated inferred point cloud unit to generate the jump flow distillation input, and the jump termination state is determined as the jump flow distillation target. The jump flow distillation input is fed into the initial 3D shape generation network to generate jump state prediction results. The jump state prediction results are compared with the jump flow distillation target to generate shape reconstruction error. The jump state prediction results corresponding to adjacent jump flow intervals are continued according to the state transformation order of the three-dimensional shape state transformation trajectory to generate cross-interval prediction trajectory. Then, the cross-interval prediction trajectory and the three-dimensional shape state transformation trajectory are compared to generate flow matching consistency error. The shape reconstruction error and flow matching consistency error are combined to generate a skip flow distillation loss. The initial 3D shape generation network is updated based on the skip flow distillation loss to generate a 3D shape generation network.

[0015] Optionally, the initial 3D shape generation network includes a shared shape decoding backbone and a component type-specific determination head corresponding to each component type. The 3D shape generation network generated after updating the initial 3D shape generation network according to the jump flow distillation loss retains the shared shape decoding backbone and the component type-specific determination head. The process of determining generative 3D shape parameters by the 3D shape generation network includes: using a shared shape decoding backbone to perform shape decoding based on point cloud conditional features to generate a shared 3D shape representation, and determining a target component type-specific determination head based on the component type corresponding to the generated inferred point cloud unit. The shared 3D shape representation is input into a target component type-specific determination head to generate generative 3D shape parameters corresponding to the generated inferred point cloud units.

[0016] Optionally, the step of jointly verifying the fitted three-dimensional shape parameters and generative three-dimensional shape parameters according to the spatial adjacency relationship of components to generate shape continuity verification results includes: Based on the spatial adjacency relationship of components, the component pairs to be verified are determined from the components corresponding to the fitted three-dimensional shape parameters and the generative three-dimensional shape parameters, and then the component connection type corresponding to the component pairs to be verified is determined. Extract shape continuity features from the fitted three-dimensional shape parameters or generative three-dimensional shape parameters of the component to be verified on both sides respectively; Based on the component connection type, determine the shape connection comparison rules and shape connection tolerance, and compare the shape connection characteristics of the component to be verified on both sides according to the shape connection comparison rules to generate shape connection residuals; The shape continuity residual is compared with the shape continuity tolerance to generate the shape continuity status. Then, the shape continuity status of each component to be verified is summarized to generate the shape continuity verification result.

[0017] Optionally, adjusting the processing path of the point cloud unit of the failed component based on the shape continuity verification result includes: Based on the shape continuity verification results, identify the fitted 3D shape parameters that failed the shape continuity verification, and then trace the component point cloud units corresponding to the fitted 3D shape parameters that failed the shape continuity verification to generate component point cloud units that failed the fitting verification. The processing path for the point cloud units of the component that failed the fitting verification was changed from the geometric fitting path to the inference generation path, so as to generate the fitting branch adjustment results. Based on the shape continuity verification results, identify the generative 3D shape parameters that fail the shape continuity verification, and then trace the component point cloud units corresponding to the generative 3D shape parameters that fail the shape continuity verification to generate the component point cloud units that fail the verification. The generation inference path corresponding to the point cloud unit of the failed component is maintained to generate the generation branch hold result. The fitting branch adjustment result and the generation branch hold result are combined to generate the processing path adjustment result, so that the processing path adjustment result participates in the generation of updating three-dimensional shape parameters.

[0018] Optionally, associating the candidate component type set with the corresponding component point cloud unit to determine the component type corresponding to the component point cloud unit includes: The corresponding component point cloud units are fitted according to each candidate component type in the candidate component type set to generate candidate fitted 3D shape parameters. The 3D shape generation network is then used to process the corresponding component point cloud units according to each candidate component type to generate candidate generative 3D shape parameters. Combine candidate fitted 3D shape parameters and candidate generative 3D shape parameters to generate a set of candidate 3D shape parameters; Based on each candidate 3D shape parameter in the candidate 3D shape parameter set, the candidate 3D surface point cloud is reconstructed. The candidate 3D surface point cloud is compared with the corresponding component point cloud unit to generate the consistency of candidate point cloud reconstruction. The candidate 3D shape parameters are verified according to the spatial adjacency relationship of the component to generate the consistency of candidate shape continuity. Combining the consistency of candidate point cloud reconstruction and the consistency of candidate shape continuity, each candidate 3D shape parameter in the candidate 3D shape parameter set is filtered, and the filtered candidate 3D shape parameters are determined as category-confirmed 3D shape parameters. The category-confirmed component type is determined according to the candidate component type corresponding to the category-confirmed 3D shape parameter, and then the component type corresponding to the component point cloud unit is updated using the category-confirmed component type.

[0019] Optionally, the step of generating component type failure states according to component type and updating the 3D shape generation network based on component type failure states includes: Based on the shape continuity verification results, identify the generative 3D shape parameters that fail the shape continuity verification, and then trace the generative inference point cloud units corresponding to the generative 3D shape parameters that fail the shape continuity verification to generate point cloud units of the generative verification failed components. The frequency of verification failures of point cloud units of failed components is generated according to component type, so as to generate component type failure status, and the active distillation component type is determined based on the component type failure status. Extract historical point cloud units of the same type as the active distillation component from a pre-built historical point cloud sample library, and combine the point cloud units of the active distillation component type that failed to generate verification with the historical point cloud units of the same type to generate active distillation samples. The teacher multi-step flow model is used to perform state transformations based on active distillation samples to generate supplementary 3D shape state transformation trajectories. The component type-specific determination head corresponding to the active distillation component type is updated based on the supplementary 3D shape state transformation trajectory to generate an updated 3D shape generation network.

[0020] Optionally, the step of generating updated 3D shape parameters for the failed component point cloud units from the updated 3D shape generation network, iteratively verifying the updated 3D shape parameters to generate a verified 3D shape parameter set, and then continuously verifying the verified 3D shape parameter set according to the spatial adjacency relationship of the components to generate a point cloud parameterized 3D model includes: Based on the processing path adjustment results, the fitting verification failed component point cloud units and the generation verification failed component point cloud units are determined. The fitting verification failed component point cloud units and the generation verification failed component point cloud units are input into the updated 3D shape generation network to generate updated 3D shape parameters. The fitted 3D shape parameters corresponding to the point cloud units of the failed fitting verification component and the generative 3D shape parameters corresponding to the point cloud units of the failed generation verification component are replaced with updated 3D shape parameters to generate an updated set of 3D shape parameters to be verified. Repeat the shape continuity check on the updated set of three-dimensional shape parameters to be checked until the updated set of three-dimensional shape parameters to be checked meets the check termination condition, so as to generate a set of three-dimensional shape parameters to be checked and confirmed. Based on the 3D point cloud, the point cloud coordinate domain is determined, and the verified 3D shape parameter set is transformed into the point cloud coordinate domain to generate a coordinate-unified 3D shape parameter set. The coordinate-unified 3D shape parameter set is then continued according to the spatial adjacency relationship of the components to generate a point cloud parameterized 3D model.

[0021] Technical advantages of the technical solution provided in this application The point cloud to parametric model generation method provided in this application addresses the technical shortcomings of existing solutions that use relatively fixed modeling methods for components with different point cloud observation states. This leads to unnecessary computational burdens on components with sufficient observation, while components with insufficient observation struggle to obtain suitable shape parameters. The method first segments the 3D point cloud to generate component point cloud units, then determines the point cloud coverage confidence based on these units, and further divides them into geometric fitting point cloud units and generative inference point cloud units according to their coverage confidence. Thus, components with different point cloud coverage states enter processing paths adapted to their observation conditions. Geometric fitting point cloud units form fitted 3D shape parameters through fitting, while generative inference point cloud units form generative 3D shape parameters through a 3D shape generation network. Compared to uniformly applying geometric fitting or generative inference to all components, this differentiated processing reduces the need for complex generative calculations on components with sufficient observation and mitigates shape deviations caused by directly performing geometric fitting on components with insufficient observation due to a lack of point cloud constraints.

[0022] To address the technical drawback of accumulated inference time during multi-round generative inference modeling of large-scale components, this application proposes a method where a teacher multi-step flow model generates 3D shape state transformation trajectories for generative inference point cloud units. Then, skip-flow distillation is performed based on these 3D shape state transformation trajectories to generate a 3D shape generation network. The 3D shape state transformation trajectories contain state evolution information from which the teacher multi-step flow model progressively forms 3D shape parameters. Skip-flow distillation uses this state evolution information to train the generative capability of the 3D shape generation network, enabling it to determine generative 3D shape parameters based on the generative inference point cloud units. Compared to performing complete multi-step state transformations of the teacher multi-step flow model in the actual modeling process of each component, transferring multi-step state transformation capabilities to the 3D shape generation network reduces the dependence of the generative inference stage on continuous multi-round computation, thereby shortening the overall processing time required for a large number of generative inference point cloud units to sequentially generate 3D shape parameters.

[0023] To address the shortcomings of existing solutions that generate shape parameters for each component separately, making it difficult to promptly detect misalignments or mismatches in the inclusion relationships of adjacent components, this application uses a joint verification method based on the spatial adjacency relationship of components to fit 3D shape parameters and generative 3D shape parameters, thereby generating shape continuity verification results. The spatial adjacency relationship of components means that the 3D shape parameters of each component are no longer evaluated solely based on the matching state between a single component and its point cloud, but are further constrained by the shape continuity relationship between adjacent components. Adjusting the processing path of the point cloud units of failed component verifications based on the shape continuity verification results allows for the selection of a new processing path and generation of new 3D shape parameters when the 3D shape parameters generated by the original processing path do not meet the component continuity requirements. This reduces the situation where individual component parameters are valid but cannot be reasonably connected in the overall model, improving the consistency of shape continuity between components in the point cloud parameterized 3D model.

[0024] To address the technical shortcomings of existing generative models that remain fixed and struggle to adaptively correct for persistent generation deviations in specific component types, this application generates component type failure states based on component type and updates the 3D shape generation network accordingly. The component type failure states associate generation failures detected during shape continuity verification with corresponding component types, enabling the 3D shape generation network to focus on component types that consistently fail verification, rather than indiscriminately updating all component types. The updated 3D shape parameters of the point cloud units of the failed component types generated by the updated 3D shape generation network correspond to the adjustments in generation capabilities and the component type deviations exposed during actual verification, thereby improving the targeting of model updates and reducing the possibility of unrelated component types being affected by the update process.

[0025] Based on this, this application iteratively verifies and updates the 3D shape parameters to generate a verified 3D shape parameter set, and then verifies and confirms the 3D shape parameter set according to the spatial adjacency relationship of the components to generate a point cloud parameterized 3D model. Updating the 3D shape parameters requires re-entering the joint verification, so that the results of processing path adjustment and 3D shape generation network update are both constrained by the spatial adjacency relationship of the components again, avoiding the update results from directly entering the final model without overall continuity verification. This forms a closed-loop technical process with point cloud coverage confidence branch, fitting and generation, shape continuity verification, processing path adjustment, network update and iterative verification in sequence, which can take into account the processing efficiency, local component shape recovery ability and overall component continuity status in large-scale point cloud parameterized modeling. Attached Figure Description

[0026] Figure 1 This application provides an embodiment of a point cloud to parameterized model generation method based on joint optimization of jump flow distillation and confidence level. Figure 2 This application provides an embodiment of a point cloud to parameterized model generation device based on joint optimization of jump flow distillation and confidence level. Figure 3 An electronic device is described in an embodiment of this application; Figure 4 This application provides an embodiment of a computer-readable storage medium. Figure 5 This application provides a point cloud to parameterized model generation system based on joint optimization of jump flow distillation and confidence level. Detailed Implementation

[0027] To make the technical solution and its implementation process clearer, the following description, in conjunction with the accompanying drawings and embodiments, further illustrates a point cloud to parametric model generation method based on joint optimization of jump flow distillation and confidence level. The following embodiments use laser-scanned point clouds of existing buildings as examples. The three-dimensional point cloud includes point cloud points corresponding to walls, beams, columns, doors, windows, and decorative components. The three-dimensional point cloud can also originate from image reconstruction, structured light scanning, or other acquisition methods capable of generating three-dimensional spatial sampling points. The specific number of components, iterations, thresholds, and network layers used in each embodiment are used to illustrate the implementation process and do not constitute a limitation on the scope of protection of this application.

[0028] In one embodiment, each point in the 3D point cloud includes at least 3D spatial coordinates. Points may also include one or more of the following: reflection intensity, color, normal vector, acquisition time, or acquisition viewpoint. Component spatial adjacency relationships are used to characterize the connection, adjacency, or containment relationships between different components in 3D space, such as end connections between adjacent walls, containment relationships between doors / windows and their respective walls, and connection relationships between beams and columns. Component spatial adjacency relationships can be read from the component index data accompanying the 3D point cloud, or they can be determined after the component point cloud units are generated, based on the spatial distance, boundary proximity, and spatial containment state between different component point cloud units. The determined component spatial adjacency relationships are then associated with the corresponding component point cloud units for subsequent shape continuity verification and point cloud parametric 3D model generation.

[0029] like Figure 1 As shown, this embodiment of the present application provides a method for generating point clouds into parameterized models based on joint optimization of skip flow distillation and confidence levels, comprising: Obtain the spatial adjacency relationship between the 3D point cloud and the component, segment the 3D point cloud to generate component point cloud units, determine the point cloud coverage confidence based on the component point cloud units, and divide the component point cloud units into geometric fitting point cloud units and generation inference point cloud units according to the point cloud coverage confidence. The fitting geometric fitting point cloud unit generates the fitting three-dimensional shape parameters; the teacher multi-step flow model generates the three-dimensional shape state transformation trajectory for the generated inference point cloud unit, and the jump flow distillation is performed based on the three-dimensional shape state transformation trajectory to generate the three-dimensional shape generation network, and the generative three-dimensional shape parameters are determined by the three-dimensional shape generation network. The shape continuity verification results are generated by jointly verifying the fitted 3D shape parameters and generative 3D shape parameters according to the spatial adjacency relationship of the components. The processing path of the point cloud unit of the failed component is adjusted according to the shape continuity verification results. The component type failure status is generated according to the component type, and the 3D shape generation network is updated according to the component type failure status. The updated 3D shape parameters of the point cloud units of the failed component are generated by the updated 3D shape generation network. The updated 3D shape parameters are iteratively verified to generate a set of verified 3D shape parameters. The set of verified 3D shape parameters is then continuously verified according to the spatial adjacency relationship of the components to generate a parameterized 3D model of the point cloud.

[0030] Specifically, in this application, the component point cloud unit is component-level point cloud data formed by aggregating point cloud points corresponding to the same component instance. Point cloud coverage confidence is used to represent the sufficiency of the observation of the corresponding component surface by the component point cloud unit. Geometric fitting point cloud units correspond to component point cloud units with relatively sufficient observation coverage that can provide more spatial constraints for fitting component shape primitives; generating inference point cloud units correspond to component point cloud units with relatively insufficient observation coverage or shape fitting residuals exceeding the shape fitting tolerance. By first determining the point cloud coverage confidence and then configuring different processing paths according to the point cloud coverage confidence, the same three-dimensional shape parameter determination method can be avoided for all component point cloud units.

[0031] Specifically, in this application, both fitted 3D shape parameters and generative 3D shape parameters are used to characterize the parametric shape of the component in 3D space. The fitted 3D shape parameters are formed through a fitting process between component shape primitives and geometrically fitted point cloud units; the generative 3D shape parameters are formed by a 3D shape generation network based on generated and inferred point cloud units. The fitted and generative 3D shape parameters may include component position, component size, and component orientation. For different component types, the fitted and generative 3D shape parameters may also include parameters adapted to the component shape, such as the thickness and height of walls, the cross-sectional dimensions and axial direction of columns, and the opening width and height of doors and windows.

