Three-dimensional model building training method and three-dimensional model building method

Abstract

The embodiment of the invention provides a three-dimensional building model training method and a three-dimensional model building method, and the three-dimensional building model training method comprises the steps: obtaining a first sample set; at least one training sample in the first sample set comprises first geometric texture field data of one first three-dimensional model; wherein the first geometric texture field data is used for reflecting symbol distance function values and texture values of effective sampling points of the first three-dimensional model; the distance between the effective sampling points and the surface of the three-dimensional model is smaller than a preset length; training an auto-encoder model by using the first sample set to obtain a target auto-encoder model; acquiring a second sample set; at least one training sample in the second sample set comprises implicit vector data corresponding to a second three-dimensional model and an image corresponding to the second three-dimensional model; and training an image feature extraction model and an implicit vector generation model in the three-dimensional construction model by using the second sample set to obtain a trained target three-dimensional construction model.

Description

TECHNICAL FIELD

[0001] Embodiments of the present specification relate to the technical field of computer technology, and particularly relate to a three-dimensional construction model training method and a three-dimensional model construction method. BACKGROUND

[0002] In the e-commerce guide, especially in the home decoration and furniture industry, consumers generally face the pain point of "imaginary and reality do not match". The high cost of return and exchange seriously affects the shopping experience and the profit of the merchant. Traditional product pictures cannot intuitively show the display effect of the product in three-dimensional space. For example: furniture pictures cannot intuitively show the size, style matching and layout effect of furniture products in three-dimensional space. Therefore, three-dimensional modeling for products can effectively reduce the decision-making cost of consumers and improve the shopping experience. However, the traditional three-dimensional modeling by artificial way is time-consuming and laborious, and the cost is high, which is difficult to cover a large number of products. Therefore, an effective solution is needed to solve the above problems. SUMMARY

[0003] Therefore, the embodiments of the present specification provide a three-dimensional construction model training method. One or more embodiments of the present specification also relate to a three-dimensional model construction method, a three-dimensional construction model training device, a three-dimensional model construction device, a computing device, a computer-readable storage medium, and a computer program product to solve the technical defects in the prior art.

[0004] According to a first aspect of the embodiments of the present specification, a three-dimensional construction model training method is provided, the three-dimensional construction model is used to convert an image into a three-dimensional model, and the three-dimensional construction model includes an image feature extraction model, a latent vector generation model, a self-encoder model, and an output layer. The self-encoder model includes an encoder and a decoder. The method comprises: obtaining a first sample set; at least one training sample in the first sample set contains first geometric texture field data of a first three-dimensional model; the first geometric texture field data is used to reflect the signed distance function value and the texture value of the effective sampling point of the first three-dimensional model; training the self-encoder model using the first sample set to obtain a target self-encoder model; A second sample set is obtained; at least one training sample in the second sample set contains latent vector data corresponding to a second 3D model and an image corresponding to the second 3D model; the latent vector data corresponding to the second 3D model is obtained by processing the second geometric texture field data of the second 3D model using the encoder of the target autoencoder model; the second geometric texture field data is used to reflect the signed distance function value and texture value of the effective sampling points of the second 3D model; the effective sampling points are sampling points whose distance from the surface of the first 3D model or the surface of the second 3D model is less than a preset length; The image feature extraction model and the latent vector generation model in the 3D construction model are trained using the second sample set to obtain the trained target 3D construction model.

[0005] According to a second aspect of the embodiments of this specification, a method for constructing a three-dimensional model is provided, comprising: Acquire the target image of the target object; The target image is input into the image feature extraction model in the target 3D construction model to obtain the image feature data corresponding to the target object output by the image feature extraction model; wherein, the target 3D construction model is trained by the above method; The image feature data is input into the latent vector generation model in the target 3D construction model to obtain the latent vector data corresponding to the target object output by the latent vector generation model. The latent vector data is input into the decoder of the autoencoder model in the target 3D model to obtain the geometric texture field data corresponding to the target object output by the decoder; the geometric texture field data is used to reflect the signed distance function value and texture value of each point in the unit coordinate system space where the 3D model corresponding to the target object is located; Based on the geometric texture field data corresponding to the target object, a three-dimensional model of the target object is constructed.

[0006] According to a third aspect of the embodiments of this specification, a three-dimensional model training apparatus is provided. The three-dimensional model is used to convert an image into a three-dimensional model. The three-dimensional model includes an image feature extraction model, a latent vector generation model, an autoencoder model, and an output layer. The autoencoder model includes an encoder and a decoder. The apparatus includes: The first acquisition module is configured to acquire a first sample set; at least one training sample in the first sample set contains first geometric texture field data of a first three-dimensional model; the first geometric texture field data is used to reflect the signed distance function value and texture value of the effective sampling points of the first three-dimensional model. The first training module is configured to train the autoencoder model using the first sample set to obtain the target autoencoder model. The second acquisition module is configured to acquire a second sample set; at least one training sample in the second sample set contains latent vector data corresponding to a second three-dimensional model and an image corresponding to the second three-dimensional model; the latent vector data corresponding to the second three-dimensional model is obtained by processing the second geometric texture field data of the second three-dimensional model using the encoder of the target autoencoder model; the second geometric texture field data is used to reflect the signed distance function value and texture value of the effective sampling points of the second three-dimensional model; the effective sampling points are sampling points whose distance from the surface of the first three-dimensional model or the surface of the second three-dimensional model is less than a preset length; The second training module is configured to train the image feature extraction model and the latent vector generation model in the 3D construction model using the second sample set, so as to obtain the trained target 3D construction model.

[0007] According to a fourth aspect of the embodiments of this specification, a three-dimensional model building apparatus is provided, comprising: The acquisition module is configured to acquire the target image of the target object. The first input module is configured to input the target image into the image feature extraction model in the target 3D construction model, and obtain the image feature data corresponding to the target object output by the image feature extraction model; wherein, the target 3D construction model is trained by the above method; The second input module is configured to input the image feature data into the latent vector generation model in the target 3D construction model to obtain the latent vector data corresponding to the target object output by the latent vector generation model. The third input module is configured to input the latent vector data into the decoder of the autoencoder model in the target 3D construction model to obtain the geometric texture field data corresponding to the target object output by the decoder; the geometric texture field data is used to reflect the signed distance function value and texture value of each point in the unit coordinate system space where the 3D model corresponding to the target object is located; The construction module is configured to construct a three-dimensional model of the target object based on the geometric texture field data corresponding to the target object.

[0008] According to a fifth aspect of the embodiments of this specification, a computing device is provided, comprising: Memory and processor; The memory is used to store computer-executable instructions, and the processor is used to execute the computer-executable instructions. When the computer-executable instructions are executed by the processor, they implement the steps of the above-described 3D model training method or 3D model construction method.

[0009] According to a sixth aspect of the embodiments of this specification, a computer-readable storage medium is provided that stores computer-executable instructions, which, when executed by a processor, implement the steps of the above-described three-dimensional model training method or three-dimensional model construction method.

[0010] According to a seventh aspect of the embodiments of this specification, a computer program product is provided, including a computer program or instructions that, when executed by a processor, implement the steps of the above-described three-dimensional model training method or three-dimensional model construction method.

