Vehicle three-dimensional model generation method and device, equipment and medium

Abstract

The invention provides a vehicle three-dimensional model generation method and device, equipment and a medium. The method comprises the following steps: acquiring at least one input image of a target vehicle; inputting the input image into the cross-domain attitude estimation network, and outputting a corresponding camera attitude parameter; based on the camera attitude parameters, performing three-dimensional reconstruction on the input image to obtain an initial three-dimensional model; semantic post-processing is conducted on the initial three-dimensional model, a vehicle three-dimensional model supporting editing is obtained, and the vehicle three-dimensional model has component-level semantic segmentation information. The scheme is used for solving the problems that the editability of a vehicle three-dimensional model generated from an image is insufficient, and engineering simulation and design requirements cannot be met in the related technology.

Description

Technical Field

[0001] This application relates to the field of vehicle simulation technology, and in particular to a method, apparatus, device and medium for generating three-dimensional vehicle models. Background Technology

[0002] In fields such as automotive industrial design, virtual reality demonstrations, and autonomous driving simulation testing, high-precision, editable 3D vehicle models are fundamental for product design iteration, enriching simulation scene libraries, and enhancing user experience. Traditional 3D model generation primarily relies on manual modeling by professional designers, but this method is inefficient, costly, and struggles to quickly generate massive numbers of models or respond to personalized design needs.

[0003] With the development of artificial intelligence technology, automated vehicle 3D model generation technology has become the mainstream of research. However, automated model generation methods rely heavily on a large amount of pre-annotated image information as input, and the output model lacks the inherent structural constraints of the vehicle, resulting in poor model editability. Summary of the Invention

[0004] The vehicle 3D model generation method, apparatus, device, and medium provided in this application are intended to solve the problems in related technologies where the editability of vehicle 3D models generated from images is insufficient and cannot meet the needs of engineering simulation and design.

[0005] In a first aspect, embodiments of this application provide a method for generating a three-dimensional vehicle model, including:

[0006] Obtain at least one input image of the target vehicle;

[0007] The input image is fed into a cross-domain pose estimation network, which outputs the corresponding camera pose parameters.

[0008] Based on the camera pose parameters, the input image is reconstructed in three dimensions to obtain an initial three-dimensional model;

[0009] The initial 3D model is semantically post-processed to obtain an editable 3D vehicle model, which contains component-level semantic segmentation information.

[0010] In one embodiment of this disclosure, an input image is input into a cross-domain pose estimation network, and corresponding camera pose parameters are output. This includes: extracting geometric features of the input image based on a pre-trained cross-domain pose estimation network, wherein the cross-domain pose estimation network is trained on a simulation image dataset containing a vehicle 3D model and ground truth camera parameters, and includes a cross-domain adaptive module; performing domain difference compensation on the geometric features based on the cross-domain adaptive module to obtain supplemented geometric features, wherein the compensation processing is used to adapt to the actual image domain of the input image; and performing regression processing on the compensated geometric features to obtain the camera pose parameters.

[0011] In one embodiment of this disclosure, a semantic post-processing is performed on an initial 3D model to obtain an editable 3D vehicle model, including: segmenting the surface of the initial 3D model based on a component semantic segmentation model to obtain at least two semantic regions corresponding to different components in the vehicle; optimizing the geometry of the semantic regions based on the component type corresponding to each semantic region and predefined engineering geometric constraints; and generating a 3D vehicle model based on the optimized geometry, wherein the engineering geometric constraints include at least one of the following: symmetry constraints applied to symmetrical components belonging to the left and right sides of the vehicle; surface smoothing processing performed on semantic regions belonging to body panels; and parametric dimensional adjustment constraints performed on semantic regions belonging to wheels.

[0012] In one embodiment of this disclosure, the geometry of the semantic region is optimized based on the component type corresponding to each semantic region and predefined engineering geometric constraints, including: if the component type of the semantic region is a car window, the semantic region is optimized by plane fitting; if the component type of the semantic region is a wheel hub, the semantic region is optimized by cylindrical surface fitting; if the component type of the semantic region is a car body panel, the semantic region is optimized by non-uniform rational B-spline surface reconstruction.

[0013] In one embodiment of this disclosure, the geometry of a semantic region is optimized based on the component type corresponding to each semantic region and predefined engineering geometric constraints, including: establishing an editable parameter set for a three-dimensional vehicle model, the parameter set being associated with the geometric dimensions of at least one semantic region; and automatically updating the geometry of the corresponding semantic region in response to modifications to the parameters in the parameter set.

[0014] In one embodiment of this disclosure, after semantic post-processing of the initial 3D model to obtain an editable vehicle 3D model, the method further includes: rendering at least one feedback image based on the vehicle 3D model; determining the difference between the feedback image and the original design intent; and adjusting the parameters of the input image generation process or semantic post-processing based on the difference to iteratively optimize the vehicle 3D model.

[0015] In one embodiment of this disclosure, acquiring at least one input image of a target vehicle includes: receiving a text description or design sketch corresponding to the target vehicle; and generating at least one input image based on a generative artificial intelligence model according to the text description or design sketch.

[0016] Secondly, embodiments of this application provide a vehicle three-dimensional model generation apparatus, comprising:

[0017] The acquisition module is used to acquire at least one input image of the target vehicle;

[0018] The determination module is used to input the input image into the cross-domain pose estimation network and output the corresponding camera pose parameters;

[0019] The generation module is used to perform 3D reconstruction on the input image based on the camera pose parameters to obtain an initial 3D model;

[0020] The optimization module is used to perform semantic post-processing on the initial 3D model to obtain an editable 3D vehicle model, which has component-level semantic segmentation information.

[0021] Optionally, the determination module is specifically used to: extract geometric features of the input image based on a pre-trained cross-domain pose estimation network, wherein the cross-domain pose estimation network is trained on a simulation image dataset containing a vehicle 3D model and the ground truth values ​​of camera parameters, and the cross-domain pose estimation network includes a cross-domain adaptive module; based on the cross-domain adaptive module, perform domain difference compensation on the geometric features to obtain supplemented geometric features, wherein the compensation processing is used to adapt to the actual image domain of the input image; and perform regression processing on the compensated geometric features to obtain camera pose parameters.

