Generative Three-Dimensional Shape Compressed Representation Method, Apparatus, Device, Medium and Product

Abstract

The present application relates to a method, apparatus, device, medium and product for generating a compressed representation of a three-dimensional shape. The method includes: obtaining three-dimensional shape data, and normalizing the three-dimensional shape data to obtain a feature space; randomly sampling the three-dimensional shape data in the feature space to obtain an initial signed distance value; numerically truncating the initial signed distance value according to a preset range space to obtain a candidate signed distance value; iteratively optimizing the candidate signed distance value according to a pre-constructed loss optimization model to obtain orthogonal tri-vector data; and the orthogonal tri-vector data is used to compress and represent the three-dimensional shape data. By using this method, it is possible to be simply and efficiently applied to the three-dimensional shape generation task while maintaining geometric details.

Description

Technical Field

[0001] The present application relates to the technical field of three-dimensional shape processing, and in particular, to a method, apparatus, device, medium, and product for generating a compressed representation of three-dimensional shapes. Background Art

[0002] With the development of three-dimensional shape processing technology, the demand for efficiently generating high-quality three-dimensional models is increasing day by day. However, traditional three-dimensional shape generation methods usually adopt the expression of three-dimensional shapes based on signed distance fields. This method faces challenges of high-dimensional data complexity and computational cost. At the same time, the acquisition of high-quality three-dimensional data is relatively expensive and time-consuming. Therefore, how to generate three-dimensional shapes more concisely and efficiently while maintaining geometric details has become an urgent problem to be solved. Summary of the Invention

[0003] Based on this, in view of the above technical problems, it is necessary to provide a method, apparatus, computer device, computer-readable storage medium, and computer program product for generating a compressed representation of three-dimensional shapes, which can effectively solve the balance problem between parameter efficiency and geometric fidelity and can be efficiently applied to the compression and generation tasks of three-dimensional shapes.

[0004] In a first aspect, the present application provides a method for generating a compressed representation of three-dimensional shapes, including:

[0005] Obtain three-dimensional shape data, and normalize the three-dimensional shape data to obtain a feature space;

[0006] Randomly sample the three-dimensional shape data in the feature space to obtain an initial signed distance value;

[0007] Perform numerical truncation on the initial signed distance value according to a preset range space to obtain a candidate signed distance value;

[0008] Iteratively optimize the candidate signed distance value according to a pre-constructed loss optimization model to obtain orthogonal tri-vector data; the orthogonal tri-vector data is used to compress and represent the three-dimensional shape data.

[0009] In one embodiment, the initial signed distance value includes: a first signed distance value, a second signed distance value, and a third signed distance value; the randomly sampling the three-dimensional shape data in the feature space to obtain an initial signed distance value includes:

[0010] Randomly sample the surface area of the three-dimensional shape data in the feature space to obtain the first signed distance value;

[0011] Randomly sample the area near the surface of the three-dimensional shape data in the feature space to obtain the second signed distance value;

[0012] Perform random sampling within the feature space to obtain the third symbol distance value.

[0013] In one embodiment, the numerical truncation of the initial symbol distance value according to the preset range space to obtain the candidate symbol distance value includes:

[0014] Screen the initial symbol distance value according to the numerical range corresponding to the preset range space to obtain the symbol distance value to be truncated;

[0015] Perform numerical truncation on the symbol distance value to be truncated in the initial symbol distance value to obtain the candidate symbol distance value.

[0016] In one embodiment, the iterative optimization of the candidate symbol distance value according to the pre-constructed loss optimization model to obtain the orthogonal three-vector data; the orthogonal three-vector data is used to compress and represent the three-dimensional shape data, including:

[0017] Randomly sample from a preset point set to obtain sampling point data;

[0018] Interpolate in the candidate symbol distance value according to the coordinate data of the sampling point data to obtain an interpolation result;

[0019] Perform loss optimization on the interpolation result and the true symbol distance value of the sampling point data according to the loss optimization model to obtain the orthogonal three-vector data.

[0020] In one embodiment, the method further includes:

[0021] Perform shape reorganization on the orthogonal three-vector data to obtain one-dimensional feature vector data;

[0022] Iteratively train the one-dimensional feature vector data according to the pre-constructed initial generation model to obtain a candidate generation model;

[0023] Establish a training distribution change mapping according to the training data of the candidate generation model; wherein, the training distribution change mapping is used to represent the mapping from the standard normal distribution to the training data distribution;

[0024] Perform loss optimization on the candidate generation model according to the change of the loss value of the distribution change mapping to obtain a target generation model; wherein, the target generation model is used to generate a three-dimensional shape represented by the orthogonal three-vector data.

[0025] In one embodiment, the method further includes:

[0026] Obtain the random noise to be restored; the random noise to be restored is a random noise with the same shape as the orthogonal three-vector data;

[0027] Iteratively restore the random noise to be restored and the preset position encoding to obtain the restored noise;

[0028] Perform variable sampling processing on the restored noise to obtain the restored one-dimensional vector data;

[0029] Reconstruct the restored one-dimensional vector data to obtain the signed distance value grid data;

[0030] Restore the signed distance value grid through the cube tracing model to obtain the triangular mesh data.

