Implicit parameterized CAD modeling method based on autoregression point cloud deconstruction
By generating modeling latent variables through an autoregressive point cloud deconstruction method and introducing a discriminant module, the problems of low efficiency and semantic inconsistency in point cloud data model generation in existing technologies are solved, realizing efficient and interpretable parametric modeling that is suitable for CAD systems.
Patent Information
- Application Number
- CN202511218574.0
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-08-28
- Publication Date
- 2026-01-20
AI Technical Summary
Existing technologies suffer from low efficiency, strong reliance on specialized knowledge, difficulty in generalization, semantic inconsistency, and structural disorder when generating structured, editable parametric models from point cloud data, especially when dealing with complex structures or composite models.
An autoregressive point cloud deconstruction method is adopted, which generates latent variables for modeling structure through autoregression and introduces a discriminant module to filter effective modeling instructions. The Transformer decoder is used to generate sketch parameters and stretching parameters, and a unified discrete modeling instruction vocabulary is constructed to realize the automated process from point cloud to parametric modeling.
It achieves efficient alignment from point cloud to parametric modeling, generates a clear and semantically reasonable sequence of modeling instructions, is applicable to mainstream CAD systems, has good interpretability and editability, and is suitable for fields such as industrial product modeling and structural design automation.
Smart Images
Figure CN121365578A_ABST
Abstract
Description
TECHNICAL FIELD
[0001] The present application relates to an implicit parameterization modeling method based on autoregressive point cloud decomposition, belonging to the technical field of computer vision, computer graphics and computer-aided design. BACKGROUND
[0002] In recent years, with the continuous progress of three-dimensional perception and generation technology, three-dimensional modeling based on point cloud has become a key technical path to promote the development of intelligent design, digital manufacturing and virtual reality applications. As an important data form for obtaining three-dimensional geometric information, point cloud has been widely used in reverse modeling, intelligent manufacturing, building information modeling (BIM), automatic driving perception system and other fields. Especially in the computer-aided design (CAD) scene, how to convert unstructured three-dimensional point cloud data into structured and editable parameterized models is one of the core challenges to realize the automation from perception to design.
[0003] Traditional three-dimensional modeling methods usually rely on manual modeling tools or rule-based modeling systems, which have problems such as low efficiency, strong dependence on professional knowledge, and difficulty in generalization. With the development of deep learning, more and more research attempts to use neural networks to directly learn three-dimensional structures from point cloud data and generate parameterized modeling instructions. However, due to the irregularity of point cloud itself and the lack of topological information, it is still difficult to extract high-level design semantics and modeling logic from it.
[0004] Current mainstream point cloud-based parameter modeling methods mainly include the following categories:
[0005] 1. Modeling method based on overall encoding
[0006] This kind of method usually uses an encoding-decoding structure to encode the entire point cloud into global hidden variables, and decodes it into a sequence of modeling instructions through a neural network model. These methods can better capture the overall geometric shape, but often ignore the composition relationship in the structure of the object, resulting in a lack of semantic interpretability of the generated results, and the problem of insufficient generalization when facing complex component combinations or modeling logic.
[0007] 2. Modeling method based on local block decomposition
[0008] To improve the structural expression ability of modeling, some methods attempt to spatially block or semantically segment the point cloud, and respectively model each sub-block. However, the division of local blocks often relies on artificial rules or external supervision, lacks adaptability, and at the same time, when modeling across blocks, it is easy to have inconsistent connections or broken modeling logic.
[0009] 3. Modeling method based on sketch condition generation
[0010] Such methods take sketch images or two-dimensional contours as input, combine point cloud information to perform three-dimensional modeling prediction, which can effectively improve the consistency between the generated shape and the user's intention. However, the modeling process usually relies on additional sketch input, lacks the ability to understand the modeling semantics from the point cloud itself, and is not suitable for modeling scenarios without prior constraints.
[0011] 4. Modeling method based on autoregressive framework
[0012] To better simulate the artificial modeling process, some studies attempt to represent CAD modeling as a sequence of step-by-step instructions, using an autoregressive framework to generate modeling commands step by step. Although these methods improve modeling interpretability and operation reproducibility, they often lack explicit decomposition mechanisms at the structural level, making it difficult to effectively modularize modeling and control semantics for complex geometries.