[0022] Optionally, the optimization module is specifically used to: segment the surface of the initial 3D model based on the component semantic segmentation model to obtain at least two semantic regions corresponding to different components in the vehicle; optimize the geometry of the semantic regions based on the component type and predefined engineering geometric constraints corresponding to each semantic region; and generate a 3D model of the vehicle based on the optimized geometry. The engineering geometric constraints include at least one of the following: symmetry constraints applied to symmetrical components belonging to the left and right sides of the vehicle; surface smoothing processing performed on semantic regions belonging to body panels; and parametric dimensional adjustment constraints performed on semantic regions belonging to wheels.

[0023] Optionally, the optimization module is specifically used to: optimize the semantic region by plane fitting if the component type of the semantic region is a car window; optimize the semantic region by cylindrical surface fitting if the component type of the semantic region is a wheel hub; and optimize the semantic region by non-uniform rational B-spline surface reconstruction if the component type of the semantic region is a car body panel.

[0024] Optionally, the optimization module is specifically used to establish an editable parameter set for the vehicle's 3D model, the parameter set being associated with the geometric dimensions of at least one semantic region; and to automatically update the geometry of the corresponding semantic region in response to modifications to the parameters in the parameter set.

[0025] Optionally, the optimization module is also used to: render and generate at least one feedback image based on the vehicle 3D model; determine the difference between the feedback image and the original design intent; and adjust the parameters of the input image generation process or semantic post-processing based on the difference to iteratively optimize the vehicle 3D model.

[0026] Optionally, the acquisition module is specifically used to receive a text description or design sketch corresponding to the target vehicle; and to generate at least one input image based on the text description or design sketch using a generative artificial intelligence model.

[0027] Thirdly, embodiments of this application provide an electronic device, including:

[0028] Processor, memory, communication interface;

[0029] Memory is used to store the processor's executable instructions;

[0030] The processor is configured to execute the vehicle 3D model generation method of any of the first aspects by executing executable instructions.

[0031] Fourthly, embodiments of this application provide a readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements the vehicle 3D model generation method of any one of the first aspects.

[0032] Fifthly, embodiments of this application provide a computer program product, including a computer program, which, when executed by a processor, is used to implement the vehicle 3D model generation method according to any one of the first aspects.

[0033] The vehicle 3D model generation method, apparatus, device, and medium provided in this application acquire vehicle images, automatically calculate camera parameters using a cross-domain pose estimation network, and then perform 3D reconstruction and semantic post-processing to ultimately generate an editable 3D model with component-level semantic information. This combines robust cross-domain pose estimation with engineering-oriented structured post-processing, solving the problems of existing technologies that rely on precise image annotation and cannot handle diverse image sources, and the lack of semantic and engineering constraints that prevents direct editing of the output model. It achieves fully automated generation of high-quality, structured vehicle 3D models from single or multiple images, directly usable for subsequent industrial design and simulation, significantly reducing the professional threshold and time cost. Attached Figure Description

[0034] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with this application and, together with the description, serve to explain the principles of this application.

[0035] Figure 1 An application scenario diagram of the vehicle 3D model generation method provided in this application;

[0036] Figure 2 This is a flowchart illustrating an embodiment of a vehicle 3D model generation method provided in this application.

[0037] Figure 3 This is a flowchart illustrating an embodiment of a vehicle 3D model generation method provided in another embodiment of this application.

[0038] Figure 4 A schematic diagram of the structure of an embodiment of a vehicle 3D model generation device provided in another embodiment of this application;

[0039] Figure 5 This is a schematic diagram of the structure of an electronic device provided in this application.

[0040] The accompanying drawings have illustrated specific embodiments of this application, which will be described in more detail below. These drawings and descriptions are not intended to limit the scope of the concept in any way, but rather to illustrate the concept of this application to those skilled in the art through reference to specific embodiments. Detailed Implementation

[0041] Exemplary embodiments will now be described in detail, examples of which are illustrated in the accompanying drawings. When the following description relates to the drawings, unless otherwise indicated, the same numbers in different drawings denote the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with this application. Rather, they are merely examples of apparatuses and methods consistent with some aspects of this application as detailed in the appended claims.

[0042] The terms “first,” “second,” “third,” “fourth,” etc. (if present) in the specification, claims, and accompanying drawings of this application are used to distinguish similar objects and are not necessarily used to describe a particular order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments of this application described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms “comprising” and “having,” and any variations thereof, are intended to cover a non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0043] In fields such as automotive industrial design, virtual reality demonstrations, and autonomous driving simulation, the ability to automatically generate high-precision, engineering-usable 3D vehicle models from images is crucial for accelerating design iterations and building simulation environments. However, achieving this goal faces two fundamental obstacles: first, the model needs to be adaptable to diverse and unannotated internet or generative images, i.e., possess cross-domain robustness; second, the generated model must not be a simple geometric shell, but rather have clear component divisions and conform to basic engineering design specifications to meet subsequent editing, adjustment, and simulation needs.

[0044] In related technologies, multi-view geometry-based methods heavily rely on image texture and precise viewpoints, failing with smooth vehicle surfaces and unlabeled images. While methods based on neural radiation fields can reconstruct appearances, their output is an unstructured implicit representation that cannot be directly converted into a model of separable parts. Traditional parametric models lack flexibility and cannot derive reasonable parameters from arbitrary images. Therefore, existing technologies are either limited by input conditions or fail to meet engineering requirements in terms of output results, resulting in a long-standing technological gap.

[0045] The vehicle 3D model generation method provided in this application breaks through the dependence on pre-labeled data by using a cross-domain pose estimation network to robustly recover the geometric information required for 3D reconstruction from any image. Then, by introducing semantic post-processing, the initially reconstructed geometric model is transformed into a digital representation with component-level structure and incorporating engineering constraints. Thus, by combining cross-domain adaptability with engineering semantic injection, the problem of automatically generating a usable model from any image is systematically solved.

[0046] Figure 1 This is a schematic diagram illustrating an application scenario of the vehicle 3D model generation method provided in this application, such as... Figure 1 As shown, during the process of generating a vehicle 3D model, the server 100 obtains the vehicle image 110 of the model to be generated input by the user, and can generate a vehicle 3D model 120 that supports editing processing to guide the vehicle to travel stably along a preset path.

[0047] It should be noted that, Figure 1 The scenario shown includes a server, vehicle images, and a 3D vehicle model, which are only used as examples of one or a specific number. However, this disclosure is not limited to this. That is to say, the number of servers, vehicle images, and 3D vehicle models can be arbitrary.

[0048] The technical solution of this application and how the technical solution of this application solves the above-mentioned technical problems are described in detail below with specific embodiments. These specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described again in some embodiments. The embodiments of this application will be described below with reference to the accompanying drawings.