[0032] Specifically, in this application, the shape continuity verification result not only reflects the relationship between the fitted 3D shape parameters or generative 3D shape parameters of a single component and the corresponding component point cloud unit, but also reflects the spatial matching status of the fitted 3D shape parameters or generative 3D shape parameters of adjacent components at the connection position. Based on the shape continuity verification result, the processing path of the failed component point cloud unit is adjusted so that the fitted 3D shape parameters generated by the original geometric fitting path are transferred to the generation inference path when the shape continuity requirements are not met. The component type failure status formed according to component type is further used to determine the component types that need to be updated, so that the update range of the 3D shape generation network corresponds to the component types that actually exhibit generation deviations.

[0033] Optionally, the segmented 3D point cloud generating component point cloud unit includes: The semantic features of each point cloud point are extracted from the 3D point cloud using a component semantic segmentation network, and the spatial adjacency relationship of the point cloud is determined based on the spatial position of each point cloud point in the 3D point cloud. Point cloud points with the same semantic features and spatial adjacency are aggregated according to the spatial adjacency relationship of the point cloud to generate component point cloud units; The component type confidence score is generated from the semantic features of the point cloud contained in the component point cloud unit, and the component type corresponding to the component point cloud unit is determined based on the component type confidence score; When the component type confidence score cannot uniquely determine the component type corresponding to the component point cloud unit, a set of candidate component types is generated based on the component type confidence score. The set of candidate component types is then associated with the corresponding component point cloud unit to determine the component type corresponding to the component point cloud unit.

[0034] Specifically, in this application, the component semantic segmentation network is pre-trained based on component point cloud samples labeled with component types. The component point cloud samples are formed by aggregating component point clouds and their corresponding component type labels from historical projects. Each sample point in the component point cloud sample is associated with a corresponding component type label, which can include walls, beams, columns, doors, windows, and decorative components. After receiving the 3D point cloud, the component semantic segmentation network extracts local geometric information step-by-step according to the spatial neighborhood between point cloud points, and combines the local geometric information at different spatial scales into point cloud semantic features corresponding to each point cloud point. These point cloud semantic features participate in point cloud aggregation and component type confidence generation, thereby establishing a correspondence between component point cloud units and component types.

[0035] In one embodiment, the component semantic segmentation network includes a point cloud input mapping layer, a local neighborhood aggregation layer, a spatial scale transformation layer, a context encoding layer, a semantic recovery layer, and a point-level classification layer. The point cloud input mapping layer converts the 3D spatial coordinates of point cloud points and associated additional information into an initial point cloud representation, which serves as the input to the local neighborhood aggregation layer. The local neighborhood aggregation layer determines neighboring point cloud points based on the spatial distance between point cloud points and aggregates the relative positions of neighboring point cloud points with the initial point cloud representation to generate a local geometric representation of the point cloud. The spatial scale transformation layer performs hierarchical sampling of the local geometric representation of the point cloud to generate point cloud scale representations corresponding to different spatial scales, enabling the component semantic segmentation network to represent component edges, local surfaces, and the overall shape of the component separately. The context encoding layer fuses point cloud scale representations corresponding to different spatial scales to generate a point cloud context representation. The semantic recovery layer maps the point cloud context representation back to the original point cloud point positions to recover the point cloud semantic features corresponding to each point cloud point. The point-level classification layer generates classification probabilities corresponding to each component type based on the point cloud semantic features; these classification probabilities participate in the generation of component type confidence.

[0036] Point cloud spatial adjacency indicates whether points in a 3D point cloud belong to the same continuous region in space. When determining point cloud spatial adjacency based on the spatial position of each point in the 3D point cloud, the spatial distance between each point and its neighboring points can be calculated, and points whose spatial distance is not greater than an adjacency distance threshold are associated as neighboring points. The adjacency distance threshold is configured based on the average point spacing of the 3D point cloud, ensuring that 3D point clouds with different acquisition densities adopt an adjacency range adapted to their point cloud density. The configured adjacency distance threshold participates in the determination of point cloud spatial adjacency, and the point cloud spatial adjacency further participates in the generation of component point cloud units.

[0037] When aggregating point cloud points with the same semantic features and spatial adjacency based on point cloud spatial adjacency, the process first assigns candidate component types to each point cloud point based on the classification probability output by the point-level classification layer. Then, starting from any point cloud point, the process expands along the point cloud spatial adjacency relationship, grouping consecutively adjacent point cloud points with the same candidate component type into the same component point cloud unit. For point cloud points with the same candidate component type but spatially separated, different component point cloud units are formed respectively, thereby avoiding merging two spatially different components of the same type into the same component point cloud unit.

[0038] The component type confidence score is determined based on the semantic features of the point cloud contained in the component point cloud unit. In one embodiment, the classification probabilities of each point cloud point within the component point cloud unit for the same candidate component type are first summarized. Then, the component type confidence score corresponding to the candidate component type is generated based on the average of the classification probabilities and the consistency of the classification results within the component point cloud unit. The higher the component type confidence score, the more concentrated the judgments of each point cloud point within the component point cloud unit are for the corresponding candidate component type. The component type confidence scores corresponding to each candidate component type are used to determine the component type corresponding to the component point cloud unit or to generate a set of candidate component types.

[0039] When the difference between the highest component type confidence score and the confidence scores of other component types reaches the type confirmation threshold, the candidate component type corresponding to the highest component type confidence score is determined as the component type corresponding to the component point cloud unit. The type confirmation threshold is pre-configured based on the classification distribution of the component semantic segmentation network in the validation data. The configured type confirmation threshold is used to determine whether the component type confidence score can uniquely determine the component type corresponding to the component point cloud unit. When the difference between the confidence scores of multiple component types does not reach the type confirmation threshold, candidate component types with low confidence scores but still having shape interpretation capabilities are not directly discarded. Instead, these candidate component types are grouped into a candidate component type set, and the candidate component type set is associated with the corresponding component point cloud unit. For example, when a component point cloud unit in an old building has partial geometric representations of both wall piers and columns in spatial morphology, both wall piers and columns can be retained in the candidate component type set. Subsequently, candidate 3D shape parameters corresponding to each candidate component type are formed, and the component type corresponding to the component point cloud unit is determined by combining the consistency of candidate point cloud reconstruction and the consistency of candidate shape continuity.

[0040] In an exemplary embodiment, the aforementioned component-level segmentation is performed on a 3D point cloud of an existing industrial building. The 3D point cloud includes thousands of wall, beam, column, door, and window components. The component semantic segmentation network first determines the semantic features of each point cloud point, and then forms multiple component point cloud units according to the spatial adjacency relationship of the point cloud. For component point cloud units whose shape is between load-bearing wall piers and decorative pilasters, the load-bearing wall piers and decorative pilasters are associated as a set of candidate component types, so that candidate 3D shape parameters can be generated separately and the component type corresponding to the component point cloud unit can be determined.

[0041] Optionally, determining the point cloud coverage confidence based on component point cloud units includes: Read the component shape prior based on the component type corresponding to the component point cloud unit; Compare the component point cloud units with the component shape priors to generate the point cloud surface coverage state, point cloud distribution state, and shape feature support state; The point cloud surface coverage state, point cloud distribution state, and shape feature support state are integrated to generate a point cloud coverage state, and then the point cloud coverage confidence degree corresponding to the component point cloud unit is determined based on the point cloud coverage state.

[0042] Specifically, in this application, the component shape prior is formed by summarizing historical component shape data of the same component type and stored in a component shape prior library. The component shape prior does not limit the component to having fixed dimensions, but rather records the surface composition, shape parameter types, component shape prior parameter ranges, and key shape locations typically present in the corresponding component type. For example, the component shape prior for a wall may include two relatively parallel main surfaces, the wall end boundary, and the thickness direction; the component shape prior for a column may include the column axis, the column side surface, and the column cross-section; the component shape prior for doors and windows may include the rectangular opening boundary and the plane where the opening is located. The component shape prior participates in determining the expected surface of the component and, through the expected surface of the component, participates in the generation of the point cloud surface coverage state, point cloud distribution state, and shape feature support state.

[0043] Based on the component type corresponding to the component point cloud unit, the component shape prior is retrieved from the component shape prior library and then transformed to the component's local coordinate domain where the component point cloud unit is located. The component's local coordinate domain is determined based on the spatial center, principal direction, and point cloud extent of the component point cloud unit. After the component shape prior is transformed to the component's local coordinate domain, the component's expected surface is formed. The component's expected surface represents the complete surface extent corresponding to the corresponding component type at the current spatial location of the component point cloud unit. The component's expected surface participates in the generation of the point cloud surface coverage state, point cloud distribution state, and shape feature support state.

[0044] The point cloud surface coverage state represents the proportion of the component's expected surface that is actually observed by the component's point cloud units. To determine the point cloud surface coverage state, the point cloud points in the component's point cloud units are projected onto the component's expected surface, and the expected surface is divided into multiple surface partitions according to a preset surface partitioning scale. The preset surface partitioning scale is configured based on the average point spacing of the 3D point cloud, and the configured preset surface partitioning scale is used to ensure that each surface partition has a spatial range adapted to the 3D point cloud sampling density. The number of surface partitions with projected points is counted, and the proportion of surface partitions with projected points to the total number of surface partitions is calculated to generate the point cloud surface coverage state. For components obscured by equipment, pipelines, or adjacent components, there are more surface partitions on the component's expected surface that do not form projected points, and the point cloud surface coverage state is correspondingly reduced.

[0045] Point cloud distribution status represents the uniformity of the distribution of observed point cloud points on the expected surface of a component. To determine the point cloud distribution status, the number of point cloud points in each surface partition is counted, and the distribution status is determined based on the differences in the number of point cloud points in each surface partition. When point cloud points are concentrated in a local area of ​​the expected surface of the component, even if the total number of point cloud points is large, the point cloud distribution status still reflects a low degree of uniformity. When point cloud points are distributed across multiple surface partitions of the expected surface of the component, the point cloud distribution status reflects a high degree of uniformity. The point cloud distribution status and the point cloud surface coverage status jointly contribute to the generation of the point cloud coverage status, thereby reducing the bias caused by judging the sufficiency of observation solely based on the number of point cloud points.

[0046] The shape feature support state indicates whether a component point cloud unit contains the key shape features required to determine the 3D shape parameters of the corresponding component. When determining the shape feature support state, planar feature points, edge feature points, and corner feature points are extracted from the component point cloud unit. The existence and spatial distribution of these feature points are then determined based on the component type to constrain the corresponding 3D shape parameters. For example, planar feature points of a wall are used to determine the wall orientation, edge feature points at the wall ends are used to determine the wall length, and corner feature points at wall corners are used to determine the connection position between adjacent walls. If a component point cloud unit only contains local planar feature points and lacks end edge feature points, its planar orientation can be identified, but the wall length still lacks observation constraints, and the shape feature support state is correspondingly reduced. The shape feature support state, along with the point cloud surface coverage state and point cloud distribution state, participates in the generation of the point cloud coverage state.

[0047] In one embodiment, a component instance is denoted as , will component instance The corresponding point cloud coverage confidence score is denoted as Point cloud coverage confidence We can perform weighted fusion based on the normalized point cloud surface cover state, point cloud distribution state, and shape feature support state. The calculation form can be expressed as:

[0048] in, This represents a component instance corresponding to a component point cloud unit; Representing a component instance The corresponding point cloud coverage confidence level; Representing a component instance The corresponding normalized value of the point cloud surface cover state; Representing a component instance The corresponding normalized value of the point cloud distribution state; Representing a component instance The corresponding shape feature supports the normalized state value; This represents the fusion weight corresponding to the point cloud surface coverage state; The fusion weights represent the point cloud distribution state. This represents the fusion weight corresponding to the shape feature support state. Each fusion weight is a non-negative number, and... Three normalized values ​​and three fusion weights jointly contribute to the point cloud coverage confidence score. The calculation.

[0049] The fusion weights can be configured based on the component type. For walls that primarily rely on planar feature points and end edge feature points to determine their 3D shape parameters, the fusion weights corresponding to the shape feature support states can be increased. For columnar components with continuous surface curvature and where the point cloud distribution has a significant impact on the fitting results, the fusion weights corresponding to the point cloud distribution can be increased. By mapping the fusion weights to component types, the point cloud coverage confidence can reflect the different needs of different component types for observation information and participate in the division of geometrically fitted point cloud units and generative inference point cloud units.

[0050] Optionally, dividing the component point cloud units into geometrically fitted point cloud units and generative inference point cloud units according to point cloud coverage confidence includes: Obtain the 3D shape reconstruction accuracy requirements configured for the parameterized 3D model of the point cloud to be generated, and generate the point cloud coverage confidence threshold based on the component type corresponding to the component point cloud unit and the 3D shape reconstruction accuracy requirements; The confidence score of the point cloud coverage corresponding to the point cloud unit of the component is compared with the confidence score threshold of the point cloud coverage to generate a confidence score comparison result. Based on the confidence level comparison results, component point cloud units whose point cloud coverage confidence level reaches the point cloud coverage confidence level threshold are identified as geometrically fitted point cloud units, while component point cloud units whose point cloud coverage confidence level does not reach the point cloud coverage confidence level threshold are identified as generated inference point cloud units.

[0051] Specifically, in this application, the 3D shape reconstruction accuracy requirement is used to limit the allowable range of component size deviation, component position deviation, and component orientation deviation in the parameterized 3D point cloud model to be generated. The 3D shape reconstruction accuracy requirement is read from the modeling task configuration data, which is configured according to the acquisition accuracy of the 3D point cloud, the usage scenario of the parameterized 3D point cloud model, and the component type. A higher 3D shape reconstruction accuracy requirement corresponds to a higher point cloud coverage confidence threshold used to determine the component point cloud unit as a geometrically fitted point cloud unit. The configured 3D shape reconstruction accuracy requirement participates in the determination of the point cloud coverage confidence threshold and the shape fitting tolerance.

[0052] In one embodiment, the point cloud coverage confidence threshold is denoted as Based on point cloud coverage confidence confidence threshold for point cloud coverage The comparison relationship between components divides the component instances into a high-confidence component set and a low-confidence component set. The high-confidence component set corresponds to the geometrically fitted point cloud units, and the low-confidence component set corresponds to the generated inference point cloud units. This division relationship can be expressed as:

[0053]

[0054] in, This represents the set of high-confidence components corresponding to geometrically fitted point cloud units; This represents the set of low-confidence components corresponding to the generated inferred point cloud units; Represents any instance of a component; Representing a component instance The corresponding point cloud coverage confidence level; Representation and component instances The component type and the point cloud coverage confidence threshold corresponding to the required accuracy of 3D shape reconstruction. High-confidence component set. Used to determine the geometrically fitted point cloud units that need to generate fitted 3D shape parameters, a low-confidence component set. Generative inference point cloud units are used to determine the trajectories of 3D shape state transformations and generative 3D shape parameters that need to be generated.

[0055] Point cloud coverage confidence threshold Configuration can be done separately according to component type. Walls, beams, columns, and doors / windows each use a point cloud coverage confidence threshold adapted to their respective 3D shape parameter determination requirements. Component point cloud units whose point cloud coverage confidence reaches the threshold are identified as geometric fitting point cloud units because their point cloud observations provide relatively sufficient spatial constraints for fitting component shape primitives. Component point cloud units whose point cloud coverage confidence does not reach the threshold are identified as generative inference point cloud units because their point cloud observations provide relatively insufficient constraints on complete 3D shape parameters.