[0013] In summary, existing methods still have obvious shortcomings in structural understanding, semantic decomposition, and modeling instruction generation. Especially when faced with complex structures or combined models, decoding distortion, semantic inconsistency, or structural disorder may occur. Therefore, there is an urgent need for a modeling framework with structural perception ability that can adaptively extract part-level semantics from point clouds and efficiently generate structured parameter modeling instructions, thereby realizing an interpretable, editable, and highly restored three-dimensional CAD modeling process. SUMMARY
[0014] Therefore, the present application provides an implicit parameterized CAD modeling method based on autoregressive point cloud decomposition, which takes three-dimensional point cloud as input, generates modeling structure latent variables through autoregression, and decodes into a structured modeling instruction sequence, realizing an automatic modeling process from perceiving geometry to parameter modeling. This method overcomes the problem of traditional point cloud reconstruction that can only generate geometric surfaces without structured design logic, and can support modeling process reasoning and editing for complex structures.
[0015] An implicit parameterized modeling method based on autoregressive point cloud decomposition, comprising the following steps:
[0016] Step 1: Obtain three-dimensional point cloud data of the target object
[0017] Step 2: Process the input point cloud using a point cloud encoder to extract global geometric features
[0018] Step 3: Construct an autoregressive part decoder with initial features z s as the starting representation, combining global context to generate a sequence of implicit modeling part latent variables step by step where N represents the maximum number of modeling operations; zn represents the nth latent variable;
[0019] Step 4: For the latent variable sequence corresponding sketch parameters and stretch parameters
[0020] Step 5: Quantize and encode each set of modeling parameter pairs , map to a predefined discrete modeling instruction vocabulary, and construct a unified symbolic modeling space;
[0021] Step 6: Introduce a discriminative module to determine whether each latent variable z n outputs a validity probability score and combines the validity label to determine whether the nth set of modeling parameter pairs participates in the final modeling; concatenate all valid modeling parameter pairs to generate the final modeling program sequence input the sequence into the parameterized modeling environment to generate the target CAD structure.
[0022] Further, it also includes optimizing the cloud encoder, latent variable generation module, parameter decoder, and validity discriminative module; wherein the loss function for optimization is:
[0023]
[0024] wherein, and represent the multi-class cross-entropy loss of sketch parameters and stretch parameters respectively, is the binary cross-entropy loss of validity discrimination, λ skt and λ ext are weight adjustment parameters used to balance the loss weights of different supervision signals.
[0025] Preferably, in step 1, the three-dimensional point cloud data of the target object contains the geometric contour information, three-dimensional coordinates, normal vectors, and color information of the target object.
[0026] Preferably, in step 2, the point cloud encoder uses a three-layer edge convolutional graph neural network structure to aggregate local neighborhood features and extract global features from the input point cloud , and finally obtains global geometric context features
[0027] Preferably, in step 3, the autoregressive decoder is composed of four standard Transformer layers, and the global context is input as a condition to initialize a special starting vector z n , and gradually generate a sequence of latent variables where each latent variable z n characterizing one semantic component in the target structure.
[0028] Preferably, in step 4, the two independent Transformer decoders are respectively a sketch parameter decoder composed of two standard Transformer layers and a stretch parameter decoder composed of two standard Transformer layers; the two-dimensional sketch parameter for representing a closed contour composed of straight lines, circular arcs and circular basic primitives; the three-dimensional modeling parameter including the position, normal direction, stretch depth, scaling factor and Boolean operation type of the sketch plane.
[0029] Preferably, in step 6, the discrimination module is composed of a linear layer.
[0030] The present application has the following beneficial effects:
[0031] (1) The present application introduces an autoregressive structure to perform semantic deconstruction on the point cloud, which can gradually generate modeling latent variable sequences with structural significance, significantly improving the expression ability of the point cloud to the modeling program under complex structures, and realizing efficient alignment from original geometry to modeling instructions.
[0032] (2) The present application adopts parallel sketch parameter and stretch parameter decoding branches, and generates modeling instruction pairs on each latent variable through a two-way Transformer decoder, fully modeling the two-dimensional profile and three-dimensional generation attributes of the construction operation, and realizing more accurate and structurally reasonable parameterized modeling output.
[0033] (3) The present application constructs a unified discrete modeling instruction vocabulary and quantizes continuous parameters into symbolic parameters, so that the modeling process has good executability and editability, and can be directly used in mainstream CAD systems to complete programmed modeling tasks, and has high universality and industrial adaptability.