[0011] The 3D model training method provided in this embodiment aims to convert an object's image into a 3D model during the model application stage. It pre-trains a 3D model for converting images into 3D models. During training, a first sample set is acquired, where at least one training sample contains first geometric texture field data of a first 3D model. This first geometric texture field data reflects the signed distance function value and texture value of the effective sampling points of the first 3D model. After acquiring the first sample set, an autoencoder model in the 3D model can be trained using this set. The autoencoder model includes an encoder and a decoder. During training, the encoder of the autoencoder model is trained to convert the geometric texture field data of the 3D model into latent vectors, while the decoder is trained to reconstruct the geometric texture field of the 3D model based on the latent vectors. After training the autoencoder model, a target autoencoder model can be obtained. Then, a second sample set can be acquired. At least one training sample in the second sample set contains latent vector data corresponding to a second 3D model and an image corresponding to that second 3D model. The latent vector data corresponding to the second 3D model is obtained by processing the second geometric texture field data of the second 3D model using the encoder of the target autoencoder model. This second geometric texture field data reflects the signed distance function value and texture value of the effective sampling points of the second 3D model. The effective sampling points are those located less than a preset length from the surface of either the first or second 3D model. After obtaining the second sample set, the image feature extraction model and the latent vector generation model in the 3D model can be trained using this set. During training, the latent vector generation model can be trained to generate the latent vector data corresponding to the 3D model based on the image feature data extracted from the image corresponding to the 3D model by the image feature extraction model. Therefore, after training, in the application phase of the target 3D model, the image of the object is input into the target 3D model. The target 3D model can extract image feature data from the image using an image feature extraction model, and then use a latent vector generation model to generate latent vector data corresponding to the object based on the image feature data. Next, the decoder of the autoencoder model can generate geometric texture field data corresponding to the object based on the object's latent vector data. Finally, the output layer can construct the corresponding 3D model of the object based on the object's geometric texture field data, ultimately realizing the transformation from image to 3D model. The target 3D model can quickly and accurately generate the corresponding 3D model based on the object's image in the application phase, eliminating the need for manual modeling, effectively improving the modeling efficiency of 3D models, and saving modeling costs. Attached Figure Description

[0012] Figure 1This is a schematic diagram illustrating an application scenario of a three-dimensional model construction and training method provided in one embodiment of this specification. Figure 2 This is a flowchart illustrating a three-dimensional model construction and training method provided in one embodiment of this specification; Figure 3 This is a flowchart illustrating a three-dimensional model construction method provided in one embodiment of this specification; Figure 4 This is a flowchart illustrating the processing steps of a three-dimensional model construction method provided in one embodiment of this specification. Figure 5 This is a schematic diagram of the structure of a three-dimensional model building training device provided in one embodiment of this specification; Figure 6 This is a schematic diagram of the structure of a three-dimensional model building device provided in one embodiment of this specification; Figure 7 This is a structural block diagram of a computing device provided in one embodiment of this specification. Detailed Implementation

[0013] Many specific details are set forth in the following description to provide a full understanding of this specification. However, this specification can be implemented in many other ways than those described herein, and those skilled in the art can make similar extensions without departing from the spirit of this specification. Therefore, this specification is not limited to the specific implementations disclosed below.

[0014] The terminology used in one or more embodiments of this specification is for the purpose of describing particular embodiments only and is not intended to be limiting of the one or more embodiments of this specification. The singular forms “a,” “described,” and “the” as used in one or more embodiments of this specification and the appended claims are also intended to include the plural forms unless the context clearly indicates otherwise. It should also be understood that the term “and / or” as used in one or more embodiments of this specification refers to and includes any or all possible combinations of one or more associated listed items.

[0015] It should be understood that although the terms first, second, etc., may be used to describe various information in one or more embodiments of this specification, such information should not be limited to these terms. These terms are only used to distinguish information of the same type from one another. For example, first may also be referred to as second without departing from the scope of one or more embodiments of this specification, and similarly, second may also be referred to as first. Depending on the context, the word "if" as used herein may be interpreted as "when," "when," or "in response to a determination."

[0016] Furthermore, 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 used for analysis, stored data, displayed data, etc.) involved in one or more embodiments of this specification are all information and data authorized by the user or fully authorized by all parties. Moreover, the collection, use and processing of related data must comply with the relevant laws, regulations and standards of the relevant countries and regions, and corresponding operation entry points are provided for users to choose to authorize or refuse.

[0017] The technical solutions provided in this application can employ deep learning models with relatively large parameter scales. However, this large model is merely an example; this application does not limit the number of model parameters supported by the deep learning model used, aiming to meet actual needs. The deep learning models involved in this application can be artificial intelligence-based language models (LM) or multimodal models (MM). First, the terms and concepts used in one or more embodiments of this specification will be explained.

[0018] Neural network models are machine learning architectures inspired by the structure of biological nervous systems. Their core consists of numerous artificial neurons (nodes) connected by layers, typically including an input layer, hidden layers, and an output layer. They compute the output through forward propagation, and then, based on the error between the output and the true label, adjust the weights of the inter-layer connections and neuron biases using the backpropagation algorithm. This continuous optimization of model parameters reduces error, ultimately enabling them to learn patterns from data and perform tasks such as classification, regression, and generation. They are the core foundation of deep learning and are widely used in fields such as computer vision, natural language processing, and speech recognition.

[0019] Unit coordinate space is a standardized 3D space defined in computer graphics, 3D modeling, and other fields. Its core characteristic is that the values ​​of each coordinate axis (usually the X, Y, and Z axes) are all standardized within a uniform range of [-1, 1], and it does not contain actual physical units (such as meters or centimeters). As an "intermediate standard space," unit coordinate space eliminates differences in the original size and scale of different models, facilitating unified calculations and adaptations in 3D rendering, model fusion, texture mapping, and other processes. This ensures efficient collaboration of 3D assets from different sources and is commonly found in model standardization preprocessing and graphics engine pipeline processing.

[0020] UV unwrapping is a crucial step in computer graphics for mapping textures to 3D models. Its core function is to map the geometric information of a 3D model's surface onto the UV coordinate system of a 2D plane, essentially "unwrapping" the 3D model into a corresponding mesh of a 2D texture image. Here, "U" and "V" correspond to the horizontal and vertical axes of the 2D plane, analogous to the X, Y, and Z axes in 3D space. Through UV unwrapping, textures can be drawn in 2D software and accurately mapped back onto the 3D model's surface, ensuring correct texture display on the model. It is widely used in game modeling, film and television special effects, virtual simulation, and other fields.

[0021] A texture coordinate map (UV map) is a two-dimensional mesh mapping obtained after unwrapping the UV coordinates of a 3D model. Essentially, it represents the correspondence between the 3D model's surface and the 2D texture image. The texture coordinate map records the specific coordinates of each vertex of the 3D model in the 2D UV coordinate system, clarifying how each pixel of the 2D texture should be precisely attached to the corresponding surface of the 3D model. The texture coordinate map directly determines the display effect of the texture map on the 3D model; its topology should avoid stretching and overlapping as much as possible to ensure that the final texture appears natural and distortion-free.