[0031] In a second aspect, the present application also provides a generative three-dimensional shape compression representation device, including:

[0032] An acquisition module, configured to acquire three-dimensional shape data and normalize the three-dimensional shape data to obtain a feature space;

[0033] A random sampling module, configured to randomly sample the three-dimensional shape data in the feature space to obtain an initial signed distance value;

[0034] A numerical truncation module, configured to numerically truncate the initial signed distance value according to a preset range space to obtain a candidate signed distance value;

[0035] A compression representation module, configured to iteratively optimize the candidate signed distance value according to a pre-constructed loss optimization model to obtain orthogonal three-vector data; the orthogonal three-vector data is used to compress and represent the three-dimensional shape data.

[0036] In a third aspect, the present application also provides a computer device, including a memory and a processor, the memory stores a computer program, and when the processor executes the computer program, the following steps are implemented:

[0037] Acquire three-dimensional shape data and normalize the three-dimensional shape data to obtain a feature space;

[0038] Randomly sample the three-dimensional shape data in the feature space to obtain an initial signed distance value;

[0039] Numerically truncate the initial signed distance value according to a preset range space to obtain a candidate signed distance value;

[0040] Iteratively optimize the candidate signed distance value according to a pre-constructed loss optimization model to obtain orthogonal three-vector data; the orthogonal three-vector data is used to compress and represent the three-dimensional shape data.

[0041] In a fourth aspect, the present application further provides a computer-readable storage medium, on which a computer program is stored. When the computer program is executed by a processor, the following steps are implemented:

[0042] Obtain three-dimensional shape data, and normalize the three-dimensional shape data to obtain a feature space;

[0043] Randomly sample the three-dimensional shape data in the feature space to obtain an initial signed distance value;

[0044] Numerically truncate the initial signed distance value according to a preset range space to obtain a candidate signed distance value;

[0045] Iteratively optimize the candidate signed distance value according to a pre-constructed loss optimization model to obtain orthogonal tri-vector data; the orthogonal tri-vector data is used to compress and represent the three-dimensional shape data.

[0046] In a fifth aspect, the present application further provides a computer program product, including a computer program. When the computer program is executed by a processor, the following steps are implemented:

[0047] Obtain three-dimensional shape data, and normalize the three-dimensional shape data to obtain a feature space;

[0048] Randomly sample the three-dimensional shape data in the feature space to obtain an initial signed distance value;

[0049] Numerically truncate the initial signed distance value according to a preset range space to obtain a candidate signed distance value;

[0050] Iteratively optimize the candidate signed distance value according to a pre-constructed loss optimization model to obtain orthogonal tri-vector data; the orthogonal tri-vector data is used to compress and represent the three-dimensional shape data.

[0051] The above-mentioned generative three-dimensional shape compression representation method, device, computer device, computer-readable storage medium, and computer program product obtain three-dimensional shape data, normalize the three-dimensional shape data to obtain a feature space, randomly sample the three-dimensional shape data in the feature space to obtain an initial symbol distance value, perform numerical truncation on the initial symbol distance value according to a preset range space to obtain a candidate symbol distance value, and iteratively optimize the candidate symbol distance value according to a pre-constructed loss optimization model to obtain orthogonal tri-vector data. The orthogonal tri-vector data is used to compress and represent the three-dimensional shape data. Therefore, by normalizing the obtained three-dimensional shape data, a feature space of the three-dimensional shape data is obtained, and random sampling is performed within this feature space to obtain an initial symbol distance value that can represent the three-dimensional shape data. Then, according to the preset range space, numerical truncation is performed on the initial symbol distance value, and the initial symbol distance values near each shape in the three-dimensional shape data can be effectively intercepted as candidate symbol distance values. Then, by iteratively optimizing the candidate symbol distance values according to the pre-constructed loss optimization model, the symbol distance field corresponding to the candidate symbol distance values can be decomposed into orthogonal tri-vector data, greatly reducing the storage requirements while maintaining high geometric fidelity of the three-dimensional shape data. BRIEF DESCRIPTION OF THE DRAWINGS

[0052] In order to more clearly illustrate the technical solutions in the embodiments of the present application or related technologies, the following will briefly introduce the drawings required for use in the description of the embodiments of the present application or related technologies. Obviously, the drawings in the following description are only some embodiments of the present application. For those of ordinary skill in the art, without creative efforts, other related drawings can also be obtained based on these drawings.

[0053] Figure 1 FIG. is an application environment diagram of the generative three-dimensional shape compression representation method in an embodiment;

[0054] Figure 2 FIG. is a flowchart of the generative three-dimensional shape compression representation method in an embodiment;

[0055] Figure 3 FIG. is a flowchart of the generative three-dimensional shape compression representation method in an embodiment;

[0056] Figure 4 FIG. is a flowchart of the generative three-dimensional shape compression representation method in another embodiment;

[0057] Figure 5 FIG. is a flowchart of the generative three-dimensional shape compression representation method in another embodiment;

[0058] Figure 6 FIG. is a flowchart of the generative three-dimensional shape compression representation method in another embodiment;

[0059] Figure 7 It is a schematic flowchart of a generative three-dimensional shape compression representation method in another embodiment;

[0060] Figure 8 It is a schematic diagram of the reconstruction effect of a generative three-dimensional shape compression representation method in another embodiment;

[0061] Figure 9 It is a schematic diagram of the reconstruction effect of a generative three-dimensional shape compression representation method in another embodiment;

[0062] Figure 10 It is a schematic diagram of the reconstruction effect of a generative three-dimensional shape compression representation method in another embodiment;

[0063] Figure 11 It is a structural block diagram of a generative three-dimensional shape compression representation device in one embodiment;

[0064] Figure 12 It is an internal structure diagram of a computer device in one embodiment. Detailed implementation manners

[0065] In order to make the objectives, technical solutions and advantages of the present application clearer, the present application will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application and are not used to limit the present application.