[0056] In one exemplary embodiment, the point cloud coverage confidence threshold can be configured based on the 3D point cloud density and the complexity of the component shape (e.g., configured as follows). When the point cloud coverage confidence of a component instance is not lower than the point cloud coverage confidence threshold, the corresponding component point cloud unit is determined as a geometric fitting point cloud unit; when the point cloud coverage confidence of a component instance is lower than the point cloud coverage confidence threshold, the corresponding component point cloud unit is determined as a generation inference point cloud unit. This value is only used to explain the branching process of the component point cloud unit, and can be adjusted according to the accuracy requirements of 3D shape reconstruction in actual implementation.

[0057] Optionally, the fitted geometric fitting point cloud unit generates fitted three-dimensional shape parameters, including: The component shape primitives and shape fitting tolerances are determined based on the component type corresponding to the geometrically fitted point cloud units and the accuracy requirements for 3D shape reconstruction. The component shape primitives are fitted to the geometric fitting point cloud units to generate shape primitive parameters, and then the fitted three-dimensional shape parameters are determined based on the shape primitive parameters. Calculate the shape fitting residual between the fitted 3D shape parameters and the corresponding geometric fitted point cloud units, and compare the shape fitting residual with the shape fitting tolerance to generate the fitting residual comparison result. Based on the fitting residual comparison results, geometrically fitted point cloud units whose shape fitting residuals exceed the shape fitting tolerance are redefined as generative inference point cloud units, so that the redefined generative inference point cloud units can participate in the determination of generative 3D shape parameters.

[0058] Specifically, in this application, component shape primitives are geometric representations that can characterize the main three-dimensional shape of a corresponding component with fewer parameters. When determining component shape primitives based on component type, component shape primitives corresponding to the component type are read from a pre-built shape primitive library. The shape primitive library records the correspondence between component types and component shape primitives. For example, a wall can correspond to a planar primitive or a flat plate primitive with thickness, a cylindrical beam can correspond to a cylindrical surface primitive, and doors and windows can correspond to rectangular opening primitives. The three-dimensional shape reconstruction accuracy requirement is used to determine the shape fitting tolerance, which represents the allowable shape deviation range between the fitted three-dimensional shape parameters and the geometrically fitted point cloud units. Component shape primitives and shape fitting tolerance jointly participate in the generation of the fitting three-dimensional shape parameters and the comparison results of the fitting residuals.

[0059] When fitting component shape primitives to geometrically fitted point cloud units, a random sampling consensus algorithm is first used to repeatedly extract the smallest combination of point cloud points that can determine the component shape primitives from the geometrically fitted point cloud units. After each extraction, a set of candidate shape primitive parameters is calculated based on the smallest combination of point cloud points, and then the distances between the remaining point cloud points in the geometrically fitted point cloud unit and the candidate shape primitives are calculated. An interior point distance threshold is configured according to the accuracy requirements of 3D shape reconstruction. Point cloud points with a distance not greater than the interior point distance threshold are determined as candidate interior points. Candidate shape primitive parameters with a large number of candidate interior points and a relatively complete distribution range are selected as initial shape primitive parameters. These initial shape primitive parameters participate in the generation of shape primitive parameters.

[0060] After the initial shape primitive parameters are determined, least-squares fitting is performed using the candidate interior points corresponding to these parameters to reduce the overall distance between the candidate interior points and the component shape primitives, thereby generating the shape primitive parameters. The shape primitive parameters have different parameter compositions depending on the component type. For example, the shape primitive parameters corresponding to a planar primitive may include the plane normal vector, plane offset, and surface boundary; the shape primitive parameters corresponding to a cylindrical primitive may include the cylinder axis, cross-sectional dimensions, and axial range; and the shape primitive parameters corresponding to a rectangular opening primitive may include the opening center, opening width, opening height, and opening orientation. These shape primitive parameters are used to determine the component's position, dimensions, and orientation, thereby generating the fitted 3D shape parameters.

[0061] In one embodiment, a component instance The corresponding fitted 3D shape parameters are denoted as , will component instance The corresponding geometrically fitted point cloud unit is denoted as The candidate 3D shape parameters will be used to determine the shape parameters. The defined component shape primitive is denoted as Fitting 3D shape parameters This can be determined by minimizing the distance from point cloud points in the geometrically fitted point cloud unit to the component shape primitive:

[0062] in, This represents the candidate 3D shape parameters to be solved; Representing a component instance The corresponding fitted three-dimensional shape parameters; Representing a component instance The corresponding geometrically fitted point cloud unit; Represents geometrically fitted point cloud units Any point in the cloud; Represents the candidate 3D shape parameters Defined component shape primitives; Represents point cloud points To component shape primitive The geometric distance between them; This represents a robust distance function, which is used to reduce the impact of a small number of outliers on the fitted 3D shape parameters. The above factors collectively contribute to fitting the three-dimensional shape parameters. Solve for it.

[0063] Based on the fitted three-dimensional shape parameters Calculate the shape fitting residual. Instance the component. The corresponding shape fitting residual is denoted as Its calculation form can be expressed as:

[0064] in, Representing a component instance The corresponding shape fitting residuals; Represents geometrically fitted point cloud units The number of point cloud points included; Represents geometrically fitted point cloud units Any point in the cloud; This indicates the fit of three-dimensional shape parameters. Defined component shape primitives; Represents point cloud points To component shape primitive The geometric distance between them. Shape fitting residuals. The smaller the value, the better the fit of the three-dimensional shape parameters. The characterized component shape and geometrically fitted point cloud unit The smaller the average deviation between them, the better. Shape fitting residuals It participates in the generation of fitting residual comparison results.

[0065] When component instance Belongs to the set of high-confidence components First, geometric fitting is used to generate the fitting three-dimensional shape parameters. and shape fitting residuals If the shape fitting residuals If the value is not greater than the shape fitting tolerance, then the fitted three-dimensional shape parameters are retained. To fit the three-dimensional shape parameters Participate in subsequent shape continuation verification. If the shape fitting residual... If the value is greater than the shape fitting tolerance, it means that although the point cloud coverage confidence level has reached the point cloud coverage confidence threshold, the current geometric fitting point cloud unit still cannot enable the selected component shape primitive to form a fitting result that meets the accuracy requirements of 3D shape reconstruction. Therefore, the component instance is... From the high-confidence component set Reclassified into the low-confidence component set This allows the corresponding component point cloud units to participate in the determination of generative 3D shape parameters according to the generated inference point cloud units.

[0066] In one exemplary embodiment, the shape fitting tolerance can be configured according to the three-dimensional shape reconstruction accuracy requirements (e.g., configured as follows). (meters), and determine the initial shape primitive parameters through multiple random samplings (e.g., performing...) (Number of random samples). When the shape fitting residual exceeds the shape fitting tolerance, the corresponding geometrically fitted point cloud unit is redefined as the generated inference point cloud unit. The above values ​​are used to explain the backoff determination method of the geometrically fitted point cloud unit, and do not limit other embodiments to using the same values.

[0067] Optionally, the step of generating a three-dimensional shape state transformation trajectory for the generated inference point cloud unit by the teacher multi-step flow model includes: Encode and generate inference point cloud units to generate point cloud condition features, and read the component shape prior based on the component type corresponding to the generated inference point cloud units; By fusing point cloud conditional features and component shape priors, three-dimensional shape generation conditions are generated. Generate the initial state of the three-dimensional shape, and input the initial state of the three-dimensional shape and the conditions for generating the three-dimensional shape into the teacher's multi-step flow model so that the teacher's multi-step flow model can gradually update the initial state of the three-dimensional shape and generate multiple three-dimensional shape states. Multiple 3D shape states are arranged according to their update order to generate the 3D shape state transformation trajectory corresponding to the generated inference point cloud unit.

[0068] Specifically, in this application, point cloud conditional features are used to characterize the observed local geometric information in the generated inferred point cloud units. When encoding the generated inferred point cloud units, a local neighborhood is first established based on the point cloud points within the generated inferred point cloud unit, and then the relative positions of neighboring point cloud points with respect to the center point of the local neighborhood are calculated. Subsequently, the relative positions and the original attributes of the point cloud points are aggregated step-by-step to form a local point cloud representation. After globally converging multiple local point cloud representations, point cloud conditional features are generated. These point cloud conditional features simultaneously preserve the local surface changes, overall spatial extent, and observed boundary positions of the component point cloud units, constraining the direction of the three-dimensional shape state update of the teacher's multi-step flow model.

[0069] The component shape prior includes the shape parameter composition associated with the component type corresponding to the generated inferred point cloud unit, the range of component shape prior parameters, and the shape distribution of similar components. When fusing point cloud conditional features and component shape priors, the point cloud conditional features are first mapped to a 3D shape parameter space, and then the component shape priors are converted into component shape prior representations. Subsequently, the point cloud conditional features and component shape prior representations in the 3D shape parameter space are stitched together, and conditional mapping is performed on the stitching result to generate 3D shape generation conditions. The 3D shape generation conditions are used to enable the teacher's multi-step flow model to gradually form complete 3D shape parameters along the shape distribution of similar components under the constraints of partial point cloud observations.

[0070] In one embodiment, the teacher multi-step flow model includes a point cloud conditional coding layer, a shape prior coding layer, a conditional fusion layer, a state-time coding layer, a vector field prediction layer, a shape parameter projection layer, and a state integration process.

[0071] The point cloud conditional encoding layer performs multi-scale local aggregation on the generated inference point cloud units to generate point cloud conditional features. The point cloud conditional encoding layer can include multiple levels of local aggregation units (e.g., 3 levels). Each level of local aggregation unit extracts local geometric changes within the neighborhood of a point in the point cloud, while each subsequent level expands the spatial neighborhood to extract the overall spatial distribution of the components. The outputs of each level of local aggregation unit are sequentially used in the calculations of subsequent local aggregation units, with the last level outputting the point cloud conditional features.

[0072] The shape prior encoding layer converts the shape parameter composition, the range of shape prior parameters, and the shape distribution of similar components in the component shape prior into component shape prior representations. For different component types, the shape prior encoding layer organizes the component shape prior representations with the same length, enabling the point cloud conditional features to be fused with the component shape prior representations of different component types. The component shape prior representation serves as the input to the conditional fusion layer.

[0073] The conditional fusion layer receives point cloud conditional features and component shape prior representations, and performs conditional modulation on these features to generate 3D shape generation conditions. Conditional modulation adjusts the participation level of different shape parameters in the component shape prior representations based on the point cloud conditional features. For example, when a wall plane and one end have been observed in the generated inference point cloud unit, the 3D shape generation conditions impose strong constraints on the wall orientation and the observed end position, while retaining the supplementary constraints of the component shape prior on missing ends and wall length. The 3D shape generation conditions participate in the calculations of the state-time coding layer and the vector field prediction layer.

[0074] The state-temporal encoding layer jointly encodes the current 3D shape state, the current transformation time, and the 3D shape generation conditions to generate state-temporal encoded features. The current 3D shape state represents the intermediate values ​​of the 3D shape parameters at the current transformation time; the current transformation time represents the position of the current 3D shape state within the entire 3D shape state transformation trajectory. The state-temporal encoded features allow the same current 3D shape state to correspond to different 3D shape state update vectors at different transformation times. These state-temporal encoded features serve as input to the vector field prediction layer.

[0075] The vector field prediction layer determines the 3D shape state update vector corresponding to the current 3D shape state based on the state-time encoding features. The vector field prediction layer can include multiple fully connected residual units (e.g., four). Each fully connected residual unit sequentially receives the output of the previous fully connected residual unit and introduces 3D shape generation conditions into each fully connected residual unit, allowing point cloud conditional features and component shape priors to permeate multiple 3D shape state update layers. The last fully connected residual unit generates vector field prediction features, which serve as input to the shape parameter projection layer.

[0076] The shape parameter projection layer maps the predicted features of the vector field to the 3D shape parameter space to generate a 3D shape state update vector, ensuring that the dimension of the 3D shape state update vector matches the dimension of the 3D shape parameters corresponding to the component type. The 3D shape state update vector participates in the state integration process. The state integration process sequentially applies the 3D shape state update vector to the current 3D shape state at multiple transformation moments to progressively update the initial 3D shape state and generate multiple 3D shape states. After arranging the multiple 3D shape states according to their update order, a 3D shape state transformation trajectory is generated.

[0077] In one embodiment, a component instance The corresponding generated inference point cloud unit is denoted as , will component instance The corresponding component shape prior is denoted as The conditional distribution of the teacher's multi-step flow model learning is represented as follows:

[0078] in, Represents the conditional distribution of three-dimensional shape parameters; Representing a component instance The corresponding complete three-dimensional shape parameters; Representing a component instance The corresponding generated inference point cloud unit; Representing a component instance The component shape prior corresponds to the component type. This conditional distribution represents the teacher's multi-step flow model in generating and inferring point cloud units. and component shape prior Under common constraints, for complete three-dimensional shape parameters Inference is made. Conditional distributions are involved in the training of the teacher's multi-step flow model and the generation of three-dimensional shape state transformation trajectories.

[0079] The teacher's multi-step flow model begins with an initial state of a 3D shape and then proceeds with state transitions. The initial state of the 3D shape is generated through random sampling. (The last sentence appears to be incomplete and possibly refers to a separate process: "component instances..."). The corresponding initial state of the three-dimensional shape is denoted as The final three-dimensional shape state is denoted as , where subscript The index represents the starting state in the discrete state transition sequence. This represents the index of the final state in the discrete state transformation sequence. The three-dimensional shape state transformation trajectory starts from the initial state of the three-dimensional shape. To the final three-dimensional shape state It consists of multiple three-dimensional shape states arranged according to the transformation time, and the final three-dimensional shape state Corresponding component example Prediction of three-dimensional shape parameters.

[0080] In one exemplary embodiment, the teacher multi-step flow model can be trained via conditional flow matching. This involves considering the complete three-dimensional shape parameters in the training samples. The initial state of the three-dimensional shape formed by random sampling According to the normalization transformation time Constructing training three-dimensional shape states:

[0081] in, Representing a component instance At the time of normalization transformation The corresponding training 3D shape state; Indicates the time of termination To the start time The normalization transformation time between them; Represents component instances in the training samples The corresponding complete three-dimensional shape parameters; Representing a component instance The corresponding initial state of the three-dimensional shape. When near At the same time, train the three-dimensional shape state Approximately the initial state of a three-dimensional shape ;when near At the same time, train the three-dimensional shape state Approximately complete three-dimensional shape parameters The starting state index in the discrete state transformation sequence. With normalization transformation time Correspondingly, the termination state index in the discrete state transformation sequence With normalization transformation time correspond.

[0082] Let the 3D shape state update vector of the teacher's multi-step flow model be denoted as... ,in, This represents the parameters of the teacher multi-step flow model. A defined state updates the vector field; Indicates the current training 3D shape state; Indicates the current time of the normalization transformation; This represents the three-dimensional shape generation conditions formed by fusing point cloud conditional features and component shape priors. This represents the model parameters involved in training the teacher multi-step flow model. The training loss of the teacher multi-step flow model can be expressed as:

[0083] in, This represents the training loss of the teacher's multi-step flow model; This represents a statistical average of the training samples, the initial state of the 3D shape, and the time of the normalization transformation. Represents the square Euclidean norm; This represents the 3D shape state update vector predicted by the teacher's multi-step flow model; Represents a complete three-dimensional shape parameter Pointing to the initial state of the three-dimensional shape The target 3D shape state update vector is obtained by reducing the training loss. Vector field prediction layer learns under the conditions of 3D shape generation Under the constraints, determine the three-dimensional shape state update vector corresponding to different normalization transformation times.