[0034] (4) The present application introduces a modeling effectiveness discrimination module to automatically filter effective construction steps and predict the modeling termination point, further improving the expression simplicity and semantic rationality of the modeling program, and reducing the generation of redundant modeling steps.
[0035] (5) The joint optimization mechanism proposed in the present application performs end-to-end training on the point cloud encoder, latent variable decoder and modeling instruction decoder, and uses a multi-objective loss function to effectively constrain the sketch, stretch parameters and discrimination effectiveness, so that stable and high-quality CAD instruction sequence generation can be realized under different complexity scenarios.
[0036] (6) This invention is applicable to multiple fields such as industrial product modeling, structural design automation, graphic content generation, and reverse modeling. It is especially suitable for three-dimensional reconstruction tasks with discrete modeling semantics and has good scalability and practical application prospects. Attached Figure Description
[0037] Figure 1 A flowchart illustrating the architecture provided by this invention;
[0038] Figure 2 This is a schematic diagram of the autoregressive deconstruction and modeling parameter generation process of the present invention;
[0039] Figures 3(a) and 3(b) are example diagrams of the output point cloud and modeling result of a part in the implementation, respectively;
[0040] Figures 4(a) and 4(b) are example diagrams of the output point cloud and modeling results of another part in the implementation. Detailed Implementation
[0041] The present invention will now be described in detail with reference to the accompanying drawings.
[0042] like Figure 1 As shown, this invention provides an implicit parametric CAD modeling method based on autoregressive point cloud deconstruction, specifically including the following steps:
[0043] An implicit parametric modeling method based on autoregressive point cloud deconstruction includes the following steps:
[0044] Step 1: Acquire the 3D point cloud data of the target object Each point x i ∈R 3 It includes its three-dimensional coordinate information, with optional additional normal vectors or color attributes;
[0045] Step 2: Using the point cloud data obtained in Step 1 Extracting global geometric features using a point cloud encoder As a contextual semantic representation in modeling tasks;
[0046] Step 3: Based on the global features extracted in Step 2, construct an autoregressive part representation decoder and initialize a special initial representation z. s and with As a conditional input, an implicit sequence of latent variables for modeling components is generated step by step. Each z n It represents the semantic expression of a potential modeling operation, implying its corresponding local structure and modeling intent.
[0047] Step 4: For all latent variables generated in Step 3 adopt two independent Transformer decoders, and generate its corresponding sketch parameters n on each z with the stretching parameters
[0048] Step 5: Map each generated modeling parameter pair to a predefined discrete instruction vocabulary (256 dimensions) by quantization encoding of continuous modeling parameters, to construct a unified symbolic modeling instruction space;
[0049] Step 6: Introduce a discriminant module to each latent variable z n output its validity probability score and based on the discriminant label determine whether the instruction pair participates in the final configuration generation. All modeling instruction pairs that are judged to be valid will be spliced into the final modeling program sequence which can be directly input into the parametric modeling environment to reconstruct the target CAD structure;
[0050] Step 7: Use a joint training mechanism to end-to-end optimize the point cloud encoder, latent variable generation module, parameter decoder, and validity discriminant module in steps 2 to 6. The overall loss function is defined as follows:
[0051]
[0052] Step 8: Terminate the architecture optimization training process in step 7, obtain the converged model parameters, and then realize the automatic generation of high-quality parametric modeling instruction sequences from input point clouds. This sequence can be directly input into the CAD modeling environment (such as Fusion360, FreeCAD, etc.) to complete the explainable modeling and editable reconstruction of the target structure.
[0053] In step 1, the three-dimensional point cloud data of the target object is obtained The point cloud can be obtained by a depth sensor, laser scanner, etc., containing the geometric contour information of the target object; wherein each point x i includes its three-dimensional coordinates, and optionally includes normal vector or color information to enhance the representation ability.
[0054] In step 2, the point cloud encoder uses edge convolution and other graph neural network structures to aggregate local neighborhood features and extract global features from the input point cloud Finally, the global geometric context features are obtained to provide modeling semantic conditions for subsequent latent variable generation and guide the subsequent latent variable decoding process.
[0055] In step 3, the autoregressive decoder consists of four standard Transformer layers, with a global context. Initialize a special initial vector z as the conditional input. n Gradually generate a sequence of latent variables Each latent variable z n This represents a semantic component within the target structure. The autoregressive process can be implemented using a four-layer Transformer decoding network, where the generation of latent variables at each step is jointly constrained by previously generated latent variables and the global context. Here, N is the maximum number of modeling operations (representing a sketch + stretch operation).