[0066] The generative three-dimensional shape compression representation method provided by the embodiments of the present application can be applied to, for example Figure 1In the application environment shown. Among them, the terminal 102 communicates with the server 104 through the network. The data storage system can store the data that the server 104 needs to process. The data storage system can be integrated on the server 104, or can be placed on the cloud or other network servers. The server 104 obtains three-dimensional shape data, normalizes the three-dimensional shape data to obtain a feature space; randomly samples the three-dimensional shape data in the feature space to obtain an initial symbol distance value; numerically truncates the initial symbol distance value according to a preset range space to obtain a candidate symbol distance value; iteratively optimizes the candidate symbol distance value according to a pre-constructed loss optimization model to obtain orthogonal three-vector data; the orthogonal three-vector data is used to compress and represent the three-dimensional shape data. Among them, the terminal 102 can be, but is not limited to, various personal computers, laptop computers, smart phones, tablet computers, Internet of Things devices, and portable wearable devices. The Internet of Things devices can be smart speakers, smart TVs, smart air conditioners, smart in-vehicle devices, projection devices, etc. The portable wearable devices can be smart watches, smart bracelets, head-mounted devices, etc. The head-mounted device can be a virtual reality (VR) device, an augmented reality (AR) device, smart glasses, etc. The server 104 can be an independent physical server, or a server cluster or distributed system composed of multiple physical servers, or a cloud server that provides cloud computing services.

[0067] In an exemplary embodiment, as Figure 2 shown, a generative three-dimensional shape compression representation method is provided. Taking the method applied to Figure 1 the service 104 in

[0068] Step S202, obtain three-dimensional shape data, and normalize the three-dimensional shape data to obtain a feature space.

[0069] Among them, the three-dimensional shape data can be a watertight three-dimensional shape, or three-dimensional shape data of other representation types.

[0070] In some embodiments, the three-dimensional shape data can be obtained from a three-dimensional shape dataset, or the three-dimensional shape data can be obtained by real-time scanning of an entity through other third-party tools, and is not limited thereto.

[0071] In some embodiments, the three-dimensional shape data is normalized according to a preset range, so as to normalize the three-dimensional shape data into a unit cube (-0.5, 0.5) to obtain a feature space near the three-dimensional shape data, for use in subsequent more accurate random sampling steps.

[0072] Step S204: Randomly sample the three-dimensional shape data in the feature space to obtain the initial signed distance values.

[0073] Among them, random sampling refers to the process of randomly sampling in different regions of the feature space. Through such a random sampling process in different regions, the initial signed distance values of the obtained points can more accurately describe the continuous signed distance field of each three-dimensional shape in the three-dimensional shape data.

[0074] In some embodiments, the initial signed distance values include: the first signed distance value, the second signed distance value, and the third signed distance value; randomly sampling the three-dimensional shape data in the feature space to obtain the initial signed distance values includes: randomly sampling the surface area of the three-dimensional shape data in the feature space to obtain the first signed distance value; randomly sampling the area near the surface of the three-dimensional shape data in the feature space to obtain the second signed distance value; randomly sampling in the feature space to obtain the third signed distance value.

[0075] Among them, the first signed distance value, the second signed distance value, and the third signed distance value are respectively the signed distance values obtained within different ranges of each shape in the feature space.

[0076] In some embodiments, 2 million points are sampled on the shape surface of each shape of the three-dimensional shape data to obtain the first signed distance value, and the signed distance value of the first type of points is 0. 2 million points are sampled in the area near the shape surface of each shape of the three-dimensional shape data, and the corresponding signed distance values are calculated as the second signed distance value, and 2 million points are randomly sampled in the feature space and the corresponding signed distance values are calculated. A total of 6 million points are sampled for each shape and the signed distance values corresponding to all points are obtained. These numbers of points can accurately describe the continuous signed distance field of the three-dimensional shape.

[0077] In some embodiments, given an input three-dimensional shape The corresponding initial three-vector is shown in formula (1). 2 million points in the initial signed distance values are evenly distributed in the feature space and another 4 million points are concentrated near the shape surface to capture more detailed geometric details.

[0078] (1)

[0079] Step S206: Numerically truncate the initial signed distance values according to a preset range space to obtain candidate signed distance values.

[0080] Among them, the preset range space is a numerical range space close to the signed distance values of the shape surface of each shape in the three-dimensional shape data.

[0081] In some embodiments, numerical truncation is performed on the initial symbol distance values according to a preset range space to obtain candidate symbol distance values, including: screening the initial symbol distance values according to the numerical range corresponding to the preset range space to obtain symbol distance values to be truncated; performing numerical truncation on the symbol distance values to be truncated in the initial symbol distance values to obtain candidate symbol distance values.

[0082] Wherein, numerical truncation means removing symbol distance values that do not belong to the preset range space.

[0083] In some embodiments, according to the numerical range corresponding to the preset range space , the initial symbol distance values are screened, and the initial symbol distance values whose numerical values are not within are used as symbol distance values to be truncated, and the symbol distance values to be truncated among all the initial symbol distance values are truncated to obtain candidate symbol distance values within the preset range space .