[0084] During inference, the 3D shape state is updated according to a sequence of transformation times decreasing from the start time to the end time. The 3D shape state update between two adjacent transformation times can be represented as:

[0085] in, Indicates the first Each transformation moment; Indicates that it is located at Then and closer to the end time The transformation time, therefore ; Indicates the time of transformation The corresponding three-dimensional shape state; Indicates the time of change after the update. The corresponding three-dimensional shape state; This indicates the teacher's multi-step flow model at the transition time. The predicted 3D shape state update vector; This represents the time step between two adjacent transformation moments. After repeatedly performing the above 3D shape state update at multiple transformation moments, the initial state of the 3D shape is... Gradually transforming into the final three-dimensional shape state .

[0086] In an exemplary embodiment, the teacher multi-step flow model can complete the 3D shape state update at multiple transformation moments (e.g., 20 to 40 steps, or 32 steps). Each 3D shape state update retains the 3D shape state before and after the update. After arranging all the 3D shape states in the update order, a 3D shape state transformation trajectory corresponding to the generated inference point cloud unit is formed. The 3D shape state transformation trajectory not only includes the final 3D shape parameter prediction, but also retains the state change process of the teacher multi-step flow model from the initial 3D shape state to the complete 3D shape parameters, for use in subsequent jump flow distillation.

[0087] Optionally, the stepwise updating of the initial state of the 3D shape and generation of multiple 3D shape states by the teacher's multi-step flow model includes: For the same point cloud unit, multiple initial states of three-dimensional shapes are generated. The teacher multi-step flow model is used to perform state transformation according to each initial state of three-dimensional shape and the three-dimensional shape generation conditions to generate candidate three-dimensional shape state transformation trajectories. The termination state of each candidate 3D shape state transformation trajectory is converted into candidate 3D shape parameters, and the candidate 3D surface point cloud is reconstructed based on the candidate 3D shape parameters. The point cloud coordinate domain is determined based on the 3D point cloud, and each candidate 3D surface point cloud is mapped to the point cloud coordinate domain in order to calculate the point cloud reprojection residual between each candidate 3D surface point cloud and the corresponding generated inference point cloud unit. Candidate 3D shape state transformation trajectories are selected based on the point cloud reprojection residuals, and then the candidate 3D shape state transformation trajectory with the smallest point cloud reprojection residual is determined as the 3D shape state transformation trajectory corresponding to the generated inferred point cloud unit.

[0088] Specifically, in this application, partial point cloud observations corresponding to the same generated inference point cloud unit cannot completely define the 3D shape of the missing region. Therefore, when the teacher multi-step flow model starts from different initial states of 3D shapes, it can form multiple candidate 3D shape state transformation trajectories. These multiple candidate 3D shape state transformation trajectories correspond to different candidate 3D shape parameters formed under the point cloud conditional features and component shape prior constraints. By comparing the point cloud reprojection residuals between the candidate 3D shape parameters and the observed generated inference point cloud units, the 3D shape state transformation trajectory that is more consistent with the current generated inference point cloud unit can be selected from the multiple candidate 3D shape state transformation trajectories.

[0089] When generating multiple initial states of 3D shapes for the same generative inference point cloud unit, multiple random states are independently sampled from a uniform random distribution, and each random state is mapped to the 3D shape parameter space of the corresponding component type to generate multiple initial states of 3D shapes. Each initial state of a 3D shape is input into the teacher's multi-step flow model along with the same 3D shape generation condition. The teacher's multi-step flow model performs multi-step state transformations on each initial state of a 3D shape, thereby forming multiple candidate 3D shape state transformation trajectories.

[0090] When converting the final state of the candidate 3D shape state transformation trajectory into candidate 3D shape parameters, the parameter components in the final state are analyzed according to the parameter meaning corresponding to the component type. For walls, the candidate 3D shape parameters can be analyzed as wall position, wall length, wall height, wall thickness, and wall orientation; for doors and windows, the candidate 3D shape parameters can be analyzed as opening position, opening width, opening height, and opening orientation. Based on the analyzed candidate 3D shape parameters, the corresponding component surface is reconstructed, and the reconstructed component surface is sampled according to the surface sampling interval to generate candidate 3D surface point clouds. The candidate 3D surface point clouds are used in the calculation of the point cloud reprojection residual.

[0091] Candidate 3D surface point clouds are formed by sampling the surfaces of components characterized by candidate 3D shape parameters. The surface sampling interval is configured based on the average point spacing of the 3D point cloud, ensuring that the candidate 3D surface point cloud has a comparable sampling density to the 3D point cloud. The point cloud coordinate domain is jointly determined by the origin, coordinate axis directions, and length units of the 3D point cloud. When mapping the candidate 3D surface point cloud to the point cloud coordinate domain, coordinate transformation is performed based on the component position and orientation in the candidate 3D shape parameters, ensuring that the candidate 3D surface point cloud and the corresponding generated inference point cloud unit are located in the same point cloud coordinate domain. The mapped candidate 3D surface point cloud participates in the calculation of the point cloud reprojection residual.

[0092] In one embodiment, the first The candidate 3D surface point cloud formed by reconstructing the candidate 3D shape parameters is denoted as . The corresponding point cloud reprojection residual is denoted as The point cloud reprojection residual can be determined based on the nearest distance from each point in the generated inference point cloud cell to the candidate 3D surface point cloud:

[0093] in, Representing a component instance The Point cloud reprojection residuals corresponding to each candidate 3D shape parameter; The number representing the candidate 3D shape state transformation trajectory; Representing a component instance The corresponding generated inference point cloud unit; Indicates the generation of inference point cloud units The number of point cloud points included; Indicates the generation of inference point cloud units Any point in the cloud; Indicates the first Candidate 3D surface point clouds are formed by reconstructing candidate 3D shape parameters; Represents candidate 3D surface point clouds Candidate surface sampling points; Represents point cloud points With candidate surface sampling points The Euclidean distance between them. Point cloud reprojection residual. Reflecting the The spatial deviation between the candidate 3D shape parameters corresponding to the candidate 3D shape state transformation trajectory and the observed generated inferred point cloud units.

[0094] When selecting candidate 3D shape state transformation trajectories based on the point cloud reprojection residuals, the candidate 3D shape state transformation trajectory with the smallest point cloud reprojection residual is numbered and denoted as... The filtering relationship can be represented as:

[0095] in, Indicates the selected candidate 3D shape state transformation trajectory number; Indicates the trajectory number of any candidate 3D shape state transformation; Representing a component instance The The point cloud reprojection residuals corresponding to the candidate 3D shape state transformation trajectories. (The numbering is missing from the original text.) The corresponding candidate 3D shape state transformation trajectory is determined as the 3D shape state transformation trajectory for the generated inferred point cloud unit, and the termination state of this 3D shape state transformation trajectory is determined as the teacher supervision target. The teacher supervision target and the corresponding 3D shape state transformation trajectory jointly participate in the generation of subsequent jump flow distillation supervision data.

[0096] Point cloud reprojection residuals are used to constrain candidate 3D shape parameters to be consistent with the observed generated and inferred point cloud units. Component shape priors are used to constrain unobserved component regions to maintain shape distribution consistent with the same type of component of the corresponding component type. Point cloud reprojection residuals and component shape priors constrain the observed and missing regions, respectively, so that the selected 3D shape state transformation trajectories retain some spatial information provided by point cloud observations while forming complete 3D shape parameters for the corresponding component type.

[0097] In a preferred embodiment, in addition to the point cloud reprojection residual, it can also be determined whether the candidate 3D shape parameters fall within the range of the component shape prior parameters. When the candidate 3D shape parameters exceed the range of the component shape prior parameters, the corresponding candidate 3D shape state transformation trajectory is not used as the 3D shape state transformation trajectory; among the remaining candidate 3D shape state transformation trajectories, the 3D shape state transformation trajectory is selected based on the point cloud reprojection residual. This implementation ensures that the 3D shape state transformation trajectory simultaneously satisfies the observed point cloud constraints and the component shape prior parameter range constraints, providing continuous and component-type-adaptive monitoring data for subsequent skip flow distillation based on the 3D shape state transformation trajectory.

[0098] After determining the 3D shape state transformation trajectory corresponding to the generated inferred point cloud unit, the 3D shape state transformation trajectory is further divided into jump flow intervals, and the jump flow intervals are used to form jump flow distillation supervision data for the 3D shape generation network. After the 3D shape generation network is trained, the generative 3D shape parameters are determined using the 3D shape generation network, and then the fitted 3D shape parameters and the generative 3D shape parameters are jointly included in the shape continuity verification. The processing path deviation and component type deviation found in the shape continuity verification are used for processing path adjustment and 3D shape generation network update, respectively. The updated 3D shape parameters formed after the update participate in the shape continuity verification again until a set of verified and confirmed 3D shape parameters is formed.

[0099] Optionally, the step of performing jump flow distillation based on the three-dimensional shape state transformation trajectory to generate a three-dimensional shape generation network includes: Compare adjacent three-dimensional shape states in the three-dimensional shape state transformation trajectory to generate trajectory change states; The shape parameter dimension is determined based on the component type corresponding to the generated inferred point cloud unit, and the component shape change degree is determined based on the component shape prior. The shape parameter dimension and the component shape change degree are fused to generate the component type complexity. The trajectory of three-dimensional shape state transformation is divided according to the trajectory change state and component type complexity to generate jump flow intervals; The jump start state and jump end state are determined based on the jump flow interval. The jump start state and jump end state are associated to generate jump flow distillation supervision data, which is then used to participate in the jump flow distillation training of the 3D shape generation network.

[0100] Specifically, in this application, the three-dimensional shape state transformation trajectory consists of multiple three-dimensional shape states arranged in the order of state transformation. The parameter components in each three-dimensional shape state are adapted to the component type corresponding to the generated inference point cloud unit. For example, the three-dimensional shape state corresponding to a wall may include the wall center position, wall length, wall height, wall thickness, and wall orientation; the three-dimensional shape state corresponding to a door or window may include the opening center position, opening width, opening height, and opening orientation. In order to enable parameter components of different dimensions to participate in the determination of the trajectory change state, the three-dimensional shape state is first normalized according to the value span of each parameter component within the prior parameter range of the component shape. Then, adjacent three-dimensional shape states are compared to generate the state change amount between adjacent three-dimensional shape states. After arranging the state change amounts in the order of state transformation, the trajectory change state is formed, and the trajectory change state further participates in the division of the jump flow interval.

[0101] Component instance In the The three-dimensional shape state corresponding to each state transition moment is denoted as . , will component instance The corresponding normalized scaling matrix is ​​denoted as Then the state change between adjacent three-dimensional shape states can be expressed as:

[0102] in, This indicates the generation of component instances corresponding to the inferred point cloud units; This indicates the state sequence number of the three-dimensional shape state in the three-dimensional shape state transformation trajectory; Representing a component instance In the The three-dimensional shape state corresponding to each state transition moment; Representing a component instance In the The three-dimensional shape state corresponding to each state transition moment; Indicates based on component instance The normalized scaling matrix formed by the range of shape parameter values ​​for the corresponding component type; Represents the normalized scaling matrix The inverse matrix; Indicates the first The three-dimensional shape state and the first The amount of state change between three-dimensional shape states; Represents the Euclidean norm. State change quantity. Numbered according to state sequence Arranged to form component instances The corresponding trajectory change state.

[0103] The trajectory change states reflect the extent to which the teacher's multi-step flow model adjusts the 3D shape parameters at different state transition stages. When the state change between adjacent 3D shape states is large, the corresponding trajectory segment undertakes more 3D shape parameter adjustments; when the state change is small, the corresponding trajectory segment approaches a stable 3D shape parameter region. When determining the boundaries of the jump flow intervals based on the trajectory change states, denser jump flow interval boundaries are set in trajectory segments with larger state changes, and sparser jump flow interval boundaries are set in trajectory segments with smaller state changes, so that the division density of the jump flow intervals matches the shape change amplitude in the 3D shape state transition trajectory.

[0104] Specifically, in this application, the shape parameter dimension refers to the number of independent shape parameters contained in the component type corresponding to the generated inferred point cloud unit. The larger the shape parameter dimension, the more parameter coupling relationships there are in the three-dimensional shape state. For example, the shape parameters of a regular rectangular window mainly include the center position of the opening, the width of the opening, the height of the opening, and the orientation of the opening; in addition to the above shape parameters, an irregular corner window also includes the corner position, the corner direction, and the dimensions of the openings on both sides, thus having a higher shape parameter dimension.

[0105] The degree of component shape variation is determined based on the differences between the shapes of multiple historical components of the same component type in the prior knowledge of component shape. First, the shape parameters of historical components of the same component type are scaled and normalized. Then, the dispersion of each historical component shape parameter relative to the central shape parameter corresponding to that component type is calculated to generate the degree of component shape variation. The central shape parameter corresponding to that component type is formed by statistically aggregating the shape parameters of historical components of the same component type according to their parameter components, and is used to calculate the degree of component shape variation. A higher degree of component shape variation indicates more morphological changes within the same component type; a lower degree of component shape variation indicates a more concentrated distribution of shape parameters within the same component type.

[0106] The component type is denoted as , component type The corresponding shape parameter dimension is denoted as , component type The degree of change in the corresponding component shape is denoted as The component type complexity can then be expressed as:

[0107] in, Indicate component type The corresponding component type complexity; Indicate component type The corresponding shape parameter dimension; This represents the maximum shape parameter dimension among all component types participating in the jump flow distillation training; Indicate component type The corresponding degree of change in component shape; This indicates the maximum degree of component shape change among all component types participating in the jump flow distillation training; This represents the complexity fusion weight corresponding to the shape parameter dimension; The complexity fusion weight represents the degree of change in the component's shape. and Both are non-negative numbers, and their sum is Component type complexity Together with the trajectory change state, it participates in the division of the jump flow interval.

[0108] Specifically, in this application, the state change amount is accumulated along the three-dimensional shape state transformation trajectory. When the accumulated state change amount reaches the interval change threshold corresponding to the component type complexity, a jump flow interval boundary is set at the current three-dimensional shape state. The interval change threshold is configured according to the component type complexity: when the component type complexity is high, the interval change threshold is configured to a smaller value; when the component type complexity is low, the interval change threshold is configured to a larger value. The configured interval change threshold and the trajectory change state are used together to locate the jump flow interval boundary, thereby avoiding placing the drastically changing shape segments and the gently changing shape segments under the same state mapping difficulty when dividing the three-dimensional shape state transformation trajectory according to a fixed number of steps.

[0109] The third in the three-dimensional shape state transformation trajectory Each jump flow interval is represented as:

[0110] in, Representing a component instance The corresponding number A jump flow interval; Indicates the jump flow interval number; Indicates the first The jump start state corresponding to each jump flow interval; Indicates the first The jump termination state corresponding to each jump flow interval; This indicates the state sequence number of the jump start state in the three-dimensional shape state transformation trajectory; This indicates the state sequence number of the jump termination state in the 3D shape state transformation trajectory, and the jump initiation state participates in the state transformation before the jump termination state. Jump initiation state and jump termination state After association, the first Jump flow distillation monitoring data corresponding to each jump flow interval.

[0111] For adjacent jump flow intervals, the jump termination state of the previous jump flow interval corresponds to the same trajectory boundary as the jump initiation state of the next jump flow interval. Jump flow distillation supervision data is used to train the 3D shape generation network to learn the state mapping from the jump initiation state to the jump termination state, thereby compressing the multiple 3D shape state updates performed by the teacher's multi-step flow model within the jump flow interval into a single state mapping or a few state mappings.