[0022] Texture mapping is a core technology in computer graphics for adding detail and realism to 3D models. It involves attaching 2D images (textures) to the surface of a 3D model according to preset mapping rules (such as UV mapping) to simulate visual attributes such as color, texture, bumpiness, and gloss. Texture mapping can significantly improve rendering results without increasing the model's polygon count. Common types include diffuse maps (base colors), normal maps (bump detail), and specular maps (specular effects), and are widely used in game development, film and television special effects, virtual simulation, and other fields.

[0023] The Signed Distance Function (SDF) is a function that defines the shortest distance from any point in space to a specific geometric shape (or object surface). This distance is signed to distinguish between the point's interior and exterior. Specifically, if the signed distance function value of point A is greater than zero, point A is outside the object; if the signed distance function value of point A is equal to zero, point A is on the object's surface; and if the signed distance function value of point A is less than zero, point A is inside the object.

[0024] RGB values ​​are a numerical system used to represent colors. At its core, it combines the intensities of three primary colors—red, green, and blue—to create various colors, operating on an additive color mixing model. The intensity of each primary color is typically represented by an integer from 0 to 255 (or a decimal from 0 to 1). Different combinations of these three values ​​can generate over 16 million colors (256×256×256), for example, pure red is (255,0,0) and pure white is (255,255,255). It is widely used in digital devices such as monitors, cameras, and image editing software, and is one of the core standards for digital color representation.

[0025] The Moving Cubes Algorithm (MCA) is a classic 3D isosurface extraction algorithm. Its core principle is to traverse the cube cells in the 3D volume data, determine the positional relationship between the cube vertices and the target isosurface, and then construct triangular patches within the cube to approximate the isosurface based on a preset vertex state combination pattern, ultimately generating the 3D model surface. Due to its simplicity and high efficiency, this algorithm is widely used in fields such as 3D reconstruction of medical images and visualization of geological data.

[0026] An autoencoder (AE) is an unsupervised deep learning model consisting of two core parts: an encoder and a decoder. The encoder maps high-dimensional input data to a low-dimensional latent feature space (encoding process), obtaining a compressed feature representation (latent vector). The decoder, based on this feature representation, attempts to reconstruct an output that is as similar as possible to the original input (decoding process). The model is trained by minimizing the error between the input and the reconstructed output. Ultimately, it can learn the core features of the data (for dimensionality reduction and feature extraction) and also perform tasks such as data generation by manipulating latent features. It is widely used in scenarios such as anomaly detection, image denoising, and data compression.

[0027] Variational Autoencoder (VAE) is an unsupervised learning algorithm that combines probabilistic generative models with deep learning. Its core is to map the input data to a latent space (latent vector) with probability distribution constraints through an encoder, and then reconstruct the original input from the vectors sampled from the latent space by a decoder. It approximates the posterior distribution through variational inference and minimizes the loss function composed of reconstruction error and KL divergence, which ensures reconstruction accuracy and makes the latent space continuous and smooth. It is often used in scenarios such as data generation, feature extraction, dimensionality reduction and anomaly detection.

[0028] Image feature extraction models are a class of algorithms or deep learning architectures used to automatically extract representative information from image data. Their core function is to transform the original pixel matrix into more compact and discriminative feature vectors or feature maps. These models learn information at different levels, such as edges, textures, shapes, and semantics, eliminating redundant data and retaining key features, thus providing fundamental support for subsequent computer vision tasks such as image classification, object detection, image segmentation, and image retrieval.

[0029] ViT (Vision Transformer) model: A deep learning model applied in computer vision, used to extract image features. Its core principle is to break the dependence of convolutional neural networks on local receptive fields, directly performing global attention calculations on the segmented image patch sequence. The model divides the image into fixed-size patches, adds positional encoding after linear projection, and inputs these patches into an encoder to learn global features. Ultimately, it is used for tasks such as image classification, and its global modeling capabilities outperform traditional CNNs on large-scale datasets.

[0030] Diffusion models are a type of deep learning model based on probabilistic generative modeling. Their core principle is to learn the data distribution by simulating a reversible "noise addition-denoising" process. In the forward process, the diffusion model first gradually adds Gaussian noise to the original data, eventually transforming it into a pure noise distribution. Then, by training a neural network to construct the reverse process, it gradually recovers the original data distribution from the pure noise, ultimately achieving data generation. Due to its high generation quality and strong stability, this model is widely used in image, text, and audio generation tasks.

[0031] DiT (Diffusion Transformer) model: This is a deep learning model that can generate corresponding latent vectors based on image features. By learning the inverse diffusion process, the model recovers a clear image from noisy data. It features efficient capture of data dependencies and strong scalability, and is widely used in tasks such as image and video generation.

[0032] The Flood Fill Algorithm is a computer graphics algorithm for filling regions. Its core principle is to start from a specified initial pixel in an image and gradually expand outwards to neighboring pixels according to predefined color or grayscale similarity rules, filling all matching pixels with the target color. This algorithm is commonly used in image editing for color replacement, connected component labeling, image segmentation, and map generation in games. Common implementation methods include recursive and iterative methods (such as queues and stacks).

[0033] A voxel (or volume pixel) is a basic discretized unit in three-dimensional space, analogous to a pixel in a two-dimensional image. It is the smallest unit that constitutes three-dimensional volume data, possessing fixed spatial dimensions and positional coordinates. It is typically used to represent the internal structure of three-dimensional entities, and each voxel can store attribute information such as grayscale value, color, and density. Voxels are widely used in medical imaging, 3D reconstruction, 3D modeling, game engines, and volumetric visualization.

[0034] Flow Matching is a generative modeling method based on Continuously Normalized Flows (CNFs). Its core is learning a time-dependent velocity field that describes the continuous transformation of data from a simple distribution to a target distribution using ordinary differential equations. The framework first selects a probabilistic path that interpolates between the source and target distributions, then trains a neural network as the velocity field to define the flow transformation that constructs this probabilistic path. Flow Matching avoids the computationally expensive simulation steps of traditional diffusion models and is scalable to various state spaces, including discrete state spaces and Riemannian manifolds, providing an efficient and general method for generative modeling.

[0035] In practical applications, the industry employs various methods to represent 3D models, with the most mainstream being geometric mesh + surface texture + coordinate mapping. This system achieves a good balance between efficiency, performance, and flexibility, thus becoming the industry standard. However, directly learning to generate a textured 3D mesh model from a model is extremely difficult. On one hand, the structure of a 3D mesh is discrete and irregular; the model needs to predict not only the position of each mesh vertex but also the correct connection relationships between them, a challenging combinatorial optimization problem. On the other hand, UV unwrapping requires unfolding the surface of the 3D mesh model onto 2D as seamlessly and without overlap as possible, resulting in an irregular topology in the calculated texture coordinate mapping. This process introduces a discontinuous and difficult-to-optimize step. Therefore, manual 3D modeling is currently commonly used, but it is often time-consuming, labor-intensive, and costly, and it struggles to cover a wide variety of products. Therefore, there is an urgent need for a 3D modeling system that can quickly and accurately generate corresponding 3D models based on object images, improving 3D model building efficiency and reducing costs.

[0036] To address the aforementioned technical problems, this specification provides a method for training a 3D model. One or more embodiments of this specification relate to a 3D model construction method, a 3D model training device, a 3D model construction apparatus, a computing device, a computer-readable storage medium, and a computer program product, which will be described in detail in the following embodiments.