[0084] In some embodiments, for each sampling point , we calculate its true SDF value , and truncate it to within [−0.05, 0.05]. The three vectors approximate the truncated SDF value of the point through the following formula (2):

[0085] (2)

[0086] Wherein, are the elements of the vector corresponding to the coordinates of the point . For any non-integer point in three-dimensional space, the corresponding value can be obtained from each vector by linear interpolation. To optimize this representation, we minimize the difference between the reconstructed SDF value and the true SDF value .

[0087] In this embodiment, considering that most of the information of the three-dimensional shape is concentrated on the geometric surface, therefore, the symbol distance values far from the surface can be truncated to efficiently capture the key geometric patterns of each shape with a small number of parameters.

[0088] Step S208, iteratively optimize the candidate symbol distance values according to a pre-constructed loss optimization model to obtain orthogonal three-vector data; the orthogonal three-vector data is used to compress and represent three-dimensional shape data.

[0089] Among them, the orthogonal tri-vector data refers to Tri-Vectors. Compressing and representing three-dimensional shape data means that through Tri-Vectors, the core features of three-dimensional shapes can be expressed in a concise form. Also, the process of expressing three-dimensional shape data as Tri-Vectors can be regarded as the process of compressing three-dimensional shape data.

[0090] In some embodiments, the candidate symbol distance values are iteratively optimized according to a pre-constructed loss optimization model to obtain orthogonal tri-vector data. The orthogonal tri-vector data is used to compress and represent three-dimensional shape data, including: randomly sampling from a preset point set to obtain sampling point data; interpolating in the candidate symbol distance values according to the coordinate data of the sampling point data to obtain an interpolation result; and performing loss optimization on the interpolation result and the true symbol distance value of the sampling point data according to the loss optimization model to obtain orthogonal tri-vector data.

[0091] Among them, the loss optimization model is a loss optimization model used to smooth the representation of candidate symbol distance values and reduce the differences between adjacent elements within each vector. The parameters in the loss optimization model can be optimized using the gradient descent algorithm to improve the representation accuracy of the orthogonal tri-vector data.

[0092] In some embodiments, the loss optimization model includes: a first loss sub-model and a second loss sub-model, where the objective function of the first loss sub-model is defined as the quadratic distance loss as shown in the following formula (3):

[0093] (3)

[0094] Among them, is the total number of sampling points.

[0095] It should be noted that in the second loss sub-model a quadratic regularization term is introduced to suppress outliers in the representation, and a variational regularization term is introduced to make the representation smoother, as shown below:

[0096] (4)

[0097] Among them, represents the difference between adjacent elements within each vector, while represents the difference between corresponding position elements among different components. Finally, the loss optimization model in the optimization process combines the above first loss sub-model and second loss sub-model to balance the precise compression of the shape and the smoothness of Tri-Vectors, as shown in the following formula (5):

[0098] (5)

[0099] Among them, , to balance the entire optimization process. Tri-Vectors are characterized by minimizing and can finally be expressed in the form of the following formula (6):

[0100] (6)

[0101] It should be noted that by iteratively using the gradient descent algorithm to optimize these factors, a compact tri-vector representation can be obtained, which can accurately approximate a shape. This method can efficiently store and reconstruct the shape without any neural network and capture key geometric details with high fidelity.

[0102] In some embodiments, randomly sample from a preset point set to obtain any non-integer point in the feature space as sampling point data. Use linear interpolation to obtain corresponding values from each vector of the coordinate data of the sampling point data, so as to obtain an interpolation result. Optimize the loss of the interpolation result and the true signed distance value of the sampling point data according to the loss optimization model and in combination with the gradient descent algorithm to obtain orthogonal tri-vector data.

[0103] In the above generative three-dimensional shape compression representation method, by obtaining three-dimensional shape data and normalizing the three-dimensional shape data, a feature space is obtained; randomly sample the three-dimensional shape data in the feature space to obtain an initial signed distance value; perform numerical truncation on the initial signed distance value according to a preset range space to obtain a candidate signed distance value; iteratively optimize the candidate signed distance value according to a pre-constructed loss optimization model to obtain orthogonal tri-vector data; the orthogonal tri-vector data is used to compress and represent the three-dimensional shape data. Therefore, by normalizing the obtained three-dimensional shape data, the feature space of the three-dimensional shape data can be obtained, and then randomly sample within this feature space to obtain an initial signed distance value that can represent the three-dimensional shape data. Then, perform numerical truncation on the initial signed distance value according to the preset range space, and the initial signed distance values near each shape in the three-dimensional shape data can be effectively intercepted as candidate signed distance values. Then, iteratively optimize the candidate signed distance values according to the pre-constructed loss optimization model, and the signed distance field corresponding to the candidate signed distance values can be decomposed into orthogonal tri-vector data, greatly reducing the storage requirement while maintaining the high geometric fidelity of the three-dimensional shape data.

[0104] In an exemplary embodiment, as Figure 3 shown, the generative three-dimensional shape compression representation method further includes: steps S302 to S308. Among them:

[0105] Step S302, perform shape reorganization on the orthogonal tri-vector data to obtain one-dimensional feature vector data.

[0106] Among them, shape reshaping refers to converting the representation of each shape in the orthogonal three-vector data into a set of one-dimensional vectors, that is, each shape is represented by multiple one-dimensional vectors of the same size.

[0107] In some embodiments, by reshaping the representation of each shape in the orthogonal three-vector data, it is converted into a set of one-dimensional vectors to facilitate subsequent training of the generation model.