[0112] In one exemplary embodiment, for obstructing walls with low shape parameter dimensions and low component shape change, the three-dimensional shape state transformation trajectory can be divided into fewer jump flow intervals (e.g., 4 jump flow intervals); for irregular corner windows with high shape parameter dimensions and high component shape change, the three-dimensional shape state transformation trajectory can be divided into more jump flow intervals (e.g., 8 jump flow intervals). The above number of jump flow intervals is used to illustrate the impact of component type complexity on the division of jump flow intervals. In actual implementation, the number of jump flow intervals is determined based on the trajectory change state and component type complexity.

[0113] Optionally, the jump flow distillation supervision data participates in the jump flow distillation training of the 3D shape generation network, including: An initial 3D shape generation network is constructed, which combines the jump start state, point cloud condition features, and the component type corresponding to the generated inferred point cloud unit to generate the jump flow distillation input, and the jump termination state is determined as the jump flow distillation target. The jump flow distillation input is fed into the initial 3D shape generation network to generate jump state prediction results. The jump state prediction results are compared with the jump flow distillation target to generate shape reconstruction error. The jump state prediction results corresponding to adjacent jump flow intervals are continued according to the state transformation order of the three-dimensional shape state transformation trajectory to generate cross-interval prediction trajectory. Then, the cross-interval prediction trajectory and the three-dimensional shape state transformation trajectory are compared to generate flow matching consistency error. The shape reconstruction error and flow matching consistency error are combined to generate a skip flow distillation loss. The initial 3D shape generation network is updated based on the skip flow distillation loss to generate a 3D shape generation network.

[0114] Specifically, in this application, the jump start state is taken as the 3D shape state to be transformed, the point cloud conditional features are taken as the observed component point cloud constraints, and the component type corresponding to the generated inference point cloud unit is taken as the shape parameter structural constraint. To enable the initial 3D shape generation network to distinguish jump flow intervals of different spans, the state sequence numbers corresponding to the jump start state and the jump termination state are converted into interval position representations. These interval position representations are then combined with the jump start state, point cloud conditional features, and the component type corresponding to the generated inference point cloud unit to generate the jump flow distillation input. The interval position representation represents the start position, end position, and interval span of the jump flow interval in the 3D shape state transformation trajectory, and participates in the generation of jump state prediction results in the initial 3D shape generation network. The jump flow distillation objective is the jump termination state formed by the teacher's multi-step flow model at the end of the corresponding jump flow interval.

[0115] Let the initial 3D shape generation network be denoted as... ,in, This represents the network parameters used in training the initial 3D shape generation network. For component instances... The corresponding number For each jump flow interval, the jump state prediction result can be expressed as:

[0116] in, This indicates that the initial 3D shape generation network is for the first... Jump state prediction results generated for each jump flow interval; Indicated by network parameters A defined initial 3D shape generation network; Indicates the initial state of the jump; This represents the point cloud condition features corresponding to the generated inferred point cloud units; Representing a component instance Corresponding component types; The state sequence number representing the initial jump state; This indicates the state sequence number representing the jump termination state. and state sequence number Together, they form a region position representation, and enable the initial three-dimensional shape generation network to generate jump state prediction results based on the jump flow region span.

[0117] Shape reconstruction error is used to measure the parameter difference between the jump state prediction result and the jump flow distillation target. To reduce the impact of differences in the dimensions of different shape parameters on jump flow distillation training, a normalized scaling matrix is ​​used. Scale transformation is performed on the parameter differences. The shape reconstruction error can be expressed as:

[0118] in, Indicates shape reconstruction error; This represents the total number of jump flow intervals participating in the current training batch; This indicates the result of the jump state prediction; This indicates the jump termination state generated by the teacher's multi-step flow model; Representing a component instance The corresponding normalized scaling matrix; Represents the normalized scaling matrix The inverse matrix; Represents the squared Euclidean norm. Reduces shape reconstruction error. This enables the initial 3D shape generation network to learn the jump start state to jump end state mapping of the teacher's multi-step flow model in each jump flow interval.

[0119] Specifically, in this application, a cross-interval prediction trajectory is formed according to the state transformation order of the three-dimensional shape state transformation trajectory. First, the... The jump state prediction result generated in the jump flow interval is used as the first jump state prediction result. The input state of the first jump flow interval is then used by the initial 3D shape generation network to generate the second jump flow interval. The prediction results of the successive jump states corresponding to each jump flow interval are then repeated according to the state transition order. These successive jump state prediction results form the cross-interval prediction trajectory. The cross-interval prediction trajectory is used in the calculation of the flow matching consistency error.

[0120] Will be by the first The jump state prediction results for the first jump flow interval continue to the second. The prediction results of the successive jump states formed by the jump flow intervals are denoted as: ,but:

[0121] in, This indicates the prediction result of the successive jump state formed by the continuous succession of adjacent jump flow intervals; Indicates the first Prediction results of the jump state corresponding to each jump flow interval; Indicates the first The sequence number of the jump start state corresponding to each jump flow interval; Indicates the first The jump termination state sequence number corresponds to each jump flow interval; the remaining symbols have the same meaning as in the aforementioned formula. The prediction results of each successive jump state are arranged in the order of state transformation to form a cross-interval prediction trajectory.

[0122] Flow matching consistency error measures the difference in continuity between the predicted trajectory across intervals and the 3D shape-state transformation trajectory generated by the teacher's multi-step flow model at the boundaries of the skip flow intervals. Flow matching consistency error can be expressed as:

[0123] in, Indicates the stream matching consistency error; This indicates the number of adjacent skip flow interval pairs participating in the current training batch; This indicates the prediction result for the subsequent jump state; The teacher multi-step flow model is represented in the first step. The jump termination state is formed by the endpoint of each jump flow interval; the remaining symbols have the same meaning as in the aforementioned formula. This reduces flow matching consistency error. This enables the initial 3D shape generation network to learn not only the jump termination state of a single jump flow interval, but also the evolution relationship of the 3D shape state when multiple jump flow intervals are connected consecutively.

[0124] The skip flow distillation loss is generated by combining shape reconstruction error and flow matching consistency error, and its calculation form can be expressed as:

[0125] in, This indicates the loss during jump flow distillation; Indicates shape reconstruction error; Indicates the stream matching consistency error; This represents the loss weight corresponding to the shape reconstruction error; This represents the loss weight corresponding to the flow matching consistency error. Loss Weight and loss weight All are non-negative numbers, configured based on the average span of the skip flow interval and the complexity of the component type. Skip flow distillation loss. Network parameter updates for the initial 3D shape generation network.

[0126] When updating the initial 3D shape generation network based on the jump flow distillation loss, the jump flow distillation loss is calculated relative to the network parameters. The gradient is calculated, and the network parameters are updated along the direction that reduces the loss of the hopping flow distillation. When the shape reconstruction error and flow matching consistency error corresponding to the validation sample remain stable over multiple consecutive training epochs, or when the training epochs reach the configured upper limit, the initial 3D shape generation network with updated network parameters is determined as the 3D shape generation network. The configured upper limit of the training epochs is determined by the number of training samples and the number of component types, and serves as one of the termination conditions for skip flow distillation training.

[0127] In one exemplary embodiment, the 3D shape generation network performs a forward computation on the generative inference point cloud unit to directly form generative 3D shape parameters; for generative inference point cloud units with high component type complexity, forward computation can also be performed continuously according to a few jump flow intervals. This inference method compresses the multiple 3D shape state updates of the teacher's multi-step flow model into one or a few state mappings, thereby reducing the amount of computation required for a large number of generative inference point cloud units to perform multi-step state transformations sequentially.

[0128] Optionally, the initial 3D shape generation network includes a shared shape decoding backbone and a component type-specific determination head corresponding to each component type. The 3D shape generation network generated after updating the initial 3D shape generation network according to the jump flow distillation loss retains the shared shape decoding backbone and the component type-specific determination head. The process of determining generative 3D shape parameters by the 3D shape generation network includes: using a shared shape decoding backbone to perform shape decoding based on point cloud conditional features to generate a shared 3D shape representation, and determining a target component type-specific determination head based on the component type corresponding to the generated inferred point cloud unit. The shared 3D shape representation is input into a target component type-specific determination head to generate generative 3D shape parameters corresponding to the generated inferred point cloud units.

[0129] Specifically, in this application, the shared shape decoding backbone is used to extract the 3D shape generation information that exists in common among different component types. The shared shape decoding backbone includes a state mapping layer, a condition modulation layer, a multi-level residual shape decoding layer, and a shared shape representation output layer. The layers are connected in the aforementioned order so that the processing result formed by the previous layer participates in the shape decoding processing of the next layer.

[0130] The state mapping layer maps the jump initiation state or the initial state of the 3D shape during the inference phase into a state embedding representation. Each element in the state embedding representation represents the component position, component scale, component orientation, and the combination relationships between various shape parameters. The state embedding representation serves as the input to the conditional modulation layer and the multi-level residual shape decoding layer.

[0131] The conditional modulation layer receives point cloud conditional features and generates conditional modulation parameters corresponding to each multi-level residual shape decoding layer based on these features. These conditional modulation parameters are used to adjust the state embedding representation of each layer according to the observed spatial distribution of the point cloud, enabling the shared shape decoding backbone to form different shared 3D shape representations for different generated and inferred point cloud units under the same initial 3D shape state. The conditional modulation parameters participate in the calculation of each level of the residual shape decoding layer.

[0132] The multi-level residual shape decoding layer updates the state embedding representation level by level according to the connection relationship between levels. Each level of the residual shape decoding layer first performs a linear transformation on the state embedding representation of the previous level, then adjusts the linear transformation result using conditional modulation parameters, and finally adds the adjusted linear transformation result to the state embedding representation of the previous level to generate the current level state embedding representation. The current level state embedding representation serves as the input to the next level of the residual shape decoding layer. Residual connections enable the multi-level residual shape decoding layer to retain the existing parameter information in the jump start state, while simultaneously supplementing the shape changes indicated by the point cloud conditional features level by level.

[0133] The shared shape representation output layer transforms the final-level state embedding representation into a shared 3D shape representation. This shared 3D shape representation includes cross-component type common information such as component position, component scale, component principal orientation, and local shape variations. After entering the target component type-specific determination header, the shared 3D shape representation is further converted into generative 3D shape parameters for the corresponding component type.

[0134] In one embodiment, the shared shape decoding backbone may include multiple residual shape decoding layers (e.g., 4 to 8 layers), each residual shape decoding layer employing the same layer structure and receiving conditional modulation parameters formed from point cloud conditional features. The number of residual shape decoding layers is used to illustrate the implementation of the shared shape decoding backbone, and in actual implementation, it is configured according to the dimensionality of the three-dimensional shape parameters and the number of component types.

[0135] Specifically, in this application, each component type-specific determination header corresponds to a three-dimensional shape parameter structure for a component type. The component type-specific determination header includes a type parameter mapping layer, a parameter range constraint layer, and a three-dimensional shape parameter output layer. These layers are connected in the aforementioned order to convert the shared three-dimensional shape representation into generative three-dimensional shape parameters corresponding to the respective component type.

[0136] The type parameter mapping layer maps the shared 3D shape representation to a type parameter representation adapted to the target component type. For example, the type parameter representation for a wall includes the wall center position, wall length, wall height, wall thickness, and wall orientation; the type parameter representation for an irregular corner window also includes the corner position, corner direction, and the dimensions of the openings on both sides. The type parameter representation serves as input to the parameter range constraint layer.

[0137] The parameter range constraint layer performs range mapping on each parameter component in the type parameter representation based on the prior parameter range of the component shape. For non-negative shape parameters such as length, width, and height, the parameter range constraint layer maps the corresponding parameter components to non-negative intervals. For component orientation, the parameter range constraint layer adopts a parameter form representing periodic directions to reduce discontinuous parameter differences caused when the angle crosses the periodic boundary. The range-mapped type parameter representation serves as the input to the 3D shape parameter output layer.

[0138] The 3D shape parameter output layer converts the range-mapped type parameter representation into generative 3D shape parameters corresponding to the generated inferred point cloud units. The generative 3D shape parameters and the fitted 3D shape parameters use the same component position, component size, and component orientation expression methods, enabling the generative 3D shape parameters and the fitted 3D shape parameters to jointly participate in subsequent shape continuation verification.

[0139] During the jump flow distillation training phase, the shared shape decoding backbone is updated based on the jump flow distillation supervision data corresponding to all component types to learn the 3D shape state mapping across component types. Each component type-specific determination head is updated based on the jump flow distillation supervision data for its corresponding component type to learn the parameter composition and prior parameter range of the component shape for that type. This parameter organization ensures that when the 3D shape generation network is subsequently updated based on the component type failure state, the network parameters of the shared shape decoding backbone and other component type-specific determination heads remain unchanged; only the component type-specific determination head corresponding to the actively distilled component type is updated.

[0140] Optionally, the step of jointly verifying the fitted three-dimensional shape parameters and generative three-dimensional shape parameters according to the spatial adjacency relationship of components to generate shape continuity verification results includes: Based on the spatial adjacency relationship of components, the component pairs to be verified are determined from the components corresponding to the fitted three-dimensional shape parameters and the generative three-dimensional shape parameters, and then the component connection type corresponding to the component pairs to be verified is determined. Extract shape continuity features from the fitted three-dimensional shape parameters or generative three-dimensional shape parameters of the component to be verified on both sides respectively; Based on the component connection type, determine the shape connection comparison rules and shape connection tolerance, and compare the shape connection characteristics of the component to be verified on both sides according to the shape connection comparison rules to generate shape connection residuals; The shape continuity residual is compared with the shape continuity tolerance to generate the shape continuity status. Then, the shape continuity status of each component to be verified is summarized to generate the shape continuity verification result.

[0141] Specifically, in this application, each component spatial adjacency relationship is associated with two component point cloud units and their corresponding component connection types. Component connection types include end connection types and containment connection types. End connection types are applicable to adjacent walls, beams, columns, or other components connected at their ends; containment connection types are applicable to the spatial relationship between doors / windows and their respective walls. Component connection types participate in the determination of shape continuity features, shape continuity comparison rules, and shape continuity tolerances.

[0142] After determining the pair of components to be verified based on their spatial adjacency, the fitted 3D shape parameters or generative 3D shape parameters of the components on both sides of the pair are read. For the updated 3D shape parameters that are re-formed after the processing path adjustment, the updated 3D shape parameters are used as the 3D shape parameters of the corresponding components currently participating in the shape continuity verification, and the pair of components to be verified is formed together with the 3D shape parameters of the adjacent components based on their spatial adjacency.

[0143] Shape continuity features are extracted from 3D shape parameters based on the component connection type. For end connection types, shape continuity features include the component endpoint position, end boundary direction, and end boundary range. For containment connection types, shape continuity features include the outer boundary of the contained component and the shape containment range corresponding to the containing component. The shape containment range is determined by the 3D shape parameters of the containing component, such as the wall surface boundary and wall thickness range of the wall to which the door or window belongs. Shape continuity features are used in the calculation of shape continuity residuals.

[0144] In one embodiment, for the end connection type, the component instance... The position of the continuation endpoint is denoted as , will component instance The position of the continuation endpoint is denoted as Then the end-connection residual corresponding to the end-connection type can be expressed as:

[0145] in, Representing a component instance With component instances End-joint residuals between; Representing a component instance The corresponding connection endpoint position; Representing a component instance The corresponding connection endpoint position; Represents the Euclidean norm. End-joint residuals. This refers to the shape continuation residual corresponding to the end connection type.