[0037] See Figure 1, Figure 1 This diagram illustrates an application scenario of a three-dimensional model training method according to an embodiment of this specification. Figure 1 As shown, during the model training phase, after the model training platform 101 obtains the first sample set and the second sample set, it can use the first sample set and the second sample set to train the 3D construction model 102. After training, the target 3D construction model 103 can be obtained. During the model usage phase, the image of the target object can be input into the target 3D construction model 103, and the target 3D construction model 103 can output the 3D model of the target object.

[0038] In practical applications, the target object can be a product sold on a shopping website, or other items, without specific limitations. By using a 3D model to construct the target object, the image of the target object is transformed into a 3D model displayed in three-dimensional space, allowing users to intuitively experience the object's shape and its spatial presentation. For example, a user wants to buy a sofa on a shopping website, but the product details page only displays an image of the sofa, preventing the user from visually understanding its arrangement in 3D space. The product details page can include a "View Product 3D Effect" control. When the user triggers this control, a 3D model of the sofa, constructed using the target 3D model, is presented based on the image. This allows the user to intuitively experience the sofa's shape and its spatial arrangement by viewing the 3D model. This not only improves user shopping decision efficiency and satisfaction but also effectively reduces return rates, giving the e-commerce sector a powerful immersive experience and core competitiveness.

[0039] Figure 1 The method described above involves the model training platform acquiring a first and second sample set during the training phase. This allows for the training of a 3D model using these sample sets, resulting in the target 3D model. During the usage phase, the target 3D model can quickly and accurately generate the corresponding 3D model based on the object's image, eliminating the need for manual modeling and effectively improving modeling efficiency and saving costs. For users, the 3D model provides a more comprehensive and intuitive understanding of the target product, enhancing shopping decision-making efficiency and satisfaction, and effectively reducing return rates.

[0040] See Figure 2 , Figure 2A flowchart of a three-dimensional model training method according to an embodiment of this specification is shown, wherein the three-dimensional model is used to convert an image into a three-dimensional model. The three-dimensional model includes an image feature extraction model, a latent vector generation model, an autoencoder model, and an output layer. The autoencoder model includes an encoder and a decoder. The execution entity of this process can be a model training platform, and the process may specifically include the following steps.

[0041] Step S202: Obtain a first sample set; at least one training sample in the first sample set contains first geometric texture field data of a first three-dimensional model; the first geometric texture field data is used to reflect the symbolic distance function value and texture value of the effective sampling points of the first three-dimensional model.

[0042] The 3D model training method provided in this embodiment can be applied to 3D model construction scenarios in any field, such as e-commerce, logistics, games, etc., where 3D models of objects need to be built. The 3D model trained in this embodiment can be deployed to realize the conversion from object images to 3D models of objects in related fields, so that users can intuitively feel the display effect of objects in 3D space, which is conducive to improving the user's immersive experience.

[0043] Specifically, the 3D model refers to a neural network model that can transform an input image of an object into a corresponding 3D model of the object. It can be any neural network model deployed in scenarios such as e-commerce, logistics, and gaming; this embodiment does not impose any limitations. The 3D model can include an image feature extraction model, a latent vector generation model, an autoencoder model, and an output layer. Accordingly, the image feature extraction model in the 3D model can process the target image of the input object and output the image feature data corresponding to the target object; the latent vector generation model can process the image feature data corresponding to the target object output by the image feature extraction model and output the latent vector data corresponding to the target object; the autoencoder model in the 3D model includes an encoder and a decoder. During the model's usage phase, the decoder of the autoencoder model can process the latent vector data corresponding to the target object output by the latent vector generation model and output the geometric texture field data corresponding to the target object; the output layer of the 3D model can process the geometric texture field data corresponding to the target object output by the decoder of the autoencoder model to obtain the 3D model corresponding to the target object.

[0044] Based on this, in order to realize the transformation of an object's image into a 3D model of the object during the model application stage, a 3D construction model for transforming images into 3D models can be pre-trained. During model training, a first sample set can be obtained first, wherein at least one training sample in the first sample set contains the first geometric texture field data of a first 3D model. This first geometric texture field data is used to reflect the signed distance function value and texture value of the effective sampling points of the first 3D model. The effective sampling points are sampling points whose distance from the surface of the 3D model is less than a preset length. After obtaining the first sample set, the autoencoder model in the 3D construction model can be trained using the first sample set. The autoencoder model includes an encoder and a decoder. During training, the encoder of the autoencoder model can be trained to convert the geometric texture field data of the 3D model into latent vectors, and the decoder of the autoencoder model can be trained to reconstruct the geometric texture field of the 3D model based on the latent vectors. After training the autoencoder model, a target autoencoder model can be obtained. Then, a second sample set can be acquired. At least one training sample in the second sample set contains latent vector data corresponding to a second 3D model and an image corresponding to that model. The latent vector data corresponding to the second 3D model is obtained by processing the second geometric texture field data of the second 3D model using the encoder of the target autoencoder model. This second geometric texture field data reflects the signed distance function value and texture value of the effective sampling points of the second 3D model. After obtaining the second sample set, the image feature extraction model and latent vector generation model in the 3D model can be trained using this set. During training, the latent vector generation model can be trained to generate the latent vector data corresponding to the 3D model based on the image feature data extracted from the image corresponding to the 3D model by the image feature extraction model. Therefore, after training, in the application phase of the target 3D model, the image of the object is input into the target 3D model. The target 3D model can extract image feature data from the image using an image feature extraction model, and then use a latent vector generation model to generate latent vector data corresponding to the object based on the image feature data. Next, the decoder of the autoencoder model can generate geometric texture field data corresponding to the object based on the object's latent vector data. Finally, the output layer can construct the corresponding 3D model of the object based on the object's geometric texture field data, ultimately realizing the transformation from image to 3D model. The target 3D model can quickly and accurately generate the corresponding 3D model based on the object's image in the application phase, eliminating the need for manual modeling, effectively improving the modeling efficiency of 3D models, and saving modeling costs.

[0045] In practical applications, the first sample set may contain multiple training samples. This embodiment does not specifically limit the number of training samples in the first sample set. Each training sample in the first sample set may contain the first geometric texture field data of a first three-dimensional model. The first geometric texture field data reflects the signed distance function value and texture value of the effective sampling points of the first three-dimensional model. Effective sampling points are sampling points located less than a preset length from the surface of the three-dimensional model. Specifically, an effective sampling point of a particular first three-dimensional model in the first sample set refers to a sampling point located less than a preset length from the surface of that first three-dimensional model. In fact, an effective sampling point is actually a sampling point close to the surface of the three-dimensional model. Therefore, the preset length is usually a small value set in advance and can be set and adjusted according to the mesh division needs of the three-dimensional space and actual requirements. This embodiment does not specifically limit the value of the preset length.