[0108] Step S304: Iteratively train the one-dimensional feature vector data according to a pre-constructed initial generation model to obtain a candidate generation model.

[0109] Among them, the initial generation model is constructed based on the generation model of flow matching, but is not limited thereto.

[0110] In some embodiments, when iteratively training the one-dimensional feature vector data according to the pre-constructed initial generation model, the training process of each step is as follows: At Randomly sample a time parameter , and sample a noise of the same size as the data from the standard normal distribution , then according to Mix and to obtain , after adding positional encoding, send and time into the neural network at the same time, so that the neural network predicts to speed to obtain a candidate generation model.

[0111] Step S306: Establish a training distribution change mapping according to the training data of the candidate generation model; among them, the training distribution change mapping is used to represent the mapping from the standard normal distribution to the training data distribution.

[0112] Among them, the training data of the candidate generation model refers to the training parameters obtained after the candidate generation model trains the network parameters through a large number of iterations.

[0113] In some embodiments, after training the network parameters through a large number of iterations, a mapping from the standard normal distribution to the training data distribution can be established to more intuitively judge whether the model converges during the optimization process.

[0114] Step S308: Optimize the loss of the candidate generation model according to the change of the loss value of the distribution change mapping to obtain a target generation model; among them, the target generation model is used to generate a three-dimensional shape based on the orthogonal three-vector data representation.

[0115] In some embodiments, it is determined whether the candidate generation model converges by examining the change in the loss value of the distribution change mapping. After convergence, a target generation model can be obtained that can be used to generate a three-dimensional shape characterized by orthogonal tri-vector data Tri-Vectors.

[0116] In some embodiments, the method further includes: obtaining random noise to be restored; the random noise to be restored is random noise having the same shape as the orthogonal tri-vector data; iteratively restoring the random noise to be restored and the preset positional encoding to obtain restored noise; performing variable sampling processing on the restored noise to obtain restored one-dimensional vector data; reconstructing the restored one-dimensional vector data to obtain signed distance value grid data; and restoring the signed distance value grid through a cube tracing model to obtain triangular grid data.

[0117] In some embodiments, first, a random noise to be restored having the same shape as the Tri-Vectors shape (i.e., the orthogonal tri-vector data shape) is sampled from a standard normal distribution , and then the positional encoding is added, and it is gradually restored into restored noise through multiple iterations in the network , and a Tri-Vectors representation of a three-dimensional shape is obtained. Then, the shape of the one-dimensional vector set obtained by sampling is changed to restore it into the form of Tri-Vectors to obtain restored one-dimensional vector data. After that, the restored one-dimensional vector data is reconstructed by querying the grid center point to obtain signed distance value grid data, and thus signed distance value grid data with any resolution can be reconstructed. Finally, through the Marching Cube algorithm (cube tracing model), the signed distance grid data can be restored into a triangular grid three-dimensional shape representation to obtain triangular grid data.

[0118] In this embodiment, the optimized orthogonal tri-vector data Tri-Vectors can achieve a reconstruction effect superior to the prior art under multiple parameter configurations, and significantly reduce the number of model parameters and computational complexity. It can adapt to any resolution: The orthogonal tri-vector data Tri-Vectors is resolution-independent. By adjusting the size of the rank and the vector resolution, it can flexibly adapt to reconstruction tasks from low resolution to high resolution. This flexibility enables it to achieve high-quality expression across resolutions without relying on a specific fusion module. It can also have rich scalability: By introducing a color field or adding an additional time dimension, the compression and expression of textured shapes or dynamic shapes can be achieved. And it has robustness to noise and missing information: In the case of missing information, the reconstruction can still recover a relatively good geometric shape. Efficient and flexible reconstruction: By simply performing linear interpolation and linear combination on one-dimensional vectors, geometric shapes with any resolution and retaining details can be reconstructed. Through the above advantages, the proposed solution using orthogonal tri-vector data to generate a 3D shape generation model not only solves the problems of low efficiency, insufficient details, and poor effects for complex shapes in the prior art, but also significantly improves the reconstruction efficiency and quality of 3D shapes, becoming a 3D shape representation solution with both high efficiency, simplicity, and robustness.

[0119] To more clearly understand the solution of this application, here it is described in conjunction with Figure 4 , Figure 5 , Figure 6 , Figure 7 and Figure 8 as follows:

[0120] First, Figure 4 shows the process of compressing and expressing a 3D watertight shape into 3D orthogonal vector data Tri-Vectors. After randomly sampling and numerically truncating the 3D shape, candidate signed distance values are obtained, and then through iterative optimization, the orthogonal tri-vector data is obtained.

[0121] Then, as shown in Figure 5 , the orthogonal tri-vector data generated is reshaped to obtain the one-dimensional feature vector data shown in Figure 4 . Figure 5 Figure 5 shows the process of training a 3D shape generation model using 3D orthogonal vector data Tri-Vectors to obtain the target generation model.

[0122] In some embodiments, as shown in Figure 5 , first, the representation of each shape is converted into a set of one-dimensional vectors (i.e., each shape is represented by multiple one-dimensional vectors of the same size), and then a flow-matching-based generation model is used for training. The training process for each step is as follows: In Randomly sample a time parameter , sample a noise with the same size as the data from the standard normal distribution , and then according to , mix and to obtain . After adding positional encoding, send and time into the neural network simultaneously. We hope that the neural network predicts to the speed of . After training the network parameters through a large number of iterations, a mapping from the standard normal distribution to the training data distribution can be established. Finally, by checking the change of the loss value, it can be judged whether the model converges. After convergence, a target generation model that can be used to generate three-dimensional shapes based on Tri-Vectors representation is obtained.