[0049] Figure 2 Flowchart of the vehicle 3D model generation method provided in this application Figure 1 ,like Figure 2 As shown, the method includes:

[0050] S201. Obtain at least one input image of the target vehicle.

[0051] Specifically, this embodiment is used to illustrate the main steps of the method for generating a three-dimensional vehicle model.

[0052] In this embodiment of the disclosure, the execution entity is a computer, processor, server, etc., used for multibody dynamics simulation analysis of vehicles. For the convenience of subsequent description, they are collectively referred to as servers.

[0053] In this solution, the server is designed to be compatible with cross-domain inputs of varying quality, thereby covering a wide range of scenarios from conceptual design to real-vehicle replication.

[0054] Therefore, the input image can be a single or multiple real vehicle photos crawled from the Internet, or it can be a synthetic image transformed from user-provided design sketches or text descriptions through a generative model.

[0055] In practical implementation, when the input is a text description, the server can integrate an open-source, controllable image generation model.

[0056] At this point, the model can parse natural language descriptions such as streamlined sports cars and off-road vehicles with sharp lines, understand the style, model and structural implications contained therein, and generate vehicle images that conform to the description and have a preliminary controllable perspective.

[0057] If the input is a sketch, the server can use a deep learning-based sketch parsing module to transform the rough outline drawn by the user into a reference image with clear semantic boundaries and preliminary shading.

[0058] It should be noted that the server does not simply take the generated or acquired image as the endpoint, but rather regards it as the starting point of the 3D reconstruction process, and uses the generated or acquired image and the design intent contained therein as the original data source for model generation.

[0059] In some embodiments, the server typically performs a series of preprocessing operations after acquiring the image. For example, for network images with complex backgrounds, the server may apply algorithms based on saliency detection or instance segmentation to attempt to separate the vehicle foreground from the background. For images with uneven lighting, simple histogram equalization may be performed to enhance feature visibility.

[0060] The purpose of these preprocessing steps is not to obtain a perfect image, but to provide a relatively standardized starting point for subsequent algorithm modules that do not rely on perfect input.

[0061] S202. Input the input image into the cross-domain pose estimation network and output the corresponding camera pose parameters.

[0062] Specifically, existing model generation methods typically rely on rich, matchable corner or texture features in images to infer camera position, but this premise often fails to hold true due to the smooth surfaces of vehicles, repetitive structures, and complex backgrounds of network images.

[0063] The cross-domain pose estimation network in this step enables the server to directly learn the mapping relationship from the appearance of a vehicle image to its 3D spatial viewing angle in a data-driven manner, without relying on hand-designed features.

[0064] Specifically, the implementation principle of cross-domain pose estimation networks is based on large-scale simulation pre-training and adaptive inference.

[0065] The server first uses an open-source graphics engine in a virtual environment to programmatically generate tens of thousands of rendered images of different car models under different lighting, weather, and background conditions, with each image having its camera's extrinsic and intrinsic parameters precisely known.

[0066] The server uses this data to train a deep convolutional neural network whose goal is to learn the relationship between the inherent, texture-independent geometric cues of the vehicle object category and the camera projection.

[0067] After training, the network has the potential to estimate camera pose from images, but the knowledge it learns on virtual data has domain differences compared to real, stylistically diverse network images.

[0068] To address this issue, the server can enable a cross-domain adaptive module in the network during inference. This module can be a built-in feature alignment sub-network that analyzes the difference between the feature distribution of the input image and the feature distribution of the simulated pre-training data online, and dynamically adjusts the feature representation of the intermediate layers of the network. This suppresses interference from domain-specific information such as the difference between rendered texture and real-world photo texture, and between simulated lighting and complex real-world lighting, thereby strengthening the representation of the cross-domain invariant feature of vehicle geometry.

[0069] The cross-domain pose estimation network outputs camera parameters for each input image. For example, for a side view of a vehicle, the server may estimate parameters such as the camera being located about 1 meter to the side of the vehicle, at the same height as the center of the wheel, and with a lens focal length of 35 mm.

[0070] These parameters are used to provide geometric constraints for subsequent 3D model reconstruction.

[0071] S203. Based on the camera pose parameters, perform 3D reconstruction on the input image to obtain an initial 3D model.

[0072] Specifically, after obtaining the camera pose parameters, the server can use the known camera pose parameters to back-project or optimize the pixel information in the input image into a unified three-dimensional space, thereby determining the shape of the object.

[0073] One feasible approach is to employ a deep learning-based 3D reconstruction framework. This framework takes an image and the corresponding camera parameters as input and uses an encoder-decoder network to directly regress a discrete voxel grid or a continuous implicit function to represent the vehicle shape.

[0074] Specifically, the server can simultaneously input multiple input images and their camera pose parameters into a three-dimensional convolutional neural network.

[0075] The network first processes each image separately in the image feature extraction stage, and then uses camera parameters to "lift" these two-dimensional features into the corresponding view frustum of the hypothetical three-dimensional spatial grid. It integrates multi-view information through a feature fusion layer (such as trilinear interpolation followed by 3D convolution) and finally decodes a three-dimensional voxel model, where the value of each voxel represents the probability that an object exists at that location.

[0076] In some embodiments, the server can also train a multilayer perceptron network that maps the coordinates of points in three-dimensional space to a symbolic distance value. By using image color information as a supervision signal through camera pose parameters, the network is optimized so that its zero isosurface is the vehicle surface, thereby representing the vehicle shape through an implicit function.

[0077] Regardless of the specific architecture used, the server needs to utilize the reliable camera pose provided in the previous step to solve the core alignment problem in multi-view geometry, thereby effectively fusing visual information from different perspectives that may be blurry or noisy, and inferring a 3D model that may be rough but whose overall structure is basically correct.

[0078] For example, given multiple images of a car, the server can reconstruct a triangular mesh containing the main body of the car, the general outline of the wheel arches, and the recessed areas of the windows. However, the surface of this mesh may be noisy, the edges may be inaccurate, and the various parts may be fused together without clear boundaries, which is the initial 3D model.

[0079] S204. Perform semantic post-processing on the initial 3D model to obtain an editable 3D vehicle model.

[0080] The vehicle's 3D model contains component-level semantic segmentation information.

[0081] Specifically, as mentioned above, the initial model obtained in the previous step is a pure, undifferentiated geometric shell. Based on this, the server needs to inject engineering semantics and design constraints into it to make it "understandable" and "operable".