[0146] For containment connection types, multiple boundary points are sampled from the outer boundary of the contained component, and the distance of each boundary point beyond the shape containment range of the containing component is calculated. The contained component is denoted as... The component is denoted as The first one will be included on the outer boundary of the component. The boundary points are denoted as , will contain components The corresponding shape tolerance range is denoted as Then, the included connection residual corresponding to the included connection type can be expressed as:

[0147] in, Representing a component instance With component instances The inclusion of continuation residuals between them; Indicates the number of boundary points selected from the outer boundary of the contained component; Representing a component instance The first on the outer boundary One boundary point; Representing a component instance The corresponding shape range; Represents boundary points Beyond the shape's capacity The distance when the boundary point Within the shape accommodation range Within, distance Pick Includes continuation residuals. As the shape continuation residual corresponding to the connection type.

[0148] Based on the component connection type, end joint residuals or components containing joint residuals are identified as component instances. With component instances Shape continuity residuals Shape continuity residual The same length unit is used as the shape joint tolerance. The shape joint state can be represented as:

[0149] in, Representing a component instance With component instances The shape continuity between them; The spatial adjacency relationship of the corresponding components is verified by shape continuity. This indicates that the spatial adjacency relationship of the corresponding component failed the shape continuity check. Representing a component instance With component instances Shape continuity residuals between; This indicates the shape continuity tolerance corresponding to the component connection type and the accuracy requirements of 3D shape reconstruction. Shape Continuity Status It participates in the generation of shape continuity verification results.

[0150] When summarizing the shape continuity status of each component to be verified, the spatial adjacency relationship, point cloud unit, current 3D shape parameters, and current processing path of the components that failed the shape continuity verification are recorded to generate shape continuity verification results. The shape continuity verification results are used to distinguish between verification failures caused by fitted 3D shape parameters and verification failures caused by generative 3D shape parameters, so that subsequent processing path adjustments and component type failure status statistics can be performed respectively.

[0151] In one exemplary embodiment, the shape continuity tolerance can be configured according to the three-dimensional shape reconstruction accuracy requirements (e.g., configured as follows). (meters). When the end-joining residual between adjacent wall endpoints exceeds the shape join tolerance, the shape join status of the corresponding wall is determined to be unsuccessful; when the outer boundary of a door or window exceeds the shape tolerance of its respective wall, the shape join status between the corresponding door / window and its respective wall is determined to be unsuccessful. The above values ​​are used to explain the shape join verification process, and in actual implementation, they are configured according to the accuracy requirements of 3D shape reconstruction.

[0152] Optionally, adjusting the processing path of the point cloud unit of the failed component based on the shape continuity verification result includes: Based on the shape continuity verification results, identify the fitted 3D shape parameters that failed the shape continuity verification, and then trace the component point cloud units corresponding to the fitted 3D shape parameters that failed the shape continuity verification to generate component point cloud units that failed the fitting verification. The processing path for the point cloud units of the component that failed the fitting verification was changed from the geometric fitting path to the inference generation path, so as to generate the fitting branch adjustment results. Based on the shape continuity verification results, identify the generative 3D shape parameters that fail the shape continuity verification, and then trace the component point cloud units corresponding to the generative 3D shape parameters that fail the shape continuity verification to generate the component point cloud units that fail the verification. The generation inference path corresponding to the point cloud unit of the failed component is maintained to generate the generation branch hold result. The fitting branch adjustment result and the generation branch hold result are combined to generate the processing path adjustment result, so that the processing path adjustment result participates in the generation of updating three-dimensional shape parameters.

[0153] Specifically, in this application, both the fitted 3D shape parameters and the generative 3D shape parameters are associated with the component point cloud units that generate the corresponding 3D shape parameters and the processing path. When identifying fitted 3D shape parameters that fail the shape continuity check based on the shape continuity check results, the corresponding component point cloud units are traced along the association between the fitted 3D shape parameters and the component point cloud units, and the corresponding component point cloud units are identified as component point cloud units that failed the fitting check. Component point cloud units that failed the fitting check participate in the generation of fitting branch adjustment results.

[0154] The point cloud units of the components that failed fitting verification were identified as geometrically fitted point cloud units during the point cloud coverage confidence division stage. However, the fitted 3D shape parameters failed the shape continuity verification under the constraint of component spatial adjacency, indicating that although the currently fitted 3D shape parameters can form a relatively low shape fitting residual with the local point cloud, there is still a mismatch at the connection positions of adjacent components. Therefore, the processing path for the point cloud units of the components that failed fitting verification was adjusted from the geometric fitting path to the generation of the inference path, and the generation of the inference path was recorded as the target processing path in the fitting branch adjustment results. The fitting branch adjustment results are used in the generation of the processing path adjustment results.

[0155] For generative 3D shape parameters that fail shape continuity verification, the corresponding generative inference point cloud units are traced along the association between the generative 3D shape parameters and the generative inference point cloud units, and these units are identified as generative inference point cloud units that failed the verification. These generative inference point cloud units continue to maintain the generative inference path and do not re-enter the geometric fitting path. The generation branch retains the result records of the component type and shape continuity residual corresponding to the generative inference point cloud units that failed the verification, enabling these units to participate in the generation of component type failure states. The generation branch retains the result in the generation of the processing path adjustment results.

[0156] The results of the fitting branch adjustment and the generation branch hold-forward are combined to generate the processing path adjustment results. These results include the point cloud units of the failed verification components, the current processing path corresponding to each unit, and the target processing path used when regenerating the 3D shape parameters. The processing path adjustment results participate in the generation of updated 3D shape parameters, ensuring that both the fitted and generated point cloud units of the failed verification components enter the 3D shape generation network. However, their participation in the 3D shape generation network update differs: the fitted point cloud units use the current 3D shape generation network to generate updated 3D shape parameters; the generated point cloud units, after triggering a 3D shape generation network update due to component type failure, use the updated 3D shape generation network to regenerate updated 3D shape parameters.

[0157] In a preferred embodiment, the processing path adjustment result can also retain the shape continuity residuals of the failed component point cloud units in each round of shape continuity verification. If the shape continuity residual of the same failed component point cloud unit continues to decrease after multiple rounds of updates, the target processing path recorded in the processing path adjustment result will continue to be used; if the shape continuity verification still fails when the verification termination condition is met, the updated three-dimensional shape parameters with lower shape continuity residuals in each round of updated three-dimensional shape parameters will be retained, and the retained updated three-dimensional shape parameters will participate in the generation of the verification confirmation three-dimensional shape parameter set.

[0158] Optionally, associating the candidate component type set with the corresponding component point cloud unit to determine the component type corresponding to the component point cloud unit includes: The corresponding component point cloud units are fitted according to each candidate component type in the candidate component type set to generate candidate fitted 3D shape parameters. The 3D shape generation network is then used to process the corresponding component point cloud units according to each candidate component type to generate candidate generative 3D shape parameters. Combine candidate fitted 3D shape parameters and candidate generative 3D shape parameters to generate a set of candidate 3D shape parameters; Based on each candidate 3D shape parameter in the candidate 3D shape parameter set, the candidate 3D surface point cloud is reconstructed. The candidate 3D surface point cloud is compared with the corresponding component point cloud unit to generate the consistency of candidate point cloud reconstruction. The candidate 3D shape parameters are verified according to the spatial adjacency relationship of the component to generate the consistency of candidate shape continuity. Combining the consistency of candidate point cloud reconstruction and the consistency of candidate shape continuity, each candidate 3D shape parameter in the candidate 3D shape parameter set is filtered, and the filtered candidate 3D shape parameters are determined as category-confirmed 3D shape parameters. The category-confirmed component type is determined according to the candidate component type corresponding to the category-confirmed 3D shape parameter, and then the component type corresponding to the component point cloud unit is updated using the category-confirmed component type.

[0159] Specifically, in this application, when the component type confidence level cannot uniquely determine the component type corresponding to the component point cloud unit, the component type is not determined directly based on the highest component type confidence level. Instead, each candidate component type in the candidate component type set participates in the generation of candidate three-dimensional shape parameters.

[0160] For any candidate component type in the candidate component type set, the corresponding component shape primitive and component shape prior are read according to that candidate component type. The component shape primitive is fitted to the corresponding component point cloud unit to generate candidate fitted 3D shape parameters; simultaneously, the corresponding component point cloud unit is encoded as point cloud conditional features, and candidate generative 3D shape parameters are generated using the target component type-specific determination head corresponding to that candidate component type. Each candidate component type corresponds to candidate fitted 3D shape parameters and candidate generative 3D shape parameters, and the combination of candidate fitted 3D shape parameters and candidate generative 3D shape parameters forms a candidate 3D shape parameter set.

[0161] Candidate point cloud reconstruction consistency is determined based on the point cloud reprojection residual between the candidate 3D surface point cloud and the corresponding component point cloud unit. The smaller the point cloud reprojection residual, the higher the candidate point cloud reconstruction consistency. Candidate shape continuity consistency is determined based on the shape continuity residual between the candidate 3D shape parameters and the current 3D shape parameters of adjacent components. The smaller the shape continuity residual, the higher the candidate shape continuity consistency. Candidate point cloud reconstruction consistency and candidate shape continuity consistency jointly participate in the selection of candidate 3D shape parameters.

[0162] Let the candidate component type be denoted as Candidate component types The corresponding point cloud reprojection residual is denoted as Candidate component types The corresponding shape continuity residual is denoted as Then the joint evaluation value corresponding to the candidate component type can be expressed as:

[0163] in, Indicates candidate component type The corresponding joint evaluation quantity; Indicates candidate component type The corresponding point cloud reprojection residual; This represents the point cloud reprojection tolerance corresponding to the accuracy of 3D point cloud acquisition. Indicates candidate component type Corresponding shape continuation residual; Indicates the tolerance for shape continuity; This represents the evaluation weight corresponding to the consistency of candidate point cloud reconstruction; This represents the evaluation weight corresponding to the consistency of candidate shape succession; evaluation weight and evaluation weight Both are non-negative numbers, and their sum is Joint evaluation quantity Used to filter candidate component types and candidate 3D shape parameters.

[0164] From the set of candidate component types, select the candidate component type with the lowest joint evaluation value. Then, among the candidate fitted 3D shape parameters and candidate generative 3D shape parameters corresponding to this component type, select the candidate 3D shape parameter with the lowest combination of point cloud reprojection residual and shape continuity residual as the category confirmation 3D shape parameter. The candidate component type corresponding to the category confirmation 3D shape parameter is determined as the category confirmation component type, and then the component type corresponding to the component point cloud unit is updated using the category confirmation component type. The updated component type participates in subsequent point cloud coverage confidence determination, 3D shape parameter generation, and shape continuity verification.

[0165] This processing method determines the component type corresponding to a component point cloud unit not only based on the semantic features of the local point cloud, but also based on whether the candidate component type can form candidate 3D shape parameters that match the observed component point cloud units and are reasonably connected with adjacent components. For example, for a component point cloud unit whose shape is between a load-bearing wall pier and a decorative pilaster, candidate 3D shape parameters can be generated according to the load-bearing wall pier and the decorative pilaster, respectively. Then, the category is determined and the component type is confirmed based on the consistency of candidate point cloud reconstruction and the consistency of candidate shape continuity.

[0166] Optionally, the step of generating component type failure states according to component type and updating the 3D shape generation network based on component type failure states includes: Based on the shape continuity verification results, identify the generative 3D shape parameters that fail the shape continuity verification, and then trace the generative inference point cloud units corresponding to the generative 3D shape parameters that fail the shape continuity verification to generate point cloud units of the generative verification failed components. The frequency of verification failures of point cloud units of failed components is generated according to component type, so as to generate component type failure status, and the active distillation component type is determined based on the component type failure status. Extract historical point cloud units of the same type as the active distillation component from a pre-built historical point cloud sample library, and combine the point cloud units of the active distillation component type that failed to generate verification with the historical point cloud units of the same type to generate active distillation samples. The teacher multi-step flow model is used to perform state transformations based on active distillation samples to generate supplementary 3D shape state transformation trajectories. The component type-specific determination head corresponding to the active distillation component type is updated based on the supplementary 3D shape state transformation trajectory to generate an updated 3D shape generation network.

[0167] Specifically, in this application, the component type failure status is used to indicate the degree to which generative 3D shape parameters corresponding to the same component type fail the shape continuity verification. After determining the generated point cloud units of the failed component based on the shape continuity verification results, the number of generated point cloud units of the failed component is counted according to the component type corresponding to each generated point cloud unit, and the total number of generated inferred point cloud units corresponding to the corresponding component type participating in the current modeling task is also counted. The number of generated point cloud units of the failed component and the total number of generated inferred point cloud units are jointly used in the calculation of the component type failure frequency.

[0168] The component type is denoted as , component type The number of point cloud units of the corresponding failed generation verification component is denoted as , component type The total number of corresponding generated inference point cloud units is denoted as The failure frequency of component type can then be expressed as:

[0169] in, Indicate component type The corresponding component type failure frequency; Indicate component type The corresponding number of point cloud units of the generated verification failed; Indicate component type The total number of corresponding generated and inferred point cloud units. Component type failure frequency. Determine the failure status of the component type and the active distillation component type.

[0170] Let the active distillation trigger threshold be denoted as When the component type fails frequently Greater than the active distillation trigger threshold When, the component type Determined as an active distillation component type; when the component type failure frequency... Not greater than the active distillation trigger threshold At that time, retain component type The corresponding component type has its own unique header, which is not specific to the component type. Perform active distillation update. Active distillation trigger threshold. Configured based on the distribution of verification failures for different component types in historical modeling tasks, and used to focus active distillation updates on component types where verification failures are more concentrated.

[0171] The component type failure status includes the component type failure frequency, the shape continuity residual distribution corresponding to the point cloud units of the failed component, and the failure change status in consecutive verification rounds. The component type failure frequency is used to determine whether verification failures occur concentratedly within the same component type; the shape continuity residual distribution is used to indicate the degree of verification failure; and the failure change status in consecutive verification rounds is used to indicate the change in the shape continuity status of the corresponding component type after the 3D shape generation network is updated. The component type failure status collectively participates in the determination of the active distillation component type.

[0172] Specifically, in this application, the historical point cloud sample library is formed by collecting point cloud units of components whose component types have been confirmed and whose verified 3D shape parameters have been generated in historical modeling tasks. Each historical component point cloud unit in the historical point cloud sample library is associated with the component type, the component shape prior, and the corresponding verified 3D shape parameters. For the active distillation component type, historical component point cloud units of the same type as the active distillation component are extracted from the historical point cloud sample library, and priority is given to historical component point cloud units of the same type whose point cloud surface coverage, point cloud distribution, or component shape change is close to that of the generated verification failed component point cloud units.

[0173] Active distillation samples are generated by combining point cloud units of failed component verifications corresponding to the active distillation component type with point cloud units of the same historical type. The generated point cloud units of failed component verifications reflect the actual shape continuity deviations that occur in the current modeling task; the point cloud units of the same historical type are used to supplement the shape variation range of the corresponding component type, reducing the parameter offset caused by updating the component type-specific determination head only around a small number of failed component verification point cloud units. Active distillation samples participate in supplementing the generation of 3D shape state transformation trajectories.