[0046] Furthermore, in order to determine each training sample in the first sample set, it is necessary to determine the 3D coordinates of each effective sampling point corresponding to the 3D model based on the model data of the 3D model in each training sample. Then, based on the 3D coordinates of each effective sampling point, the signed distance function value and texture value of each effective sampling point are determined, thereby obtaining the first geometric texture field data of the 3D model contained in a training sample. In this embodiment, obtaining the first sample set includes: For any training sample in the first sample set, obtain the model data of the target 3D model; based on the model data of the target 3D model, determine the 3D coordinates of each effective sampling point of the target 3D model in the unit coordinate system space; based on the 3D coordinates of each effective sampling point, determine the signed distance function value and the texture value of each effective sampling point; based on the signed distance function value and the texture value of each effective sampling point, construct any training sample in the first sample set corresponding to the target 3D model.

[0047] Specifically, each training sample in the first sample set can correspond to a 3D model. Multiple training samples can be constructed using the model data of multiple 3D models to obtain the first sample set. Therefore, for any training sample in the first sample set, the model data of the target 3D model in that training sample can be obtained first. Then, based on the model data of the target 3D model, the 3D coordinates of each effective sampling point of the target 3D model in the unit coordinate system space can be determined. After determining the 3D coordinates of each effective sampling point, the signed distance function value and texture value of each effective sampling point can be further determined based on these coordinates. Thus, the signed distance function value and texture value of each effective sampling point can be used as the first geometric texture field data of the 3D model contained in that training sample, obtaining the training sample. Multiple training samples can be determined in this way to obtain the first sample set.

[0048] The signed distance function value of each effective sampling point near the surface of the target 3D model reflects the changes in the geometric field near the surface, while the texture value of each effective sampling point reflects the changes in the texture field near the surface. The texture value can be an RGB value or other values ​​used to reflect color; no specific limitation is made. This embodiment innovatively introduces a texture field, a function defined over the entire 3D space. The texture field maps the coordinates (x, y, z) of any 3D coordinate point P to a texture value, transforming the texture from a sparse, surface-limited definition into a dense, volumetric definition distributed throughout the entire 3D space. This allows for the use of more complete 3D supervision signals to train the model to learn the texture distribution. In this embodiment, the geometric field and texture field are coupled, enabling unified learning of geometry and texture in the same 3D domain. This utilizes complete 3D supervision signals to improve the model's learning performance, avoiding the problems of "difficulty in aligning geometric details with texture mapping during generation," "surface distortion," and "texture misalignment" caused by traditional methods that treat geometry and texture as independent branches.

[0049] Furthermore, in order to determine the 3D coordinates of each effective sampling point of the target 3D model in the unit coordinate system space, the target 3D model needs to be normalized to the unit coordinate system space before the 3D coordinates of each effective sampling point near the surface of the target 3D model can be determined in the unit coordinate system. In this embodiment, determining the 3D coordinates of each effective sampling point of the target 3D model in the unit coordinate system space based on the model data of the target 3D model includes: Based on the model data of the target 3D model, the target 3D model is normalized to a unit coordinate system space; using the preset length as a reference, the unit coordinate system space is voxelized to obtain a first number of voxels; the voxels are cubes; the side length of the voxels is equal to the preset length; the first number of voxels are intersected with the surface of the target 3D model to obtain a second number of intersecting voxels; the vertices of each intersecting voxel are determined as the effective sampling points to obtain the 3D coordinates of each effective sampling point of the target 3D model in the unit coordinate system space.

[0050] The specific steps for normalizing the target 3D model to a unit coordinate system space based on the model data of the target 3D model may include: 1. Calculating the bounding box of the target 3D model based on the model data of the target 3D model. 2. Calculating the coordinates of the center point of the bounding box, and then translating the target 3D model so that the center point of the bounding box coincides with the origin of the unit coordinate system. 3. Scaling the target 3D model to the unit coordinate system space. Alternatively, other normalization methods may be used, and this embodiment does not specifically limit them.

[0051] In practical applications, when performing voxelization on a unit coordinate system space, a preset length can be set according to the segmentation requirements. Using this preset length as a reference, the unit coordinate system space is voxelized to obtain a first number of voxels. The resulting voxels are cubes with side lengths equal to the preset length. For example, if the X, Y, and Z axes of the unit coordinate system space range from -1 to 1, and the space is divided into 1024 equal parts along the X-axis, Y-axis, and Z-axis, 1024 × 1024 × 1024 voxels can be obtained, each voxel being a cube with a side length of 1 / 512. Here, 1 / 512 is the preset length, which can be calculated based on the ranges of the X, Y, and Z axes and the number of equal parts required along each axis: 2 / 1024 = 1 / 512. The above are just examples. Actual segmentation requirements can be adjusted according to the model's refinement needs. For example, if the model's refinement requirements are high, the number of segments along each axis can be increased, which will increase the number of voxels and result in a more refined 3D model, but will also require more computing resources and computation time. If the model's refinement requirements are low, the number of segments along each axis can be appropriately reduced to decrease the number of voxels, thereby saving computing resources and reducing computation time.

[0052] Specifically, the first number of voxels are intersected with the surface of the target 3D model. An intersection detection algorithm can be used to perform an "intersection" check on each voxel. If there is no intersection, the voxel is skipped. If one or more triangular facets of the target 3D model's surface intersect with a voxel, that voxel is marked. This process continues until all voxels have been traversed. The vertices of all voxels intersecting with the target 3D model's surface are determined as valid sampling points, thus obtaining the 3D coordinates of each valid sampling point of the target 3D model in the unit coordinate system space. The intersection detection algorithm can include, but is not limited to, the AABB-triangular facet intersection detection algorithm, ray casting, etc.

[0053] In summary, by identifying the effective sampling points near the surface of the target 3D model, the first geometric texture field data of the 3D model does not need to include the signed distance function value and texture value of all sampling points (vertices of each voxel) in the unit coordinate system space during the construction of the training samples in the first sample set. Only the signed distance function value and texture value of each effective sampling point need to be calculated. This can effectively solve the problem of computational redundancy in the spatial domain of traditional volume methods, significantly reduce the amount of invalid computation, and improve computational efficiency.

[0054] Furthermore, the sign distance function value of the effective sampling point can be determined in two steps: the sign and the value. The value corresponds to the projection distance from the effective sampling point to the surface of the 3D model. The texture value of the effective sampling point is defined as being consistent with the texture value of the projection point of the effective sampling point onto the surface of the target 3D model. In this embodiment, determining the sign distance function value and the texture value of each effective sampling point based on their 3D coordinates includes: Based on the three-dimensional coordinates of each valid sampling point, the projection distance from each valid sampling point to the surface of the target three-dimensional model and the centroid coordinates of the projection points of each valid sampling point on the surface of the target three-dimensional model are determined; based on the projection distance from each valid sampling point to the surface of the target three-dimensional model, the signed distance function value of each valid sampling point is determined; based on the centroid coordinates of the projection points of each valid sampling point on the surface of the target three-dimensional model, the texture value of each valid sampling point is determined.

[0055] Specifically, determining the projection distance from each valid sampling point to the surface of the target 3D model and the centroid coordinates of the projection points of each valid sampling point on the surface of the target 3D model based on the 3D coordinates of each valid sampling point may include: For any valid sampling point, the bounding volume hierarchy method is used to determine several candidate triangular meshes on the surface of the target 3D model that are close to the valid sampling point; based on the 3D coordinates of the valid sampling point and the 3D coordinates of the vertices of each candidate triangular mesh, the projection distance from the valid sampling point to each candidate triangular mesh is calculated; based on the projection distance, the target triangular mesh is determined from the several candidate triangular meshes; the projection distance from the valid sampling point to the target triangular mesh is determined as the projection distance from the valid sampling point to the surface of the target 3D model; based on the 3D coordinates of the valid sampling point and the 3D coordinates of the vertices of the target triangular mesh, the centroid coordinates of the projection point of the valid sampling point on the surface of the target 3D model are determined.