[0082] The core principle of semantic post-processing is to combine data-driven semantic recognition with knowledge-driven geometric optimization to decompose and optimize a unified mesh into a set of components that conform to engineering specifications.

[0083] The server first performs semantic segmentation.

[0084] Specifically, the server invokes a pre-trained neural network for segmenting parts within the vehicle's 3D mesh.

[0085] This network takes the vertex coordinates and normal information of the initial 3D model as input, propagates the information on the mesh surface through graph convolutional layers, and finally predicts a semantic label for each triangle or vertex, such as "left front door", "hood", "right front wheel hub", "windshield", etc. This is equivalent to giving the model a structured digital "anatomical diagram".

[0086] This process is carried out entirely in three-dimensional space, avoiding the complexity and inconsistencies of segmenting a two-dimensional image and then mapping it to three dimensions.

[0087] After obtaining component-level semantic segmentation, the server immediately initiates rule-based geometric optimization.

[0088] This process applies different engineering priors depending on the component type. For example, for all areas labeled "windows," the server fits one or more planes and adjusts the mesh vertices to lie precisely on the fitted plane, ensuring the flatness of the glass. For major body panels, such as doors and hoods, the server checks whether they should be symmetrical and uses non-rigid registration techniques such as iterative nearest-point algorithms to force symmetrical components to maintain shape consistency, eliminating random asymmetric noise introduced by reconstruction.

[0089] At the same time, the server will perform global and local smoothing processes, using methods such as Laplacian smoothing or its variants to eliminate unreasonable bumps on the surface while preserving feature edges, making the surface transition natural and conforming to the intuitive feeling of aerodynamics or manufacturing process.

[0090] After this series of processes, the initial noisy mesh is transformed into a three-dimensional model with clear component boundaries, smooth surfaces, and conformity to basic design rules. At this point, the model has the foundation to be imported into professional simulation software and further refined, thus facilitating subsequent simulation design.

[0091] The vehicle 3D model generation method provided in this application acquires vehicle images, automatically calculates camera parameters using a cross-domain pose estimation network, and then performs 3D reconstruction and semantic post-processing to ultimately generate an editable 3D model with component-level semantic information. This method combines robust cross-domain pose estimation with engineering-oriented structured post-processing, solving the problems of existing technologies that rely on precise image annotation and cannot handle diverse image sources, and the lack of semantic and engineering constraints that prevents direct editing of the output model. It achieves fully automated generation of high-quality, structured vehicle 3D models from single or multiple images, which can be directly used for subsequent industrial design and simulation, significantly reducing the professional threshold and time cost.

[0092] Figure 3 Schematic diagram of the process for generating the 3D model of the vehicle provided in this application Figure 2 ,like Figure 3 As shown, in this embodiment... Figure 2 Based on the examples, the specific process of the vehicle 3D model generation method is described in detail, which includes:

[0093] S301. Obtain at least one input image of the target vehicle.

[0094] Specifically, the server first needs to acquire visual data representing the target vehicle.

[0095] The acquisition here generally involves receiving one or more vehicle photos uploaded by users. These photos may come from the Internet, mobile phone shooting, or professional photography, and their angles, lighting, and backgrounds vary.

[0096] After acquiring this data, the server performs standardized format parsing and basic validation.

[0097] Furthermore, depending on the application scenario, in addition to directly acquiring vehicle photos, methods for acquiring input images may also include the following:

[0098] Step A1: Receive the text description or design sketch corresponding to the target vehicle.

[0099] Specifically, when the input is a text description, the server receives natural language input from the user, such as "a pure electric coupe with a streamlined body and hidden door handles." The text contains high-level semantics such as style (streamlined), structural features (hidden door handles), and vehicle type (pure electric coupe).

[0100] When the input is a design sketch, the server receives a two-dimensional line drawing created by the user through the drawing interface. This may only contain the vehicle's outer contour and main feature lines, lacking texture, detail, and accurate perspective.

[0101] The server needs to parse the design framework and proportional intent contained in these imprecise inputs.

[0102] Step A2: Based on the generative artificial intelligence model, generate at least one input image from the text description or design sketch.

[0103] Specifically, for text input, the server can invoke a finely tuned conditional image generation model. During training, this model not only learns text-image alignment but also introduces 3D geometric consistency constraints (e.g., requiring generated images of the same vehicle from different perspectives to correspond to a consistent 3D shape), enabling it to generate a series of geometrically consistent multi-view images (such as frontal, side, frontal, and rear 3 / 4 views) based on a single text.

[0104] For sketch input, the server uses a sketch-to-image translation model based on an encoder-decoder architecture. This model first encodes the lines and structural information of the sketch using a convolutional neural network, and then combines a random noise vector (controlling unspecified detail variations) with the decoder to generate a vehicle image with realistic texture, lighting, and details.

[0105] To ensure that the generated images are suitable for subsequent 3D reconstruction, the server can embed a lightweight geometric rationality discriminator during the generation process to penalize the generation of images that are obviously abnormal in physical structure (such as severely deformed wheels), thus guiding the generation process to produce more "reconstructable" results.

[0106] In some embodiments, the generative artificial intelligence model is a controllable generative model. When the server generates vehicle images based on text or sketches, the generative model used has explicit control over attributes such as vehicle styling and size proportions. Users can guide the generative model to produce vehicle images with differentiation in specific design dimensions (such as sporty or family-oriented) by adjusting different latent control variables, thereby more accurately transforming abstract ideas into concrete visual input.

[0107] In some embodiments, the system includes conditional execution and closed-loop data flow between modules. Specifically, the generative artificial intelligence model is activated to generate an image only when the server determines that the input is text or a sketch; when the input is an existing image, the system directly proceeds to the pose estimation process.

[0108] S302. Based on a pre-trained cross-domain pose estimation network, extract the geometric features of the input image.

[0109] The cross-domain pose estimation network is trained on a simulation image dataset containing the vehicle's 3D model and the ground truth values ​​of camera parameters. The cross-domain pose estimation network includes a cross-domain adaptive module.

[0110] Specifically, the architecture of cross-domain pose estimation networks typically uses a deep convolutional neural network as the backbone (such as variants of ResNet and Vision Transformer), and its training objective is to directly regress camera pose parameters.

[0111] During training, the cross-domain pose estimation network needs to learn to ignore domain-specific appearances such as vehicle color and texture, and instead focus on understanding the common, view-related geometric features of the vehicle category (such as contours, occlusion relationships between parts, and perspective distortion).