[0174] When the teacher multi-step flow model performs state transformations based on active distillation samples, point cloud conditional features and 3D shape generation conditions are generated for each component point cloud unit in the active distillation samples. The teacher multi-step flow model then performs multi-step state transformations to generate supplementary 3D shape state transformation trajectories. These supplementary 3D shape state transformation trajectories are divided according to the aforementioned trajectory change states and component type complexity to generate supplementary jump flow intervals. Supplementary jump flow distillation supervision data is then generated based on these intervals. This supplementary jump flow distillation supervision data is used in the calculation of the active distillation loss.

[0175] During the active distillation update process, the network parameters of the shared shape decoding backbone and the component type-specific determination headers corresponding to non-actively distilled component types remain unchanged; only the component type-specific determination headers corresponding to the actively distilled component types are updated. The network parameters of the shared shape decoding backbone are denoted as... Active distillation component type The corresponding component type-specific header parameters are denoted as follows: Then active distillation update can be represented as:

[0176]

[0177] in, Indicates the first At the start of each active distillation update cycle, the active distillation component type The header parameters are specifically defined for each component type. Indicates the first Each active distillation update cycle corresponds to a unique header parameter specific to the component type. Indicates the step size for parameter updates during active distillation; This indicates the active distillation loss relative to the component type-specific determined head parameters. The gradient; This represents the active distillation loss calculated based on supplemental skip distillation monitoring data; This indicates the network parameters of the shared shape decoding backbone before active distillation updates; This represents the network parameters that share the shape decoding backbone after active distillation updates. The above formula indicates that active distillation updates only change the component-specific determination header parameters corresponding to the active distillation component type. Network parameters of the shared shape decoding backbone It remains unchanged.

[0178] Active distillation loss The method of combining shape reconstruction error and flow matching consistency error is adopted, and calculations are performed only for active distillation samples corresponding to active distillation component types. After the component type-specific determination header parameters are updated, an updated 3D shape generation network is generated. The updated 3D shape generation network retains the ability of the original 3D shape generation network to determine 3D shape parameters for other component types, while adjusting the 3D shape parameter mapping corresponding to the active distillation component type. The updated 3D shape generation network is used to regenerate the updated 3D shape parameters corresponding to the point cloud units of the component that failed the generation verification.

[0179] In one exemplary embodiment, when the failure frequency of the component type corresponding to the irregular corner window reaches or exceeds the active distillation trigger threshold (e.g., the failure frequency of the component type is...), The active distillation trigger threshold is The irregular corner window is identified as the active distillation component type. Active distillation samples can include point cloud units of generated and validated failed components from the current modeling task and historical point cloud units of the same type from the historical point cloud sample library (e.g., including 13 generated and validated failed component point cloud units and 40 historical component point cloud units of the same type). After generating supplementary 3D shape state transformation trajectories using the teacher's multi-step flow model, only the component type-specific determination header corresponding to the irregular corner window is updated; the shared shape decoding backbone and other component type-specific determination headers remain unchanged. The above values ​​are used to illustrate the active distillation update process; in actual implementation, the configuration is based on the component type failure status.

[0180] Optionally, the step of generating updated 3D shape parameters for the failed component point cloud units from the updated 3D shape generation network, iteratively verifying the updated 3D shape parameters to generate a verified 3D shape parameter set, and then continuously verifying the verified 3D shape parameter set according to the spatial adjacency relationship of the components to generate a point cloud parameterized 3D model includes: Based on the processing path adjustment results, the fitting verification failed component point cloud units and the generation verification failed component point cloud units are determined. The fitting verification failed component point cloud units and the generation verification failed component point cloud units are input into the updated 3D shape generation network to generate updated 3D shape parameters. The fitted 3D shape parameters corresponding to the point cloud units of the failed fitting verification component and the generative 3D shape parameters corresponding to the point cloud units of the failed generation verification component are replaced with updated 3D shape parameters to generate an updated set of 3D shape parameters to be verified. Repeat the shape continuity check on the updated set of three-dimensional shape parameters to be checked until the updated set of three-dimensional shape parameters to be checked meets the check termination condition, so as to generate a set of three-dimensional shape parameters to be checked and confirmed. Based on the 3D point cloud, the point cloud coordinate domain is determined, and the verified 3D shape parameter set is transformed into the point cloud coordinate domain to generate a coordinate-unified 3D shape parameter set. The coordinate-unified 3D shape parameter set is then continued according to the spatial adjacency relationship of the components to generate a point cloud parameterized 3D model.

[0181] Specifically, in this application, the updated 3D shape generation network described in the claims corresponds to the aforementioned updated 3D shape generation network. To maintain consistency in terminology in the subsequent technical description, the term "updated 3D shape generation network" will be used uniformly below. The processing path adjustment result records the fitting verification failed component point cloud units and the generation verification failed component point cloud units. For fitting verification failed component point cloud units, the fitting verification failed component point cloud units are encoded to generate point cloud conditional features, and the point cloud conditional features are input into the updated 3D shape generation network to generate the corresponding updated 3D shape parameters. For generation verification failed component point cloud units, after the active distillation update is completed, the point cloud conditional features corresponding to the generation verification failed component point cloud units are input into the updated 3D shape generation network to generate the corresponding updated 3D shape parameters.

[0182] The updated 3D shape parameters are associated one-to-one with the point cloud units of the failed component. When replacing the fitted or generative 3D shape parameters with the updated 3D shape parameters, only the 3D shape parameters corresponding to the point cloud units of the failed component are replaced; the remaining fitted and generative 3D shape parameters that have passed shape continuity verification remain unchanged. After the replacement is complete, all currently valid 3D shape parameters of all components are combined into a set of updated 3D shape parameters to be verified. This updated set of 3D shape parameters then participates in subsequent shape continuity verification.

[0183] When repeatedly performing shape continuity verification on the updated set of 3D shape parameters to be verified, the component pairs to be verified are still determined according to the spatial adjacency relationship of the components, and the shape continuity residuals are calculated according to the component connection type. When the spatial adjacency relationships of all components in the set of 3D shape parameters to be verified are updated, they satisfy... At that time, the set of three-dimensional shape parameters to be verified is updated through shape continuity verification.

[0184] The verification termination conditions include all component spatial adjacency relationships passing shape continuity verification, or the number of shape continuity verification executions reaching the configured maximum number of iterations. The maximum number of iterations is denoted as... Record the current shape continuation verification round as Then the verification termination condition can be expressed as:

[0185] in, Indicates the first Component examples in wheel shape continuation verification With component instances Shape continuity residuals between; Indicates the tolerance for shape continuity; Indicates the current shape continuation verification round; Indicates the maximum number of iterations; Represents each pair of component instances in the spatial adjacency relationship of components; This indicates that the iterative verification will terminate when either of the two verification termination conditions is met. The verification termination condition is used to determine when to generate the set of verified 3D shape parameters.

[0186] When the maximum number of iterations is reached but a small number of component spatial adjacency relationships still fail the shape continuity verification, the updated 3D shape parameters corresponding to the same component point cloud unit in each round of shape continuity verification are compared. The updated 3D shape parameter with the lower combination of shape fitting residual, point cloud reprojection residual, and shape continuity residual is selected as the verification confirmed 3D shape parameter for the corresponding component point cloud unit. The verification confirmed 3D shape parameters corresponding to each component point cloud unit are combined to form a verification confirmed 3D shape parameter set.

[0187] Specifically, in this application, the point cloud coordinate domain is determined by the origin, three coordinate axes, and length unit of the 3D point cloud. For fitted 3D shape parameters, the component positions are located within the point cloud coordinate domain; for generative and updated 3D shape parameters, the component positions and orientations are transformed to the point cloud coordinate domain based on the coordinate transformation relationship between the component local coordinate domain and the point cloud coordinate domain of the generated inferred point cloud unit. After the coordinate transformation is completed, the verified 3D shape parameters are combined into a unified coordinate 3D shape parameter set.

[0188] When unifying the 3D shape parameter set according to the spatial adjacency relationship of components, the shape continuity features corresponding to each component are read according to the component connection type, and adjacent components are assembled and positioned based on the shape continuity residual. For end connection types, the connection endpoints of adjacent components are aligned within the shape continuity tolerance range; for containment connection types, the outer boundary of the contained component is located within the shape containment range of the containing component. The unified 3D shape parameter set after assembly and positioning participates in the generation of the point cloud parametric 3D model.

[0189] In a preferred embodiment, if adjacent components that have passed shape continuity verification still have residual shape gaps at the connection position that are smaller than the shape continuity tolerance, then position adjustment amounts are allocated based on the reliability of the three-dimensional shape parameters of the components on both sides. The reliability of the three-dimensional shape parameters is determined based on the point cloud coverage confidence, shape fitting residual, or point cloud reprojection residual. Components with higher point cloud coverage confidence or lower point cloud reprojection residuals are given smaller position adjustment amounts, while components on the other side are given larger position adjustment amounts. After adjusting the component positions of adjacent components using the position adjustment amounts, the shape continuity residuals are recalculated, and the adjustment results are retained as long as the adjusted shape continuity residuals are not greater than the original shape continuity residuals.

[0190] Each 3D shape parameter in the unified coordinate 3D shape parameter set is converted into a corresponding parametric geometry. The connection, adjacency, and containment relationships between these parametric geometries are recorded according to the spatial adjacency relationships of the components to generate a point cloud parametric 3D model. Each component in the point cloud parametric 3D model retains its component type, position, size, and orientation, and is associated with the processing path information of the corresponding component's point cloud unit. The processing path information indicates whether the 3D shape parameters of the corresponding component originate from a geometric fitting path, a generation inference path, a processing path adjustment, or a candidate component type confirmation.

[0191] In one exemplary embodiment, the method of this application is applied to a 3D point cloud of an existing industrial building. The 3D point cloud is segmented to form multiple component point cloud units (e.g., approximately 6800 component point cloud units) corresponding to walls, beams, columns, and doors / windows. After segmentation according to point cloud coverage confidence, most component point cloud units enter the geometric fitting path, while those with significant point cloud occlusion enter the generation inference path. After geometric fitting, component point cloud units whose shape fitting residuals exceed the shape fitting tolerance are redefined as generation inference point cloud units. A teacher multi-step flow model generates 3D shape state transformation trajectories for the generation inference point cloud units and trains the 3D shape generation network through jump flow distillation.

[0192] After the fitted 3D shape parameters and the generative 3D shape parameters are formed, shape continuity verification is performed according to the spatial adjacency relationship of the components. Point cloud units of components that fail the fitting verification during shape continuity verification are transferred to the generation inference path, and these point cloud units are assigned a component type failure state based on the component type. When the component type failure state reaches the active distillation trigger threshold, the component type-specific determination header corresponding to the active distillation component type is updated, and the updated 3D shape parameters are regenerated using the updated 3D shape generation network. The updated 3D shape parameters participate in shape continuity verification again until a verified 3D shape parameter set is formed. Finally, the verified 3D shape parameter set is transformed to the point cloud coordinate domain to generate a coordinate-unified 3D shape parameter set, and the coordinate-unified 3D shape parameter set is continued according to the spatial adjacency relationship of the components to generate a point cloud parameterized 3D model.

[0193] Without contradiction, the three-dimensional shape state transformation trajectory division method, skip flow distillation loss, shared shape decoding backbone, component type-specific determination head, shape continuity verification method, processing path adjustment method, candidate component type confirmation method, active distillation update method, and iterative verification method in the above embodiments can be combined with each other. The numerical values ​​in each embodiment are only used to explain the corresponding processing procedures. In actual implementation, they can be configured according to the acquisition density of the three-dimensional point cloud, the distribution of component types, the complexity of component shapes, and the accuracy requirements of three-dimensional shape reconstruction.

[0194] like Figure 2 As shown, this embodiment of the present application provides a point cloud to parameterized model generation device based on joint optimization of skip flow distillation and confidence level, which includes: The point cloud branching module is used to obtain the spatial adjacency relationship between the 3D point cloud and the component, segment the 3D point cloud to generate component point cloud units, determine the point cloud coverage confidence based on the component point cloud units, and divide the component point cloud units into geometric fitting point cloud units and generated inference point cloud units according to the point cloud coverage confidence. The 3D shape parameter generation module is used to fit geometric fitting point cloud units to generate fitted 3D shape parameters, and uses the teacher multi-step flow model to generate 3D shape state transformation trajectories for the generated inference point cloud units. Based on the 3D shape state transformation trajectory, jump flow distillation is performed to generate a 3D shape generation network, and the generative 3D shape parameters are determined using the 3D shape generation network. The joint optimization module is used to jointly verify the fitted 3D shape parameters and generative 3D shape parameters according to the spatial adjacency relationship of the components to generate shape continuity verification results, adjust the processing path of the point cloud units of the failed components according to the shape continuity verification results, generate the component type failure status according to the component type, and update the 3D shape generation network according to the component type failure status. The parameterized model generation module is used to generate updated 3D shape parameters for point cloud units of failed components using the updated 3D shape generation network, iteratively verify and update the 3D shape parameters to generate a set of verified 3D shape parameters, and continue to verify and confirm the 3D shape parameter set according to the spatial adjacency relationship of the components to generate a point cloud parameterized 3D model.

[0195] like Figure 3 As shown, an electronic device according to an embodiment of this application includes a memory and a processor. The memory stores a computer program. When the processor executes the computer program, it implements the point cloud to parameterized model generation method based on jump flow distillation and confidence joint optimization as described in any one of this application.

[0196] like Figure 4 As shown, a computer-readable storage medium is provided in an embodiment of this application. The computer-readable storage medium stores a computer program. When the computer program is executed by a processor, it implements the point cloud to parameterized model generation method based on jump flow distillation and confidence joint optimization as described in any one of this application.

[0197] like Figure 5As shown in the figure, this application discloses a point cloud to parameterized model generation system based on jump flow distillation and confidence joint optimization. The system includes a point cloud acquisition device, a parameterized model generation device, and a model output device. The point cloud acquisition device is communicatively connected to the parameterized model generation device, and the parameterized model generation device is communicatively connected to the model output device. The point cloud acquisition device acquires a 3D point cloud in a target space and sends the 3D point cloud to the parameterized model generation device. The parameterized model generation device acquires the 3D point cloud and determines the spatial adjacency relationship of components, segments the 3D point cloud to generate component point cloud units, determines the point cloud coverage confidence based on the component point cloud units, and divides the component point cloud units into geometric fitting point cloud units and generative inference point cloud units according to the point cloud coverage confidence. The parameterized model generation device is also used to fit the geometric fitting point cloud units to generate fitted 3D shape parameters, uses a teacher multi-step flow model to generate 3D shape state transformation trajectories for the generative inference point cloud units, and generates 3D shape state transformation trajectories based on the 3D shape state transformation trajectories. The device performs a jump flow distillation to generate a 3D shape generation network, and uses the 3D shape generation network to determine generative 3D shape parameters. The parameterized model generation device is also used to jointly verify the fitted 3D shape parameters and the generative 3D shape parameters according to the component spatial adjacency relationship to generate shape continuity verification results. Based on the shape continuity verification results, the processing path of the verification-failed component point cloud units is adjusted, a component type failure state is generated according to the component type, and the 3D shape generation network is updated according to the component type failure state. Then, the updated 3D shape generation network is used to generate updated 3D shape parameters for the verification-failed component point cloud units. The parameterized model generation device is also used to iteratively verify the updated 3D shape parameters to generate a verified 3D shape parameter set, and connect the verified 3D shape parameter set according to the component spatial adjacency relationship to generate a point cloud parameterized 3D model. The point cloud parameterized 3D model is then sent to the model output device. The model output device is used to receive and output the point cloud parameterized 3D model.

[0198] Figures 2-5 For an exemplary description, please refer to the above. Figure 1 This will not be elaborated upon here.