[0123] After Figure 5 training to obtain the target generation model, Figure 6 and Figure 7 show the sampling process using the trained target generation model. First, sample a random noise with the same shape as the Tri-Vectors from the standard normal distribution in Figure 6 , then add positional encoding, and gradually restore it to through multiple iterations in the network to obtain a Tri-Vectors representation of a three-dimensional shape. Then change the shape of the one-dimensional vector set obtained by sampling, and restore the one-dimensional vector set to in the form of three-dimensional orthogonal vectors Tri-Vectors in Figure 7 . After that, the signed distance value grid with any resolution can be reconstructed through the method of querying the grid center point, and finally the signed distance grid can be restored to the triangular mesh three-dimensional shape expression through the MarchingCube algorithm.

[0124] In some embodiments, Figure 8 shows the result of expanding the TriVectors three-dimensional orthogonal vector data to compress and reconstruct textured three-dimensional shapes.

[0125] In some embodiments, Figure 9 shows the result of compressing and reconstructing a sequence of deformable three-dimensional shapes.

[0126] In some embodiments, Figure 10The figure shows a target generation model trained for a gallery of applications using more orthogonal tri-vectors, the reconstruction results of reconstructing the texture shapes in the gallery using the target generation model, including the reconstruction results of deformable shapes.

[0127] It should be understood that although the steps in the flowcharts involved in the above-described embodiments are shown in sequence according to the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless otherwise clearly stated in this document, the execution of these steps has no strict order limit, and these steps can be executed in other orders. Moreover, at least some of the steps in the flowcharts involved in the above-described embodiments may include multiple steps or multiple stages. These steps or stages are not necessarily executed at the same time, but can be executed at different times. The execution order of these steps or stages is not necessarily sequential, but can be executed alternately or in turn with at least some of the steps or stages in other steps or other steps.

[0128] Based on the same inventive concept, an embodiment of the present application also provides a generative three-dimensional shape compression representation device for implementing the above-mentioned generative three-dimensional shape compression representation method. The solution provided by this device for solving problems is similar to the solution described in the above method. Therefore, the specific limitations in one or more embodiments of the generative three-dimensional shape compression representation device provided below can refer to the limitations on the generative three-dimensional shape compression representation method in the above text, and will not be repeated here.

[0129] In an exemplary embodiment, as Figure 11 shown, a generative three-dimensional shape compression representation device 1100 is provided, including: an acquisition module 1101, a random sampling module 1102, a numerical truncation module 1103, and a compression representation module 1104, where:

[0130] The acquisition module 1101 is configured to acquire three-dimensional shape data and normalize the three-dimensional shape data to obtain a feature space;

[0131] The random sampling module 1102 is configured to randomly sample the three-dimensional shape data in the feature space to obtain an initial symbol distance value;

[0132] The numerical truncation module 1103 is configured to numerically truncate the initial symbol distance value according to a preset range space to obtain a candidate symbol distance value;

[0133] The compression representation module 1104 is configured to iteratively optimize the candidate symbol distance value according to a pre-constructed loss optimization model to obtain orthogonal tri-vector data; the orthogonal tri-vector data is used to compress and represent the three-dimensional shape data.

[0134] In some embodiments, the initial signed distance values include: a first signed distance value, a second signed distance value, and a third signed distance value; the random sampling module 1102 is further configured to randomly sample the surface region of the three-dimensional shape data in the feature space to obtain the first signed distance value; randomly sample the region near the surface of the three-dimensional shape data in the feature space to obtain the second signed distance value; and randomly sample in the feature space to obtain the third signed distance value.

[0135] In some embodiments, the numerical truncation module 1103 is further configured to screen the initial signed distance values according to the numerical range corresponding to the preset range space to obtain the signed distance values to be truncated; and perform numerical truncation on the signed distance values to be truncated in the initial signed distance values to obtain candidate signed distance values.

[0136] In some embodiments, the compressed representation module 1104 is further configured to randomly sample from a preset point set to obtain sampled point data; perform interpolation on the candidate signed distance values according to the coordinate data of the sampled point data to obtain an interpolation result; and perform loss optimization on the interpolation result and the true signed distance value of the sampled point data according to the loss optimization model to obtain orthogonal three-vector data.

[0137] In some embodiments, the apparatus further includes: a generation model training module, configured to perform shape reconstruction on the orthogonal three-vector data to obtain one-dimensional feature vector data; perform iterative training on the one-dimensional feature vector data according to a pre-constructed initial generation model to obtain a candidate generation model; establish a training distribution change mapping according to the training data of the candidate generation model; wherein the training distribution change mapping is used to represent the mapping from the standard normal distribution to the training data distribution; and perform loss optimization on the candidate generation model according to the change of the loss value of the distribution change mapping to obtain a target generation model; wherein the target generation model is used to generate a three-dimensional shape represented by the orthogonal three-vector data.

[0138] In some embodiments, the apparatus further includes: a restoration module, configured to obtain random noise to be restored; the random noise to be restored is random noise having the same shape as the orthogonal three-vector data; perform iterative restoration on the random noise to be restored and a preset position encoding to obtain restored noise; perform variable sampling processing on the restored noise to obtain restored one-dimensional vector data; perform reconstruction on the restored one-dimensional vector data to obtain signed distance value grid data; and restore the signed distance value grid through a cube tracing model to obtain triangular mesh data.