[0112] Therefore, when faced with a new image, the geometric features extracted by the cross-domain pose estimation network are highly abstract and compressed visual pattern representations that contribute most to subsequent pose regression.

[0113] In some embodiments, the training strategy for the cross-domain pose estimation network includes a unique simulation data construction pipeline. In a virtual simulation environment, the server programmatically controls parameters such as vehicle model, body material, ambient lighting, weather conditions, background scene, and camera pose to generate batches of highly realistic and diverse ground-value image data. This ensures that the pre-training data fully covers visual variations in the real world, guaranteeing the robustness of the cross-domain pose estimation network.

[0114] S303. Based on the cross-domain adaptive module, domain difference compensation is performed on the geometric features to obtain the supplemented geometric features.

[0115] The compensation process is used to adapt the input image to the actual image domain.

[0116] Specifically, networks trained directly from simulation data will experience performance degradation when faced with real photos or AI-generated images due to domain differences (such as the difference in glossiness between simulation renderings and noise in real photos, the difference in simplicity between simulation backgrounds and clutter in real backgrounds, etc.).

[0117] The role of the cross-domain adaptive module is to dynamically reduce this difference during inference. Its implementation principle can be based on feature alignment.

[0118] Specifically, the cross-domain adaptation module is implemented within the cross-domain pose estimation network (e.g., after some intermediate feature layer). It contains a lightweight domain discriminant subnetwork that distinguishes whether the current input features come from the "simulated domain" or the "target domain" (i.e., whether it is a real image or a generated image).

[0119] The feature extractor of the main network (i.e., the cross-domain pose estimation network) is trained to generate features that can "fool" this domain discriminator, i.e., to make the features as domain-invariant as possible.

[0120] During inference, when an image of unknown origin is input into the server, the cross-domain adaptive module prompts the feature extraction network to adjust its internal representation, actively suppressing non-geometric noise that strongly identifies the image's source (domain) (such as unique texture style and color distribution), while strengthening the expression of the vehicle's geometry.

[0121] This process is called domain difference compensation, and its output is the supplemented geometric features, that is, cleaned-up geometric features that are closer to the simulation domain distribution and thus more suitable for attitude regression.

[0122] S304. Perform regression processing on the compensated geometric features to obtain the camera pose parameters.

[0123] Specifically, the geometric features, after domain adaptive compensation, are fed into the regression head of the network.

[0124] The regression head typically consists of several fully connected layers.

[0125] It maps high-dimensional geometric feature vectors to specific camera pose parameters. For camera extrinsic parameters, it typically outputs a 6-DOF vector, where 3 values ​​represent the camera's position in the world coordinate system (T). x T y T z The other three values ​​represent the camera's rotation angle (usually expressed as a rotation vector or Euler angles, such as R). x R y R z ).

[0126] For camera intrinsic parameters, the output is typically the focal length (f) and the principal point coordinates (c). x c y )wait.

[0127] During training, the server uses the true values ​​of the simulation data as a supervision signal and employs mean squared error (MSE) or a loss function more suitable for rotation (such as geodesic loss on rotation matrices) to constrain the network output.

[0128] Therefore, when a vehicle image is input, the server obtains a set of specific numbers directly from the network through forward propagation, namely the estimated camera position and orientation, thus providing precise viewpoint constraints for 3D reconstruction.

[0129] S305. Based on the camera pose parameters, perform 3D reconstruction on the input image to obtain an initial 3D model.

[0130] Specifically, after obtaining reliable camera pose parameters, the server can perform high-quality 3D reconstruction.

[0131] The goal of the reconstruction is to obtain a triangular mesh representation of the vehicle surface. Specifically, this can be achieved using Neural Radiation Field (NeRF) or its variants.

[0132] A multilayer perceptron network is built in the server. This network takes the coordinates (x, y, z) of a point in three-dimensional space and the viewing direction (d) in the camera pose parameters as input, and outputs the color (c) and volume density (σ) of the point.

[0133] During training, the server uses the camera pose obtained in the previous step to back-project each pixel in the input image into a ray that passes through the scene. It then calculates the predicted pixel color along the ray sampling points using the volume rendering integral formula and compares it with the real pixel color to calculate the loss.

[0134] The volume rendering formula can be simply expressed as:

[0135] C(r) = ∫ T(t)σ(r(t))c(r(t), d) dt,

[0136] Where T(t) is the transmittance.

[0137] By optimizing this network, the geometry of its implicit representation (typically defined by regions with high volume density σ) becomes the vehicle surface.

[0138] After training, the server can extract a triangular mesh from the neural field using the Moving Cubes algorithm, which serves as the initial 3D model.

[0139] The model can reflect the overall shape of the vehicle, but the surface may be noisy and lack clear component boundaries.

[0140] S306. Based on the component semantic segmentation model, the surface of the initial three-dimensional model is segmented to obtain at least two semantic regions corresponding to different components in the vehicle.

[0141] Specifically, in order to make the model editable, the server can call a pre-trained 3D semantic segmentation network to further process the initial 3D model.

[0142] The network takes a discretized representation of the initial 3D model (such as a point cloud or graph structure) as input.

[0143] For example, a server can sample a triangular mesh as a point cloud, where each point contains its three-dimensional coordinates and normal vector.

[0144] The architecture of 3D semantic segmentation networks can adopt PointNet++ or graph convolution-based networks, which are able to process unstructured 3D data and capture local geometric context.

[0145] The network has been trained on a large labeled 3D vehicle dataset (such as a CAD model library containing multiple vehicle models, where each model's facets are manually labeled with categories such as "door", "hood", and "tire").

[0146] When a new initial 3D model is input, the 3D semantic segmentation network predicts a probability distribution for each vertex or face, indicating the likelihood that it belongs to each predefined part category.

[0147] The server assigns a semantic label to each basic unit of the entire grid by taking the category with the highest probability, thereby dividing the entire surface into several connected "semantic regions", each region corresponding to a specific vehicle component.

[0148] S307. Based on the component type and predefined engineering geometric constraints corresponding to each semantic region, optimize the geometry of the semantic region, and generate a three-dimensional vehicle model based on the optimized geometry.

[0149] Among them, the engineering geometric constraints include at least one of the following: symmetry constraints applied to symmetrical components belonging to the left and right sides of the vehicle; surface smoothing processing performed on semantic regions belonging to body panels; and parametric dimensional adjustment constraints performed on semantic regions belonging to wheels.