[0056] Bounding Volume Hierarchy (BVH) is a spatial partitioning technique used to accelerate collision detection or ray tracing in 3D space. Its core idea is to construct a multi-level bounding volume structure for complex objects or scenes: the top level uses a large bounding volume (such as an axis-aligned bounding box AABB, a sphere, etc.) to wrap the entire object, and the next level further divides the bounding volume into multiple sub-bounding volumes, each wrapping a different part of the object, and so on to form a tree-like hierarchy. During detection, the top level is used to quickly determine whether the high-level bounding volumes of two objects intersect. If they do not intersect, the branch is directly excluded. If they intersect, the next level is used to detect the sub-bounding volumes. By filtering layer by layer, the number of geometric elements that need to be accurately calculated is greatly reduced, thereby improving detection efficiency.

[0057] Based on this, by utilizing the bounding volume hierarchy method, one or more candidate triangular meshes that are close to a certain effective sampling point on the surface of the target 3D model can be identified. If there are multiple candidate triangular meshes, a geometric calculation function (e.g., the `point_to_triangle_distance` function) can be called to calculate the shortest distance from a single point in 3D space to the surface of a single triangle. Based on the 3D coordinates of the effective sampling point and the coordinates of the three vertices of each candidate triangular mesh, the projected distance from the effective sampling point to each candidate triangular mesh is calculated. The candidate triangular mesh with the smallest projected distance to the effective sampling point is then identified as the target triangular mesh, and the projected distance from the effective sampling point to the target triangular mesh is determined as the projected distance from the effective sampling point to the surface of the target 3D model. Furthermore, the barycentric coordinates of the projection point of the effective sampling point on the surface of the target 3D model can be determined using a geometric calculation function based on the 3D coordinates of the effective sampling point and the vertices of the target triangular mesh.

[0058] Furthermore, for a valid sampling point, after determining the projected distance from the valid sampling point to the surface of the target 3D model, this projected distance can be determined as the numerical part of the signed distance function value of the valid sampling point. Then, by determining the sign of the signed distance function value using other methods, the signed distance function value of the valid sampling point can be obtained. In this embodiment, determining the signed distance function value of each valid sampling point based on the projected distance from each valid sampling point to the surface of the target 3D model includes: For any valid sampling point, the sign of the sign distance function of the valid sampling point is determined using a flooding fill algorithm; the projection distance from the valid sampling point to the surface of the target 3D model is determined as the value of the sign distance function of the valid sampling point; the sign of the sign distance function of the valid sampling point and the value of the sign distance function of the valid sampling point are combined to obtain the sign distance function value of the valid sampling point.

[0059] For any valid sampling point, the flooding algorithm can determine whether the valid sampling point is inside, outside, or on the surface of the target 3D model. If the valid sampling point is inside the target 3D model, its signed distance function value is negative; if it is outside the target 3D model, its signed distance function value is positive; and if it is on the surface of the target 3D model, its signed distance function value is 0.

[0060] In practical applications, after determining the sign of the signed distance function for a valid sampling point using the flooding fill algorithm, the projected distance from that valid sampling point to the surface of the target 3D model can be defined as the value of the signed distance function for that valid sampling point. The sign of the signed distance function is then combined with its numerical value to obtain the final signed distance function value for that valid sampling point. For example, if the projected distance from valid sampling point A to the surface of the target 3D model is calculated to be 0.001 using geometric calculation functions, and the flooding fill algorithm determines that valid sampling point A is inside the target 3D model, then the signed distance function value for valid sampling point A is -0.001.

[0061] Furthermore, the model data of the target 3D model may also include a texture map. After determining the barycentric coordinates of the projection points of the effective sampling points on the surface of the target 3D model, the texture coordinates of the projection points can be determined based on the barycentric coordinates of the projection points. Then, based on the texture map, the texture value of the projection points is further determined, and the texture value of the projection points is then determined as the texture value of the effective sampling point. In this embodiment, the model data of the target 3D model includes a texture map; determining the texture value of each effective sampling point based on the barycentric coordinates of the projection points of each effective sampling point on the surface of the target 3D model includes: For any valid sampling point, obtain the texture coordinates of each vertex of the target triangular mesh corresponding to the valid sampling point; determine the texture coordinates of the projection point based on the centroid coordinates of the projection point of the valid sampling point on the surface of the target 3D model and the texture coordinates of each vertex of the target triangular mesh; determine the texture value of the projection point using the texture map and the texture coordinates of the projection point; and determine the texture value of the projection point as the texture value of the valid sampling point.

[0062] In practical applications, the texture coordinates of each vertex of the target triangular mesh corresponding to the valid sampling points can be obtained from the model data of the target 3D model. After determining the texture coordinates of each vertex of the target triangular mesh, the texture coordinates of the projection points can be determined by interpolation based on the centroid coordinates of the projection points. Furthermore, the texture values ​​of the projection points can be obtained by reading the corresponding texture values ​​from the texture map based on the texture coordinates of the projection points.

[0063] Step S204: Train the autoencoder model using the first sample set to obtain the target autoencoder model.

[0064] The autoencoder model consists of two parts: an encoder and a decoder. The encoder maps high-dimensional input data to a low-dimensional latent feature space, obtaining a compressed latent vector. The decoder, based on this latent vector, attempts to reconstruct an output that is as similar as possible to the original input. In this embodiment, the autoencoder model may include, but is not limited to, variational autoencoder models and adversarial autoencoder models.

[0065] In actual training, after inputting a training sample from the first sample set into the autoencoder model, the encoder of the autoencoder model can convert the geometric texture field data of the 3D model in the training sample into latent vectors. The decoder of the autoencoder model can reconstruct the geometric texture field of the 3D model based on the latent vectors output by the encoder. The training objective is to minimize the difference between the geometric texture field of the 3D model output by the encoder and the geometric texture field of the 3D model in the training sample. Training the autoencoder model can train the encoder of the autoencoder model to convert the geometric texture field data of the 3D model into latent vectors, and at the same time train the decoder of the autoencoder model to reconstruct the geometric texture field of the 3D model based on the latent vectors.

[0066] Based on this, in this embodiment, training the autoencoder model using the first sample set to obtain the target autoencoder model includes: For any training sample in the first sample set, the first geometric texture field data of a first 3D model contained in the training sample is input into the encoder of the autoencoder model to obtain the latent vector data corresponding to the first 3D model output by the encoder; the latent vector data is input into the decoder of the autoencoder model to obtain the third geometric texture field data of the first 3D model reconstructed by the decoder; the third geometric texture field data is used to reflect the symbolic distance function value and texture value of each point in the unit coordinate system space where the first 3D model reconstructed by the decoder is located; based on the third geometric texture field data of the first 3D model, the symbolic distance function value and texture value of each effective sampling point of the first 3D model are determined to obtain the fourth geometric texture field data; the fourth geometric texture field data is used to reflect the symbolic distance function value and texture value of each effective sampling point of the first 3D model reconstructed by the decoder; with the goal of reducing the difference between the first geometric texture field data and the fourth geometric texture field data, the model parameters of the autoencoder model are adjusted to obtain the trained target autoencoder model.