[0056] In step 4, two independent Transformer decoders are guided to process the entire latent variable sequence. Decoding operations are performed; the sketch parameter decoder consists of two standard Transformer layers, based on each z... n Parameters for generating 2D sketches Used to represent closed contours composed of basic primitives such as lines, arcs, and circles (represented in x and y coordinates); the stretch parameter decoder consists of two standard Transformer layers, based on the same latent variable z. n Generate 3D modeling parameters This includes the position of the sketch plane, normal direction, extrusion depth, scaling factor, and Boolean operation type (new, cut, merge, intersection, etc.).
[0057] In step 5, a parameter discretization module is used to discretize the continuous modeling parameters. This is mapped to a unified discrete instruction vocabulary. Each parameter dimension is segmented into a fixed number (256 classes) of discrete intervals, thereby enabling the modeling of instruction sequences. Symbolic representation.
[0058] In step 6, the discrimination module consists of a linear layer, which performs discrimination on each latent variable z. n Performs validity assessment. This module outputs a validity probability score. Used to estimate whether the latent variable corresponds to a real modeling operation; supervision label Automatically generated from the length of the actual modeling sequence, indicating whether the modeling step is a valid part. Finally, all are filtered out. Modeling instructions corresponding to latent variables The final modeling program is generated by piecing together the data. Used for structural reconstruction.
[0059] In step 7, the point cloud encoder, latent variable generation module, parameter decoder, and validity discrimination module are jointly trained end-to-end. The following loss function is used during the training process:
[0060]
[0061] wherein, and denote the multi-class cross-entropy loss of sketch parameters and stretch parameters respectively, is the binary cross-entropy loss of effectiveness discrimination, and λ skt and λ ext are weight adjustment parameters for balancing the loss weights of different supervision signals.
[0062] In step 8, according to the preset training round number and verification index, the optimization process is terminated when the model training converges, and a set of network parameters that can be used for modeling inference is obtained. After the training is completed, the proposed method can directly input any target point cloud data and output a modeling instruction sequence with clear structure and reasonable semantics for modeling program execution and target structure reconstruction in mainstream CAD systems.
[0063] Embodiment:
[0064] The model used in this embodiment is based on an implicit parameterized modeling method based on autoregressive point cloud deconstruction proposed by the present application, which is trained end-to-end on the DeepCAD dataset. The total number of training rounds is 200 rounds, and the training set, validation set and test set are divided according to the ratio of 9:0.5:0.5. During the training process, the AdamW optimizer is used, the initial learning rate is set to 1e-4, and periodic evaluation is performed on the validation set to ensure convergence and generalization ability. All experiments are completed in parallel on 8 NVIDIA A100-40GB GPUs.
[0065] To verify the modeling ability of the model in different structure scenarios, as shown in FIG. 3, this embodiment selects four groups of three-dimensional point clouds with different structures in the test set as input, covering different types of geometric features, including multi-outline sketches, circular arc mixed structures, geometric objects with holes and Boolean operations, etc. All point clouds are preprocessed to ensure point cloud integrity and uniformity, with a point number of 2048 and normal attributes.
[0066] According to steps 1 to 3 of the present application, the input point cloud is first input to the point cloud encoding module to extract global geometric features and randomly initialize the initial vector z s . Then an autoregressive decoder is used to gradually generate a set of modeling latent variables Each latent variable represents a potential modeling operation component and has specific local semantics and geometric partition directionality.
[0067] As shown in FIG. 3(a), FIG. 3(b), FIG. 4(a), FIG. 4(b), according to steps 4 to 6, all latent variables are guided in parallel to sketch and stretch two independent Transformer decoders, respectively outputting sketch parameters corresponding to each latent variable and stretch parameters After the parameter output is mapped to the discrete modeling vocabulary, it constitutes a token sequence of symbolic instructions, and is screened by the discriminant module to form the final modeling program
[0068] FIG. 3(a), FIG. 3(b), FIG. 4(a), FIG. 4(b) show the CAD model modeling results of the final prediction parameters. The modeling method proposed by the present application has the following advantages:
[0069] Strong structure recovery capability: For complex sketch geometry (such as the multi-outline structure in FIG. 3(a)), the internal outline, nested structure and their topological relationship can be accurately identified;
[0070] Reasonable coupling of sketch and stretch: As shown in FIG. 3(b), for a model with multiple circular arcs and long stretch paths, the generated sketch is highly matched with the corresponding stretch direction, maintaining consistent configuration;
[0071] Supporting Boolean modeling operations: FIG. 4(a) shows the results of hole and cutting operations. The generated Boolean operation can accurately reconstruct the required geometric relationship;
[0072] Clear and editable instruction semantics: All generated instructions are based on discrete modeling space representation and can be directly loaded and parsed by standard CAD environments (such as Fusion360, FreeCAD), with good interpretability and scalability.