[0150] Specifically, after obtaining semantic segmentation, the server can perform targeted geometric optimization based on the engineering characteristics of different components. Specifically:

[0151] To address symmetry constraints, the server first detects the vehicle's centerline. For semantic regions marked as symmetrical components (such as the left and right front doors), the server calculates the mirror image of one side of the region to the centerline and uses the mirrored position as the target to perform non-rigid iterative nearest point (ICP) optimization or apply elastic deformation constraints to the vertex positions of the other side of the region, forcing it to be symmetrical and eliminating asymmetry caused by reconstruction noise.

[0152] For surface smoothing, especially for body panels (such as hoods and door panels), which are theoretically smooth and continuous surfaces, the server can employ a smoothing algorithm based on Laplacian coordinates. For mesh vertex v... i Its Laplace coordinate δ i Defined as:

[0153] δ i = v i - Σ(v j ) / N,

[0154] Among them, v j It is its adjacent vertex.

[0155] Smoothing is achieved by minimizing the change in the Laplacian coordinates after vertex position adjustment, thus smoothing out high-frequency noise while maintaining the overall shape.

[0156] The objective function can be expressed as:

[0157] Σ||L(v') - δ|| 2 + λΣ ||v' - v|| 2 Minimize the result.

[0158] Where L is the Laplacian operator, v' is the optimized vertex, v is the original vertex, λ is the weight that preserves the original shape, ||·|| 2 denoted as the Euclidean norm, used to find the square root of the sum of the squares of all elements in a vector.

[0159] To address parametric sizing constraints and support subsequent editing, the server establishes a relationship between parameters (such as wheelbase and hub diameter) and the vertices of the underlying mesh. For example, the wheelbase parameter is associated with the distance between the center points of the front and rear wheels.

[0160] When a user modifies the wheelbase, the server drives the mid-point of the vehicle body associated with this parameter to undergo coordinated displacement through a predefined linear hybrid skin (LBS) or an interpolation deformation function based on radial basis functions (RBF), while keeping the wheels themselves undeformed, thus achieving overall scaling that conforms to physical intuition.

[0161] Furthermore, the method for optimizing based on component type includes the following steps:

[0162] Step B1: If the component type of the semantic region is a car window, perform planar fitting optimization on the semantic region.

[0163] Specifically, the car window is theoretically a flat piece of glass. The server uses principal component analysis or least squares to fit an optimal plane to all vertices within that semantic region.

[0164] The server then projects these vertices onto the fitted plane, forcing the window area to form a flat surface.

[0165] Step B2: If the component type of the semantic region is a wheel hub, perform cylindrical surface fitting optimization on the semantic region.

[0166] Specifically, wheel hubs typically have cylindrical sections.

[0167] The server uses a cylinder fitting algorithm based on Random Sample Consensus (RANSAC) for the vertices in this region. The RANSAC algorithm estimates the parameters of a cylinder model (including the axis direction, a point on the axis, and the radius) by randomly sampling a minimum set of points (e.g., 7 points), calculates the distance from all points to the cylinder surface, and counts the number of in-circuit points (distances less than a threshold).

[0168] After multiple iterations, the model with the most interior points was selected as the optimal cylinder fit.

[0169] Then, the server snaps the vertices of the hub region onto the optimal cylindrical surface, regularizing its geometry.

[0170] Step B3: If the component type of the semantic region is a car body panel, perform non-uniform rational B-spline surface reconstruction optimization on the semantic region.

[0171] Specifically, for complex vehicle body surfaces, the server can use industry-standard non-uniform rational B-spline surfaces (abbreviated as NURBS surfaces) for high-precision reconstruction.

[0172] First, the server extracts a quadrilateral topology from the mesh of the semantic region as the parameter domain. Then, through parameterized inverse calculation, a set of NURBS control points, weights, and node vectors are solved so that the reconstructed NURBS surface most closely approximates the original mesh vertices in the least-squares sense (the specific reconstruction operation is a prior art in this field and will not be described in detail here).

[0173] After reconstruction, the component is represented by a smooth, continuous NURBS surface with precise mathematical expressions, which greatly facilitates subsequent simulation design operations.

[0174] Furthermore, considering that semantic regions typically comprise multiple regions, the specific steps in the optimization process can also be as follows:

[0175] Step C1: Establish a set of parameters that support editing of the vehicle's 3D model.

[0176] The parameter set is associated with the geometric dimensions of at least one semantic region.

[0177] Specifically, after completing the geometric optimization, the server actively analyzes the model, extracts and exposes a set of high-level parameters that users can intuitively edit.

[0178] For example, it can automatically calculate and establish parameters such as total length, total width, total height, wheelbase, front overhang length, rear overhang length, wheel hub diameter, and ground clearance.

[0179] Each parameter does not exist in isolation, but is bound to the coordinates of a set of vertices in a specific semantic region through predefined mathematical relationships. For example, the relationship between the wheelbase parameter and the X coordinates of the front and rear wheel center points on the chassis can be expressed as: wheelbase = difference in X coordinates of the front and rear wheel center points.

[0180] In some embodiments, the server may provide a user interface to modify these parameter values.

[0181] Step C2: In response to modifications to the parameters in the parameter set, automatically update the geometry of the corresponding semantic region.

[0182] Specifically, when a user modifies a parameter (such as changing the wheelbase from 2700mm to 2800mm), the server will trigger the parametric deformation engine.

[0183] The engine calculates the new position of each affected vertex based on the binding relationship between parameters and vertices.

[0184] The deformation process must maintain geometric rationality. For example, when the wheelbase is lengthened, the middle of the car body should be stretched uniformly, while the proportions of the doors and the positions of the wheels should be adjusted accordingly, and auxiliary components such as windows and lights should also be coordinated. This can be achieved by solving an optimization problem with parameter variations as constraints and the goal of minimizing the overall deformation energy of the model (such as preserving local details as much as possible).

[0185] After the update, the entire model quickly responded to design changes while maintaining the semantics of the components and the optimized geometric quality.

[0186] In some embodiments, the post-3D reconstruction processing engine can work in deep collaboration with a parametric solver. Specifically, after obtaining the semantically segmented and fundamentally optimized model, the server can invoke a built-in parametric solver. This solver can reverse-engineer a set of high-level design parameters from the optimized mesh model, such as the vehicle's length, width, height, wheelbase, and track width, thereby transforming the mesh model into a parametric template and enabling rapid variant generation within a preset design space.