[0067] It should be noted that the third geometric texture field data of the first 3D model reconstructed by the decoder of the autoencoder model reflects the signed distance function value and texture value of each point in the unit coordinate system space where the first 3D model is located. Specifically, each point can refer to the vertex of each voxel in the unit coordinate system space. In contrast, the first geometric texture field data in the corresponding training samples reflects the signed distance function value and texture value of each effective sampling point in the unit coordinate system space where the first 3D model is located. Therefore, the number of points included in the first geometric texture field data differs from the number of points included in the third geometric texture field data, making it impossible to directly calculate the difference between the two. Therefore, based on the third geometric texture field data of the first 3D model, the signed distance function value and texture value of each effective sampling point of the first 3D model are first determined to obtain the fourth geometric texture field data. The fourth geometric texture field data is used to reflect the signed distance function value and texture value of each effective sampling point of the first 3D model reconstructed by the decoder. By adjusting the model parameters of the autoencoder model with the goal of reducing the difference between the first geometric texture field data and the fourth geometric texture field data, the encoder of the autoencoder model can be trained to convert the geometric texture field data of the three-dimensional model into latent vectors, and the decoder of the autoencoder model can be trained to accurately reconstruct the geometric texture field of the three-dimensional model based on the latent vectors.

[0068] Step S206: Obtain a second sample set; at least one training sample in the second sample set contains latent vector data corresponding to a second three-dimensional model and an image corresponding to the second three-dimensional model; the latent vector data corresponding to the second three-dimensional model is obtained by processing the second geometric texture field data of the second three-dimensional model using the encoder of the target autoencoder model; the second geometric texture field data is used to reflect the symbolic distance function value and texture value of the effective sampling points of the second three-dimensional model; the effective sampling points are sampling points that are less than a preset length from the surface of the first three-dimensional model or the surface of the second three-dimensional model.

[0069] In practical applications, the second sample set may contain multiple training samples. This embodiment does not specifically limit the number of training samples in the second sample set. Each training sample in the second sample set may contain latent vector data corresponding to a second 3D model and an image corresponding to that 3D model. The image corresponding to the 3D model may refer to an image of the object corresponding to that 3D model, and there may be one or more images; this is not specifically limited. The second 3D model in the second sample set may be the same 3D model as the first 3D model in the first sample set, or it may be a different 3D model; this is not specifically limited.

[0070] Specifically, when constructing training samples in the second sample set, for a training sample, the encoder of the trained target autoencoder model can first be used to process the second geometric texture field data of a second three-dimensional model to obtain the latent vector data corresponding to the second three-dimensional model. Then, the latent vector data corresponding to the second three-dimensional model and the image corresponding to the second three-dimensional model are combined to obtain a training sample.

[0071] In the embodiments of this specification, a valid sampling point can refer to a sampling point that is less than a preset length from the surface of the 3D model. Specifically, a valid sampling point of a first 3D model in the first sample set refers to a sampling point that is less than a preset length from the surface of the first 3D model; a valid sampling point of a second 3D model in the second sample set refers to a sampling point that is less than a preset length from the surface of the second 3D model. For a target 3D model, any sampling point that is less than a preset length from the surface of the target 3D model can be used as a valid sampling point of the target 3D model.

[0072] Step S208: Use the second sample set to train the image feature extraction model and the latent vector generation model in the 3D construction model to obtain the trained target 3D construction model.

[0073] The image feature extraction model in the 3D construction model is used to extract feature data from the input image of the target object. Specifically, the ViT (Vision Transformer) model can be used as the image feature extraction model, or other models with image feature extraction capabilities can be used; there are no specific limitations.

[0074] In the 3D construction model, the latent vector generation model is used to generate latent vector data corresponding to the target object based on the image feature data of the target object output by the image feature extraction model. Specifically, the DiT (DiffusionTransformer) model can be used as the latent vector generation model, or other diffusion models with corresponding functions can also be used; there is no specific limitation on this.

[0075] During the model training process, the image feature extraction model does not need to be trained; therefore, the parameters of the image feature extraction model can be frozen during training. Based on this, in this embodiment, training the image feature extraction model and the latent vector generation model in the 3D construction model using the second sample set to obtain the trained target 3D construction model includes: For any training sample in the second sample set, the image feature extraction model with frozen image input parameters corresponding to the second 3D model contained in the training sample is used to obtain image feature data output by the image feature extraction model; the image feature data is input into the latent vector generation model to obtain latent vector data output by the latent vector generation model; the model parameters of the latent vector generation model are adjusted with the goal of reducing the difference between the latent vector data output by the latent vector generation model and the latent vector data corresponding to the second 3D model contained in the training sample, to obtain the trained target 3D construction model.

[0076] Specifically, the second sample set is mainly used to train the latent vector generation model. During training, the latent vector generation model can be trained to generate latent vector data corresponding to the 3D model based on image feature data extracted from the image corresponding to the 3D model by the image feature extraction model. The training objective is to make the latent vector data output by the latent vector generation model as close as possible to the latent vector data corresponding to the 3D model contained in the training samples. Therefore, the model parameters of the latent vector generation model can be adjusted with the goal of reducing the difference between the latent vector data output by the latent vector generation model and the latent vector data corresponding to the 3D model contained in the training samples. After training the autoencoder model and the latent vector generation model in the 3D construction model, the trained target 3D construction model can be obtained. The target 3D construction model is used in the model usage stage.

[0077] Specifically, the latent vector generation model can be trained using the flow matching framework, or other frameworks can be used; no specific limitation is made. Compared to traditional diffusion models, flow matching improves training stability through deterministic path planning. The process of training the model using the flow matching framework is a well-known existing technique, therefore, it will not be described in detail here.

[0078] In the actual use of the 3D model, the decoder of the autoencoder model in the target 3D model outputs the geometric texture field data corresponding to the target object, not the 3D model corresponding to the target object. Therefore, it is necessary to construct the 3D model corresponding to the target object based on the geometric texture field data. Based on this, in this embodiment, the output layer of the target 3D model is used to construct the target 3D model based on the geometric texture field data of the target 3D model reconstructed by the decoder. The geometric texture field data reflects the signed distance function value and texture value of each point in the unit coordinate system space where the target 3D model reconstructed by the decoder resides.