[0073] In summary, the embodiment verifies the modeling stability and universality of the implicit parameterization modeling method based on autoregressive point cloud reconstruction under different types of point cloud input. Compared with traditional point cloud reconstruction methods, this method not only generates CAD models with clear structure and reasonable semantics, but also has good engineering usability and subsequent editing capability, suitable for various application scenarios such as reverse engineering, structure reconstruction and industrial modeling.
[0074] It should be understood that the above is only a preferred embodiment of the present application, and is not intended to limit the protection scope of the present application. Any modification, equivalent replacement, improvement, etc. within the spirit and principles of the present application shall be included in the protection scope of the present application.
Claims
1. An implicit parametric modeling method based on autoregressive point cloud deconstruction, characterized in that, Comprising the following steps: Step 1: Obtain three-dimensional point cloud data of the target object Step 2: processing the input point cloud with a point cloud encoder to extract global geometric features Step 3: Constructing an autoregressive component decoder to initial features z s For the initial representation, incorporate global context Step 2: Generate the sequence of implicit modeling component latent variables where N denotes the maximum number of modeling operations; z n denotes the n-th latent variable; Step 4: Conditioning on the latent variable sequence Corresponding generation of sketch parameters and stretch parameters Step 5: Quantization encoding each set of modeling parameters pair to a predefined discrete modeling instruction vocabulary, constructing a unified symbolic modeling space; Step 6: Introduce discriminative module to each latent variable z n Output the validity probability score And combine with the validity label Determine whether the nth modeling parameter pair participates in the final modeling; splice all valid modeling parameter pairs to generate the final modeling program sequence Input the sequence into the parameterized modeling environment to generate the target CAD structure.
2. The implicit parameterization modeling method based on autoregressive point cloud decomposition according to claim 1, characterized in that, Further comprising optimizing the cloud encoder, the latent variable generation module, the parameter decoder and the validity discrimination module; wherein the loss function for optimization is: In the formula, and respectively represent the multi-class cross-entropy loss of sketch parameters and stretching parameters, is the binary cross-entropy loss of effectiveness discrimination, λ skt and λ ext is a weight adjustment parameter used to balance the loss weights of different supervision signals.
3. The implicit parameterization modeling method based on autoregressive point cloud deconstruction of claim 2, wherein, In the step 1, the three-dimensional point cloud data of the target object contains geometric contour information, three-dimensional coordinates, normal vectors and color information of the target object.
4. The implicit parameterization modeling method based on autoregressive point cloud deconstruction of claim 2, wherein, In the step 2, the point cloud encoder adopts a three-layer edge convolutional graph neural network structure to encode the input point cloud carries out local neighborhood feature aggregation and global feature extraction, and finally obtains global geometric context features 5. The implicit parameterization modeling method based on autoregressive point cloud deconstruction of claim 2, wherein, In step 3, the autoregressive decoder consists of four standard Transformer layers with global context as input, initializing a special starting vector z n , generating the latent variable sequence step by step where each latent variable z n represents one semantic component in the target structure.
6. The implicit parameterization modeling method based on autoregressive point cloud deconstruction of claim 2, wherein, In the step 4, two independent Transformer decoders are sketch parameter decoder composed of two standard Transformer layers, and stretch parameter decoder composed of two standard Transformer layers; two-dimensional sketch parameters For representing a closed contour composed of straight lines, circular arcs, and circular basic primitives; three-dimensional modeling parameters Including the position, normal direction, stretch depth, scaling factor, and Boolean operation type of the sketch plane.
7. The implicit parameterization modeling method based on autoregressive point cloud deconstruction of claim 2, wherein, In the step 6, the discrimination module is composed of a linear layer.