[0187] S308: Based on the vehicle's 3D model, render and generate at least one feedback image.

[0188] Specifically, in order to evaluate the consistency between the generated model and the original intent and to achieve closed-loop optimization, the server needs to import the 3D model back into the 2D image space for comparison.

[0189] The server can use a differentiable renderer to achieve this process.

[0190] Given an optimized 3D vehicle model (mesh or NURBS surface), material properties, and a virtual camera pose (which can be the same as the input image pose or newly set), the renderer generates a simulated 2D image through rasterization or ray tracing calculations.

[0191] Since the renderer is differentiable, it means that there are computable gradients from image pixel colors to model vertex positions, material parameters, etc., which provides a basis for subsequent automatic optimization.

[0192] The server can render multi-view images for comprehensive evaluation as needed.

[0193] S309. Determine the difference between the feedback image and the original design intent.

[0194] Specifically, if the original design intent (i.e., the input image) is a real image, the difference can be directly calculated as the pixel-level error between the feedback image and the corresponding viewpoint input image, such as mean squared error (MSE) or perceptual loss (such as the difference after extracting features through a pre-trained VGG network).

[0195] If the original input is a text description, the server uses a text-image matching model (such as the CLIP model) to calculate the embedding vectors of the feedback image and the original text, and then calculates the cosine distance or Euclidean distance between these two vectors as a measure of discrepancy. The larger the distance, the greater the deviation between the model and the text intent.

[0196] S310. Based on the differences, adjust the parameters of the input image generation process or semantic post-processing to iteratively optimize the vehicle 3D model.

[0197] Specifically, the server uses the difference value calculated in the previous step and its backpropagation gradient to adjust the parameters of the upstream process.

[0198] The process of adjusting the input image generation specifically includes:

[0199] If the discrepancy stems from poor image quality, the server will propagate the difference gradient back to the image generation model in step A2 through the differentiable renderer, the 3D model, and so on, adjusting its internal parameters to generate an image that is easier to reconstruct and more in line with the intent.

[0200] The process of adjusting the parameters of semantic post-processing includes:

[0201] If the discrepancy stems from improper geometric optimization (such as excessive smoothing leading to feature loss), the server can propagate the gradient to the constraint weights in S307 (such as the weight λ in Laplacian smoothing), fitting threshold, etc., and automatically adjust the optimization intensity.

[0202] Through several iterations of "generation-reconstruction-rendering-comparison-adjustment", the server can make the final output 3D vehicle model continuously approach the original design intent.

[0203] The vehicle 3D model generation method provided in this disclosure constructs a complete and feasible technical solution by refining the generation of multi-view images from text / sketch, robust pose estimation based on simulation pre-training and domain adaptation, reconstruction of the initial model using neural radiation fields, component-level semantic segmentation of the model, and application of targeted engineering geometry optimization (such as symmetry processing and NURBS surface reconstruction) and closed-loop iteration. This effectively solves the problems of viewpoint generation, cross-domain pose estimation, structural reconstruction, and engineering post-processing encountered in the process of generating high-quality 3D models from creative input or irregular images. It achieves fully automatic generation of vehicle 3D models with clear structure, separable components, standardized geometry, and support for parametric editing, starting from extremely flexible and uncertain inputs, realizing automated generation from conceptual ideas to engineering-usable models.

[0204] Figure 4 This is a schematic diagram of the structure of the vehicle 3D model generation device provided in this application, as shown below. Figure 4 As shown, the vehicle 3D model generation device 400 provided in this embodiment includes:

[0205] The acquisition module 410 is used to acquire at least one input image of the target vehicle;

[0206] The determination module 420 is used to input the input image into the cross-domain pose estimation network and output the corresponding camera pose parameters;

[0207] The generation module 430 is used to perform three-dimensional reconstruction on the input image based on the camera pose parameters to obtain an initial three-dimensional model;

[0208] The optimization module 440 is used to perform semantic post-processing on the initial 3D model to obtain an editable 3D vehicle model, wherein the 3D vehicle model has component-level semantic segmentation information.

[0209] Optionally, the determining module 420 is specifically used to: extract geometric features of the input image based on a pre-trained cross-domain pose estimation network, wherein the cross-domain pose estimation network is trained on a simulation image dataset containing a vehicle 3D model and the ground truth values ​​of camera parameters, and the cross-domain pose estimation network includes a cross-domain adaptive module; based on the cross-domain adaptive module, perform domain difference compensation on the geometric features to obtain supplemented geometric features, wherein the compensation processing is used to adapt to the actual image domain of the input image; and perform regression processing on the compensated geometric features to obtain camera pose parameters.

[0210] Optionally, the optimization module 440 is specifically used to: segment the surface of the initial 3D model based on the component semantic segmentation model to obtain at least two semantic regions corresponding to different components in the vehicle; optimize the geometry of the semantic regions based on the component type corresponding to each semantic region and predefined engineering geometric constraints; and generate a 3D model of the vehicle based on the optimized geometry. The engineering geometric constraints include at least one of the following: symmetry constraints applied to symmetrical components belonging to the left and right sides of the vehicle; surface smoothing processing performed on semantic regions belonging to body panels; and parametric dimensional adjustment constraints performed on semantic regions belonging to wheels.

[0211] Optionally, the optimization module 440 is specifically used to: optimize the semantic region by plane fitting if the component type of the semantic region is a car window; optimize the semantic region by cylindrical surface fitting if the component type of the semantic region is a wheel hub; and optimize the semantic region by non-uniform rational B-spline surface reconstruction if the component type of the semantic region is a car body panel.

[0212] Optionally, the optimization module 440 is specifically used to establish an editable parameter set for the vehicle's 3D model, the parameter set being associated with the geometric dimensions of at least one semantic region; and to automatically update the geometry of the corresponding semantic region in response to modifications to the parameters in the parameter set.

[0213] Optionally, the optimization module 440 is also used to: render and generate at least one feedback image based on the vehicle 3D model; determine the difference between the feedback image and the original design intent; and adjust the parameters of the input image generation process or semantic post-processing based on the difference to iteratively optimize the vehicle 3D model.

[0214] Optionally, the acquisition module 410 is specifically used to receive a text description or design sketch corresponding to the target vehicle; and generate at least one input image based on the text description or design sketch using a generative artificial intelligence model.

[0215] The vehicle 3D model generation device provided in this embodiment can execute the method provided in the above method embodiment. Its implementation principle and technical effect are similar, and will not be described in detail here.