[0139] In the above generative three-dimensional shape compression representation device, by acquiring three-dimensional shape data and normalizing the three-dimensional shape data, a feature space is obtained; in the feature space, random sampling is performed on the three-dimensional shape data to obtain an initial symbol distance value; according to a preset range space, numerical truncation is performed on the initial symbol distance value to obtain a candidate symbol distance value; according to a pre-constructed loss optimization model, iterative optimization is performed on the candidate symbol distance value to obtain orthogonal tri-vector data; the orthogonal tri-vector data is used to compress and represent the three-dimensional shape data. Therefore, by normalizing the acquired three-dimensional shape data, a feature space of the three-dimensional shape data is obtained, and random sampling is performed within this feature space to obtain an initial symbol distance value that can represent the three-dimensional shape data. Then, according to the preset range space, numerical truncation is performed on the initial symbol distance value, and the initial symbol distance values near each shape in the three-dimensional shape data can be effectively intercepted as candidate symbol distance values. Then, according to the pre-constructed loss optimization model, iterative optimization is performed on the candidate symbol distance values, and the symbol distance field corresponding to the candidate symbol distance values can be decomposed into orthogonal tri-vector data, greatly reducing the storage requirements while maintaining the high geometric fidelity of the three-dimensional shape data.

[0140] Each module in the above generative three-dimensional shape compression representation device can be implemented in whole or in part by software, hardware, and their combination. The above modules can be embedded in the processor in the computer device in hardware form or be independent of it, or can be stored in the memory in the computer device in software form to facilitate the processor to call and execute the operations corresponding to the above modules.

[0141] In an exemplary embodiment, a computer device is provided. The computer device can be a server, and its internal structure diagram can be as Figure 12 shown. The computer device includes a processor, a memory, an input / output interface (Input / Output, abbreviated as I / O), and a communication interface. Among them, the processor, the memory, and the input / output interface are connected through a system bus, and the communication interface is connected to the system bus through the input / output interface. Among them, the processor of the computer device is used to provide computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system, a computer program, and a database. The internal memory provides an environment for the operation of the operating system and the computer program in the non-volatile storage medium. The database of the computer device is used to store three-dimensional shape data. The input / output interface of the computer device is used to exchange information between the processor and external devices. The communication interface of the computer device is used to communicate with external terminals through a network connection. When the computer program is executed by the processor, it realizes a generative three-dimensional shape compression representation method.

[0142] Those skilled in the art can understand, Figure 12The structure shown is only a block diagram of some of the structures related to the solution of this application, and does not constitute a limitation on the computer device to which the solution of this application is applied. The specific computer device may include more or fewer components than those shown in the figure, or combine some components, or have different component arrangements.

[0143] In an exemplary embodiment, a computer device is provided, including a memory and a processor. A computer program is stored in the memory, and when the processor executes the computer program, the steps in the above method embodiments are implemented.

[0144] In an embodiment, a computer-readable storage medium is provided, on which a computer program is stored, and when the computer program is executed by a processor, the steps in the above method embodiments are implemented.

[0145] In an embodiment, a computer program product is provided, including a computer program, and when the computer program is executed by a processor, the steps in the above method embodiments are implemented.

[0146] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data for analysis, stored data, displayed data, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties, and the collection, use, and processing of relevant data need to comply with relevant regulations.

[0147] Those of ordinary skill in the art can understand that all or part of the processes in the methods of the above embodiments can be completed by instructing relevant hardware through a computer program. The computer program can be stored in a non-volatile computer-readable storage medium. When the computer program is executed, it can include the processes of the embodiments of the above methods. Among them, any reference to a memory, database, or other medium used in the embodiments provided in this application can include at least one of non-volatile memory and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, optical memory, high-density embedded non-volatile memory, resistive random access memory (ReRAM), magnetoresistive random access memory (MRAM), ferroelectric random access memory (FRAM), phase change memory (PCM), graphene memory, etc. Volatile memory can include random access memory (RAM) or external cache memory, etc. By way of illustration and not limitation, RAM can be in various forms, such as static random access memory (SRAM) or dynamic random access memory (DRAM), etc. The databases involved in the embodiments provided in this application can include at least one of relational databases and non-relational databases. Non-relational databases can include distributed databases based on blockchain, etc., without limitation. The processors involved in the embodiments provided in this application can be general-purpose processors, central processing units, graphics processing units, digital signal processors, programmable logic devices, data processing logics based on quantum computing, artificial intelligence (AI) processors, etc., without limitation.

[0148] The technical features of the above embodiments can be combined arbitrarily. For the sake of brevity of description, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, it should be considered as within the scope recorded in this application.

[0149] The above-described embodiments merely represent several implementation manners of the present application. The description thereof is relatively specific and detailed, but it should not be construed as a limitation on the scope of the patent of the present application. It should be noted that for those of ordinary skill in the art, without departing from the concept of the present application, several modifications and improvements can still be made, and these all fall within the protection scope of the present application. Therefore, the protection scope of the present application shall be subject to the appended claims.