Claims

1. A method for generating point cloud-to-parameterized models based on joint optimization of skip flow distillation and confidence level, characterized in that, include: Obtain the spatial adjacency relationship between the 3D point cloud and the component, segment the 3D point cloud to generate component point cloud units, determine the point cloud coverage confidence based on the component point cloud units, and divide the component point cloud units into geometric fitting point cloud units and generation inference point cloud units according to the point cloud coverage confidence. The fitting geometric fitting point cloud unit generates the fitting three-dimensional shape parameters; the teacher multi-step flow model generates the three-dimensional shape state transformation trajectory for the generated inference point cloud unit, and the jump flow distillation is performed based on the three-dimensional shape state transformation trajectory to generate the three-dimensional shape generation network, and the generative three-dimensional shape parameters are determined by the three-dimensional shape generation network. The shape continuity verification results are generated by jointly verifying the fitted 3D shape parameters and generative 3D shape parameters according to the spatial adjacency relationship of the components. The processing path of the point cloud unit of the failed component is adjusted according to the shape continuity verification results. The component type failure status is generated according to the component type, and the 3D shape generation network is updated according to the component type failure status. The updated 3D shape parameters of the point cloud units of the failed component are generated by the updated 3D shape generation network. The updated 3D shape parameters are iteratively verified to generate a set of verified 3D shape parameters. The set of verified 3D shape parameters is then continuously verified according to the spatial adjacency relationship of the components to generate a parameterized 3D model of the point cloud.

2. The point cloud to parameterized model generation method based on joint optimization of skip flow distillation and confidence level according to claim 1, characterized in that, The segmented 3D point cloud generating component point cloud unit includes: The semantic features of each point cloud point are extracted from the 3D point cloud using a component semantic segmentation network, and the spatial adjacency relationship of the point cloud is determined based on the spatial position of each point cloud point in the 3D point cloud. Point cloud points with the same semantic features and spatial adjacency are aggregated according to the spatial adjacency relationship of the point cloud to generate component point cloud units; The component type confidence score is generated from the semantic features of the point cloud contained in the component point cloud unit, and the component type corresponding to the component point cloud unit is determined based on the component type confidence score; When the component type confidence score cannot uniquely determine the component type corresponding to the component point cloud unit, a set of candidate component types is generated based on the component type confidence score. The set of candidate component types is then associated with the corresponding component point cloud unit to determine the component type corresponding to the component point cloud unit.

3. The point cloud to parameterized model generation method based on joint optimization of skip flow distillation and confidence level according to claim 2, characterized in that, The determination of point cloud coverage confidence based on component point cloud units includes: Read the component shape prior based on the component type corresponding to the component point cloud unit; Compare the component point cloud units with the component shape priors to generate the point cloud surface coverage state, point cloud distribution state, and shape feature support state; The point cloud surface coverage state, point cloud distribution state, and shape feature support state are integrated to generate a point cloud coverage state, and then the point cloud coverage confidence degree corresponding to the component point cloud unit is determined based on the point cloud coverage state.

4. The point cloud to parameterized model generation method based on joint optimization of skip flow distillation and confidence level according to claim 3, characterized in that, The method of dividing the component point cloud unit into geometrically fitted point cloud unit and generated inference point cloud unit according to the point cloud coverage confidence includes: Obtain the 3D shape reconstruction accuracy requirements configured for the parameterized 3D model of the point cloud to be generated, and generate the point cloud coverage confidence threshold based on the component type corresponding to the component point cloud unit and the 3D shape reconstruction accuracy requirements; The confidence score of the point cloud coverage corresponding to the point cloud unit of the component is compared with the confidence score threshold of the point cloud coverage to generate a confidence score comparison result. Based on the confidence level comparison results, component point cloud units whose point cloud coverage confidence level reaches the point cloud coverage confidence level threshold are identified as geometrically fitted point cloud units, while component point cloud units whose point cloud coverage confidence level does not reach the point cloud coverage confidence level threshold are identified as generated inference point cloud units.

5. The point cloud to parameterized model generation method based on joint optimization of skip flow distillation and confidence level according to claim 4, characterized in that, The fitted geometric point cloud unit generates fitted 3D shape parameters, including: The component shape primitives and shape fitting tolerances are determined based on the component type corresponding to the geometrically fitted point cloud units and the accuracy requirements for 3D shape reconstruction. The component shape primitives are fitted to the geometric fitting point cloud units to generate shape primitive parameters, and then the fitted three-dimensional shape parameters are determined based on the shape primitive parameters. Calculate the shape fitting residual between the fitted 3D shape parameters and the corresponding geometric fitted point cloud units, and compare the shape fitting residual with the shape fitting tolerance to generate the fitting residual comparison result. Based on the fitting residual comparison results, geometrically fitted point cloud units whose shape fitting residuals exceed the shape fitting tolerance are redefined as generative inference point cloud units, so that the redefined generative inference point cloud units can participate in the determination of generative 3D shape parameters.

6. The point cloud to parameterized model generation method based on joint optimization of skip flow distillation and confidence level according to claim 4, characterized in that, The process of generating a 3D shape state transformation trajectory for the generated inference point cloud unit using the teacher's multi-step flow model includes: Encode and generate inference point cloud units to generate point cloud condition features, and read the component shape prior based on the component type corresponding to the generated inference point cloud units; By fusing point cloud conditional features and component shape priors, three-dimensional shape generation conditions are generated. Generate the initial state of the three-dimensional shape, and input the initial state of the three-dimensional shape and the conditions for generating the three-dimensional shape into the teacher's multi-step flow model so that the teacher's multi-step flow model can gradually update the initial state of the three-dimensional shape and generate multiple three-dimensional shape states. Multiple 3D shape states are arranged according to their update order to generate the 3D shape state transformation trajectory corresponding to the generated inference point cloud unit.

7. The point cloud to parameterized model generation method based on joint optimization of skip flow distillation and confidence level according to claim 6, characterized in that, The process of progressively updating the initial state of the 3D shape and generating multiple 3D shape states using a multi-step flow model by the teacher includes: For the same point cloud unit, multiple initial states of three-dimensional shapes are generated. The teacher multi-step flow model is used to perform state transformation according to each initial state of three-dimensional shape and the three-dimensional shape generation conditions to generate candidate three-dimensional shape state transformation trajectories. The termination state of each candidate 3D shape state transformation trajectory is converted into candidate 3D shape parameters, and the candidate 3D surface point cloud is reconstructed based on the candidate 3D shape parameters. The point cloud coordinate domain is determined based on the 3D point cloud, and each candidate 3D surface point cloud is mapped to the point cloud coordinate domain in order to calculate the point cloud reprojection residual between each candidate 3D surface point cloud and the corresponding generated inference point cloud unit. Candidate 3D shape state transformation trajectories are selected based on the point cloud reprojection residuals, and then the candidate 3D shape state transformation trajectory with the smallest point cloud reprojection residual is determined as the 3D shape state transformation trajectory corresponding to the generated inferred point cloud unit.

8. The point cloud to parameterized model generation method based on joint optimization of skip flow distillation and confidence level according to claim 6, characterized in that, The method of performing jump flow distillation based on the three-dimensional shape state transformation trajectory to generate a three-dimensional shape generation network includes: Compare adjacent three-dimensional shape states in the three-dimensional shape state transformation trajectory to generate trajectory change states; The shape parameter dimension is determined based on the component type corresponding to the generated inferred point cloud unit, and the component shape change degree is determined based on the component shape prior. The shape parameter dimension and the component shape change degree are fused to generate the component type complexity. The trajectory of three-dimensional shape state transformation is divided according to the trajectory change state and component type complexity to generate jump flow intervals; The jump start state and jump end state are determined based on the jump flow interval. The jump start state and jump end state are associated to generate jump flow distillation supervision data, which is then used to participate in the jump flow distillation training of the 3D shape generation network.

9. The point cloud to parameterized model generation method based on joint optimization of skip flow distillation and confidence level according to claim 8, characterized in that, The jump flow distillation supervision data participates in the jump flow distillation training of the 3D shape generation network, including: An initial 3D shape generation network is constructed, which combines the jump start state, point cloud condition features, and the component type corresponding to the generated inferred point cloud unit to generate the jump flow distillation input, and the jump termination state is determined as the jump flow distillation target. The jump flow distillation input is fed into the initial 3D shape generation network to generate jump state prediction results. The jump state prediction results are compared with the jump flow distillation target to generate shape reconstruction error. The jump state prediction results corresponding to adjacent jump flow intervals are continued according to the state transformation order of the three-dimensional shape state transformation trajectory to generate cross-interval prediction trajectory. Then, the cross-interval prediction trajectory and the three-dimensional shape state transformation trajectory are compared to generate flow matching consistency error. The shape reconstruction error and flow matching consistency error are combined to generate a skip flow distillation loss. The initial 3D shape generation network is updated based on the skip flow distillation loss to generate a 3D shape generation network.

10. The point cloud to parameterized model generation method based on joint optimization of skip flow distillation and confidence level according to claim 9, characterized in that: The initial 3D shape generation network includes a shared shape decoding backbone and a component type-specific determination head corresponding to each component type. The 3D shape generation network generated after updating the initial 3D shape generation network according to the jump flow distillation loss retains the shared shape decoding backbone and the component type-specific determination head. The process of determining generative 3D shape parameters by the 3D shape generation network includes: using a shared shape decoding backbone to perform shape decoding based on point cloud conditional features to generate a shared 3D shape representation, and determining a target component type-specific determination head based on the component type corresponding to the generated inferred point cloud unit. The shared 3D shape representation is input into a target component type-specific determination head to generate generative 3D shape parameters corresponding to the generated inferred point cloud units.

11. The point cloud to parameterized model generation method based on joint optimization of skip flow distillation and confidence level according to claim 1, characterized in that, The method of jointly verifying and fitting three-dimensional shape parameters and generative three-dimensional shape parameters based on the spatial adjacency relationship of components to generate shape continuity verification results includes: Based on the spatial adjacency relationship of components, the component pairs to be verified are determined from the components corresponding to the fitted three-dimensional shape parameters and the generative three-dimensional shape parameters, and then the component connection type corresponding to the component pairs to be verified is determined. Extract shape continuity features from the fitted three-dimensional shape parameters or generative three-dimensional shape parameters of the component to be verified on both sides respectively; Based on the component connection type, determine the shape connection comparison rules and shape connection tolerance, and compare the shape connection characteristics of the component to be verified on both sides according to the shape connection comparison rules to generate shape connection residuals; The shape continuity residual is compared with the shape continuity tolerance to generate the shape continuity status. Then, the shape continuity status of each component to be verified is summarized to generate the shape continuity verification result.

12. The point cloud to parameterized model generation method based on joint optimization of skip flow distillation and confidence level according to claim 11, characterized in that, The process of adjusting the processing path of the point cloud unit of the failed component based on the shape continuity verification result includes: Based on the shape continuity verification results, identify the fitted 3D shape parameters that failed the shape continuity verification, and then trace the component point cloud units corresponding to the fitted 3D shape parameters that failed the shape continuity verification to generate component point cloud units that failed the fitting verification. The processing path for the point cloud units of the component that failed the fitting verification was changed from the geometric fitting path to the inference generation path, so as to generate the fitting branch adjustment results. Based on the shape continuity verification results, identify the generative 3D shape parameters that fail the shape continuity verification, and then trace the component point cloud units corresponding to the generative 3D shape parameters that fail the shape continuity verification to generate the component point cloud units that fail the verification. The generation inference path corresponding to the point cloud unit of the failed component is maintained to generate the generation branch hold result. The fitting branch adjustment result and the generation branch hold result are combined to generate the processing path adjustment result, so that the processing path adjustment result participates in the generation of updating three-dimensional shape parameters.

13. The point cloud to parameterized model generation method based on joint optimization of skip flow distillation and confidence level according to claim 2, characterized in that, Associating the candidate component type set with the corresponding component point cloud unit to determine the component type corresponding to the component point cloud unit includes: The corresponding component point cloud units are fitted according to each candidate component type in the candidate component type set to generate candidate fitted 3D shape parameters. The 3D shape generation network is then used to process the corresponding component point cloud units according to each candidate component type to generate candidate generative 3D shape parameters. Combine candidate fitted 3D shape parameters and candidate generative 3D shape parameters to generate a set of candidate 3D shape parameters; Based on each candidate 3D shape parameter in the candidate 3D shape parameter set, the candidate 3D surface point cloud is reconstructed. The candidate 3D surface point cloud is compared with the corresponding component point cloud unit to generate the consistency of candidate point cloud reconstruction. The candidate 3D shape parameters are verified according to the spatial adjacency relationship of the component to generate the consistency of candidate shape continuity. Combining the consistency of candidate point cloud reconstruction and the consistency of candidate shape continuity, each candidate 3D shape parameter in the candidate 3D shape parameter set is filtered, and the filtered candidate 3D shape parameters are determined as category-confirmed 3D shape parameters. The category-confirmed component type is determined according to the candidate component type corresponding to the category-confirmed 3D shape parameter, and then the component type corresponding to the component point cloud unit is updated using the category-confirmed component type.

14. The point cloud to parameterized model generation method based on joint optimization of skip flow distillation and confidence level according to claim 10, characterized in that, The step of generating component type failure states according to component type and updating the 3D shape generation network based on component type failure states includes: Based on the shape continuity verification results, identify the generative 3D shape parameters that fail the shape continuity verification, and then trace the generative inference point cloud units corresponding to the generative 3D shape parameters that fail the shape continuity verification to generate point cloud units of the generative verification failed components. The frequency of verification failures of point cloud units of failed components is generated according to component type, so as to generate component type failure status, and the active distillation component type is determined based on the component type failure status. Extract historical point cloud units of the same type as the active distillation component from a pre-built historical point cloud sample library, and combine the point cloud units of the active distillation component type that failed to generate verification with the historical point cloud units of the same type to generate active distillation samples. The teacher multi-step flow model is used to perform state transformations based on active distillation samples to generate supplementary 3D shape state transformation trajectories. The component type-specific determination head corresponding to the active distillation component type is updated based on the supplementary 3D shape state transformation trajectory to generate an updated 3D shape generation network.

15. The point cloud to parameterized model generation method based on joint optimization of skip flow distillation and confidence level according to claim 12, characterized in that, The process involves generating updated 3D shape parameters for failed component point cloud units using an updated 3D shape generation network, iteratively verifying and updating these parameters to generate a set of verified 3D shape parameters, and then sequentially verifying and confirming these parameters according to the spatial adjacency relationships of the components to generate a parameterized 3D model of the point cloud. This includes: Based on the processing path adjustment results, the fitting verification failed component point cloud units and the generation verification failed component point cloud units are determined. The fitting verification failed component point cloud units and the generation verification failed component point cloud units are input into the updated 3D shape generation network to generate updated 3D shape parameters. The fitted 3D shape parameters corresponding to the point cloud units of the failed fitting verification component and the generative 3D shape parameters corresponding to the point cloud units of the failed generation verification component are replaced with updated 3D shape parameters to generate an updated set of 3D shape parameters to be verified. Repeat the shape continuity check on the updated set of three-dimensional shape parameters to be checked until the updated set of three-dimensional shape parameters to be checked meets the check termination condition, so as to generate a set of three-dimensional shape parameters to be checked and confirmed. Based on the 3D point cloud, the point cloud coordinate domain is determined, and the verified 3D shape parameter set is transformed into the point cloud coordinate domain to generate a coordinate-unified 3D shape parameter set. The coordinate-unified 3D shape parameter set is then continued according to the spatial adjacency relationship of the components to generate a point cloud parameterized 3D model.

Images