[0216] Figure 5 This is a schematic diagram of the structure of an electronic device provided in this application. Figure 5 As shown, the electronic device 50 includes:

[0217] Processor 51, memory 52, and communication interface 53;

[0218] Memory 52 is used to store executable instructions of processor 51;

[0219] The processor 51 is configured to execute the technical solutions in any of the foregoing method embodiments by executing the aforementioned executable instructions.

[0220] Optionally, the memory 52 can be either standalone or integrated with the processor 51.

[0221] Optionally, when the memory 52 is a device independent of the processor 51, the electronic device 50 may further include:

[0222] Bus 54, memory 52 and communication interface 53 are connected to processor 51 through bus 54 and complete communication with each other. Communication interface 53 is used to communicate with other devices.

[0223] Optionally, the communication interface 53 can be implemented using a transceiver. The communication interface is used to enable communication between the database access device and other devices (e.g., clients, read-write databases, and read-only databases). The memory may include random access memory (RAM) and may also include non-volatile memory, such as at least one disk drive.

[0224] Bus 54 can be a Peripheral Component Interconnect (PCI) bus or an Extended Industry Standard Architecture (EISA) bus, etc. Buses can be categorized as address buses, data buses, control buses, etc. For ease of representation, only one thick line is used in the diagram, but this does not indicate that there is only one bus or one type of bus.

[0225] The processors mentioned above can be general-purpose processors, including central processing units (CPUs), network processors (NPs), etc.; they can also be digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, or discrete hardware components.

[0226] The electronic device is used to execute the technical solutions in any of the foregoing method embodiments. Its implementation principle and technical effect are similar, and will not be described again here.

[0227] This application also provides a readable storage medium storing a computer program thereon, which, when executed by a processor, implements the technical solutions provided in any of the foregoing method embodiments.

[0228] This application also provides a computer program product, including a computer program, which, when executed by a processor, is used to implement the technical solutions provided in any of the foregoing method embodiments.

[0229] Those skilled in the art will understand that all or part of the steps of the above-described method embodiments can be implemented by hardware related to program instructions. The aforementioned program can be stored in a computer-readable storage medium. When executed, the program performs the steps of the above-described method embodiments; and the aforementioned storage medium includes various media capable of storing program code, such as ROM, RAM, magnetic disks, or optical disks.

[0230] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features therein. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of this application.

Claims

1. A method for generating a three-dimensional vehicle model, characterized in that, include: Obtain at least one input image of the target vehicle; The input image is fed into a cross-domain pose estimation network, which outputs the corresponding camera pose parameters. Based on the camera pose parameters, the input image is reconstructed in three dimensions to obtain an initial three-dimensional model; The initial 3D model is subjected to semantic post-processing to obtain an editable 3D vehicle model, wherein the 3D vehicle model has component-level semantic segmentation information.

2. The method according to claim 1, characterized in that, The step of inputting the input image into a cross-domain pose estimation network and outputting the corresponding camera pose parameters includes: Geometric features of the input image are extracted based on a pre-trained cross-domain pose estimation network. The cross-domain pose estimation network is trained on a simulation image dataset containing a vehicle 3D model and camera parameter ground truth values. The cross-domain pose estimation network includes a cross-domain adaptive module. Based on the cross-domain adaptive module, the geometric features are compensated for domain differences to obtain supplemented geometric features, wherein the compensation process is used to adapt to the actual image domain of the input image. The compensated geometric features are subjected to regression processing to obtain the camera pose parameters.

3. The method according to claim 1, characterized in that, The semantic post-processing of the initial 3D model to obtain an editable vehicle 3D model includes: Based on the component semantic segmentation model, the surface of the initial three-dimensional model is segmented to obtain at least two semantic regions corresponding to different components in the vehicle; Based on the component type corresponding to each semantic region and predefined engineering geometric constraints, the geometry of the semantic region is optimized, and the vehicle 3D model is generated based on the optimized geometry. The engineering geometric constraints include at least one of the following: Symmetry constraints are applied to symmetrical components belonging to the left and right sides of the vehicle; Surface smoothing processing is performed on semantic regions belonging to body panels; Parametric dimensional adjustment constraints are applied to the semantic region belonging to the wheel.

4. The method according to claim 3, characterized in that, The optimization of the geometry of each semantic region based on the component type and predefined engineering geometric constraints includes: If the component type of the semantic region is a car window, then the semantic region is optimized by plane fitting. If the component type of the semantic region is a wheel hub, then the semantic region is optimized by cylindrical surface fitting; If the component type of the semantic region is a vehicle body panel, the semantic region is optimized by non-uniform rational B-spline surface reconstruction.

5. The method according to claim 3, characterized in that, The optimization of the geometry of each semantic region based on the component type and predefined engineering geometric constraints includes: Establish an editable parameter set for the vehicle's 3D model, the parameter set being associated with the geometric dimensions of at least one of the semantic regions; In response to modifications to the parameters in the parameter set, the geometry of the corresponding semantic region is automatically updated.

6. The method according to claim 3, characterized in that, After semanticizing the initial 3D model to obtain an editable vehicle 3D model, the process further includes: Based on the vehicle's 3D model, at least one feedback image is rendered and generated. Determine the difference between the feedback image and the original design intent; Based on the differences, the parameters of the input image generation process or the semantic post-processing are adjusted to iteratively optimize the vehicle 3D model.

7. The method according to any one of claims 1 to 6, characterized in that, The acquisition of at least one input image of the target vehicle includes: Receive a text description or design sketch corresponding to the target vehicle; Based on a generative artificial intelligence model, at least one input image is generated from the text description or design sketch.

8. A vehicle three-dimensional model generation device, characterized in that, include: The acquisition module is used to acquire at least one input image of the target vehicle; The determination module is used to input the input image into the cross-domain pose estimation network and output the corresponding camera pose parameters; The generation module is used to perform three-dimensional reconstruction on the input image based on the camera pose parameters to obtain an initial three-dimensional model; An optimization module is used to perform semantic post-processing on the initial 3D model to obtain an editable 3D vehicle model, wherein the 3D vehicle model has component-level semantic segmentation information.

9. An electronic device, characterized in that, include: Processor, memory, communication interface; The memory is used to store the executable instructions of the processor; The processor is configured to execute the vehicle 3D model generation method according to any one of claims 1 to 7 by executing the executable instructions.

10. A readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the vehicle three-dimensional model generation method according to any one of claims 1 to 7.

Images