Claims

1. A generative three-dimensional shape compression representation method, characterized in that: The method comprises: Acquiring three-dimensional shape data, and normalizing the three-dimensional shape data to obtain a feature space; Randomly sampling the three-dimensional shape data in the feature space to obtain an initial signed distance value; Numerically truncating the initial symbol distance value according to a preset range space to obtain a candidate symbol distance value; Iteratively optimizing the candidate symbol distance values ​​according to a pre-constructed loss optimization model to obtain orthogonal three-vector data; the orthogonal three-vector data is used to compress and represent the three-dimensional shape data; Reshape the orthogonal three-vector data to obtain one-dimensional feature vector data; Iteratively training the one-dimensional feature vector data according to a pre-constructed initial generation model to obtain a candidate generation model; Establishing a training distribution change map based on the training data of the candidate generation model; wherein the training distribution change map is used to characterize the mapping from the standard normal distribution to the training data distribution; Performing loss optimization on the candidate generation model according to the change of the loss value of the distribution change map to obtain a target generation model; wherein the target generation model is used to generate a three-dimensional shape represented by the orthogonal three-vector data; Obtaining random noise to be restored; the random noise to be restored is random noise with a shape consistent with the orthogonal three-vector data; Iteratively restore the random noise to be restored and the preset position code to obtain restored noise; Performing a sampling change process on the restored noise to obtain restored one-dimensional vector data; Reconstructing the restored one-dimensional vector data to obtain signed distance value grid data; The symbol distance value grid is restored through a cube tracing model to obtain triangle mesh data.

2. The method according to claim 1, characterized in that The initial signed distance value includes: a first signed distance value, a second signed distance value and a third signed distance value; the random sampling of the three-dimensional shape data in the feature space to obtain the initial signed distance value includes: Randomly sampling the surface area of ​​the three-dimensional shape data in the feature space to obtain the first signed distance value; Randomly sampling a region near a surface of the three-dimensional shape data in the feature space to obtain the second signed distance value; Random sampling is performed in the feature space to obtain the third symbol distance value.

3. The method according to claim 1, characterized in that: The step of numerically truncating the initial symbol distance value according to a preset range space to obtain a candidate symbol distance value includes: The initial symbol distance value is screened according to the numerical range corresponding to the preset range space to obtain the symbol distance value to be truncated; The to-be-truncated symbol distance value in the initial symbol distance value is numerically truncated to obtain the candidate symbol distance value.

4. The method according to claim 1, characterized in that: The candidate symbol distance values ​​are iteratively optimized according to the pre-constructed loss optimization model to obtain orthogonal three-vector data; The orthogonal three-vector data is used to compress and represent the three-dimensional shape data, including: Randomly sample from a preset point set to obtain sampling point data; Interpolate the candidate symbol distance values ​​according to the coordinate data of the sampling point data to obtain an interpolation result; The interpolation result and the true sign distance value of the sampling point data are subjected to loss optimization according to the loss optimization model to obtain the orthogonal three-vector data.

5. A generative three-dimensional shape compression representation device, characterized in that: The device comprises: An acquisition module, used for acquiring three-dimensional shape data and normalizing the three-dimensional shape data to obtain a feature space; A random sampling module, used for randomly sampling the three-dimensional shape data in the feature space to obtain an initial signed distance value; A numerical truncation module, used for numerically truncating the initial signed distance value according to a preset range space to obtain a candidate signed distance value; A compression characterization module, used for iteratively optimizing the candidate symbol distance value according to a pre-built loss optimization model to obtain orthogonal three-vector data; the orthogonal three-vector data is used for compressing and characterizing the three-dimensional shape data; A generative model training module is used to reshape the orthogonal three-vector data to obtain one-dimensional feature vector data; iteratively train the one-dimensional feature vector data according to a pre-constructed initial generative model to obtain a candidate generative model; establish a training distribution change mapping according to the training data of the candidate generative model; wherein the training distribution change mapping is used to characterize the mapping from the standard normal distribution to the training data distribution; perform loss optimization on the candidate generative model according to the change of the loss value of the distribution change mapping to obtain a target generative model; wherein the target generative model is used to generate a three-dimensional shape based on the representation of the orthogonal three-vector data; A restoration module is used to obtain the random noise to be restored; the random noise to be restored is a random noise with the same shape as the orthogonal three-vector data; the random noise to be restored and the preset position code are iteratively restored to obtain the restored noise; the restored noise is subjected to sampling change processing to obtain the restored one-dimensional vector data; the restored one-dimensional vector data is reconstructed to obtain the signed distance value grid data; the signed distance value grid is restored through a cube tracking model to obtain the triangular grid data.

6. The device according to claim 5, characterized in that The initial signed distance value includes: a first signed distance value, a second signed distance value and a third signed distance value; the random sampling module is also used to randomly sample the surface area of ​​the three-dimensional shape data in the feature space to obtain the first signed distance value; randomly sample the area near the surface of the three-dimensional shape data in the feature space to obtain the second signed distance value; and randomly sample in the feature space to obtain the third signed distance value.

7. The device according to claim 5, characterized in that The numerical truncation module is further used to screen the initial signed distance value according to the numerical range corresponding to the preset range space to obtain the signed distance value to be truncated; The to-be-truncated symbol distance value in the initial symbol distance value is numerically truncated to obtain the candidate symbol distance value.

8. A computer device comprising a memory and a processor, wherein the memory stores a computer program, wherein: When the processor executes the computer program, the steps of the method according to any one of claims 1 to 4 are implemented.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the method according to any one of claims 1 to 4 are implemented.

10. A computer program product, comprising a computer program, characterized in that When the computer program is executed by a processor, the steps of the method according to any one of claims 1 to 4 are implemented.

Images