[0079] See Figure 2The schematic diagram illustrates the 3D model training method provided in this embodiment. To enable the conversion of an object's image into a 3D model during the model application stage, a 3D model for converting images into 3D models can be pre-trained. During model training, a first sample set can be acquired. At least one training sample in the first sample set contains first geometric texture field data of a first 3D model. This first geometric texture field data reflects the signed distance function value and texture value of the effective sampling points of the first 3D model. Effective sampling points are those located less than a preset distance from the surface of the 3D model. After acquiring the first sample set, an autoencoder model in the 3D model can be trained using this set. The autoencoder model includes an encoder and a decoder. During training, the encoder of the autoencoder model can be trained to convert the geometric texture field data of the 3D model into latent vectors, while the decoder of the autoencoder model can be trained to reconstruct the geometric texture field of the 3D model based on the latent vectors. After training the autoencoder model, a target autoencoder model can be obtained. Then, a second sample set can be acquired. At least one training sample in the second sample set contains latent vector data corresponding to a second 3D model and an image corresponding to that model. The latent vector data corresponding to the second 3D model is obtained by processing the second geometric texture field data of the second 3D model using the encoder of the target autoencoder model. This second geometric texture field data reflects the signed distance function value and texture value of the effective sampling points of the second 3D model. After obtaining the second sample set, the image feature extraction model and latent vector generation model in the 3D model can be trained using this set. During training, the latent vector generation model can be trained to generate the latent vector data corresponding to the 3D model based on the image feature data extracted from the image corresponding to the 3D model by the image feature extraction model. Therefore, after training, in the application phase of the target 3D model, the image of the object is input into the target 3D model. The target 3D model can extract image feature data from the image using an image feature extraction model, and then use a latent vector generation model to generate latent vector data corresponding to the object based on the image feature data. Next, the decoder of the autoencoder model can generate geometric texture field data corresponding to the object based on the object's latent vector data. Finally, the output layer can construct the corresponding 3D model of the object based on the object's geometric texture field data, ultimately realizing the transformation from image to 3D model. The target 3D model can quickly and accurately generate the corresponding 3D model based on the object's image in the application phase, eliminating the need for manual modeling, effectively improving the modeling efficiency of 3D models, and saving modeling costs.

[0080] See Figure 3 , Figure 3A flowchart of a three-dimensional model construction method according to an embodiment of this specification is shown, which may specifically include the following steps.

[0081] Step S302: Obtain the target image of the target object.

[0082] Step S304: Input the target image into the image feature extraction model in the target 3D construction model to obtain the image feature data corresponding to the target object output by the image feature extraction model; wherein, the target 3D construction model is trained by the above method.

[0083] Step S306: Input the image feature data into the latent vector generation model in the target 3D construction model to obtain the latent vector data corresponding to the target object output by the latent vector generation model.

[0084] Step S308: Input the latent vector data into the decoder of the autoencoder model in the target 3D construction model to obtain the geometric texture field data corresponding to the target object output by the decoder; the geometric texture field data is used to reflect the signed distance function value and texture value of each point in the unit coordinate system space where the 3D model corresponding to the target object is located.

[0085] Step S310: Based on the geometric texture field data corresponding to the target object, construct a three-dimensional model corresponding to the target object.

[0086] The three-dimensional model construction method provided in this embodiment can be referred to in the same or corresponding descriptions in the above embodiments for any content not described in detail. This embodiment will not elaborate further here.

[0087] Specifically, in this embodiment, constructing a three-dimensional model of the target object based on the geometric texture field data corresponding to the target object includes: Using the moving cube algorithm, based on the signed distance function values ​​of each point in the unit coordinate system space where the 3D model corresponding to the target object is located, the vertex coordinates of each triangular mesh on the surface of the 3D model corresponding to the target object are determined; based on the texture values ​​of each point in the unit coordinate system space where the 3D model corresponding to the target object is located, the texture values ​​of the vertices of each triangular mesh are determined; based on the vertex coordinates of each triangular mesh on the surface of the 3D model and the texture values ​​of the vertices of each triangular mesh, the 3D model corresponding to the target object is constructed.

[0088] In practical applications, the Moving Cube algorithm is used to determine the vertex coordinates of each triangular mesh on the surface of the 3D model corresponding to the target object, based on the signed distance function values ​​of each point in the unit coordinate system space of the 3D model corresponding to the target object. The fundamental principle is to determine whether a voxel intersects with the surface of the 3D model based on the signed distance function values ​​of each point (vertices of each voxel) in the unit coordinate system space of the 3D model corresponding to the target object. If they intersect, the precise coordinates of the intersection point are calculated by linear interpolation based on the sign change of the SDF values ​​of adjacent vertices, generating a local triangular mesh set. This process is repeated until all voxels are traversed, thus obtaining the vertex coordinates of each triangular mesh constituting the 3D model.

[0089] Furthermore, the texture values ​​of each vertex of a triangular mesh can be determined based on the texture values ​​of each point in the unit coordinate system space where the 3D model corresponding to the target object is located. Specifically, the texture value of the vertex closest to a vertex of a triangular mesh among the vertices of each voxel in the unit coordinate system space can be determined as the texture value of that vertex of the triangular mesh. Alternatively, the texture values ​​of the two vertices (voxel vertices) adjacent to that vertex of the triangular mesh can be determined, and the texture value of that vertex of the triangular mesh can be obtained through linear interpolation.

[0090] Finally, after determining the vertex coordinates and texture values ​​of each triangular mesh that makes up the surface of the 3D model, the 3D model corresponding to the target object can be constructed.

[0091] In practical applications, the target object may include the target product. After constructing a 3D product model corresponding to the target product, the display effect of the target product in virtual space can be presented to the user. Based on this, in this embodiment, after constructing the 3D model corresponding to the target object based on the geometric texture field data corresponding to the target object, the method further includes: A three-dimensional product model corresponding to the target product is displayed to the user in a virtual space to present the display effect of the target product in the virtual space.

[0092] In summary, by utilizing a target 3D model to transform the image of a target product into a 3D product model displayed in a 3D virtual space, users can intuitively experience the form of the target product and its display effect in virtual space. For users, the 3D model of the target product allows for a more comprehensive and intuitive understanding, which helps improve the efficiency of user shopping decisions and shopping satisfaction, and can also effectively reduce return rates.

[0093] See Figure 3The flowchart shown illustrates the 3D model construction method provided in this embodiment. The method involves inputting an image of the target object into a target 3D model. The target 3D model utilizes an image feature extraction model to extract image feature data from the image. Then, a latent vector generation model generates latent vector data corresponding to the target object based on the image feature data. Next, an autoencoder model's decoder generates geometric texture field data corresponding to the target object based on the latent vector data. Finally, the output layer constructs the 3D model of the target object based on the geometric texture field data, thus achieving the conversion from image to 3D model. This target 3D model can quickly and accurately generate the corresponding 3D model from the object's image during the application stage, eliminating the need for manual modeling, effectively improving the modeling efficiency of 3D models, and saving modeling costs.

[0094] The following is in conjunction with the appendix Figure 4 Taking the application of the 3D model construction method provided in this specification in a product display scenario as an example, the 3D model construction will be further explained. Figure 4 The present specification shows a flowchart of a three-dimensional model construction method according to an embodiment, which includes the following steps.

[0095] Step S402: Obtain the target image of the target product.

[0096] Step S404: Input the target image into the image feature extraction model in the target 3D construction model to obtain the image feature data corresponding to the target product output by the image feature extraction model; wherein, the target 3D construction model is trained by using the above-mentioned 3D construction model training method.

[0097] Step S406: Input the image feature data into the latent vector generation model in the target 3D construction model to obtain the latent vector data corresponding to the target product output by the latent vector generation model.

[0098] Step S408: Input the latent vector data into the decoder of the autoencoder model in the target 3D construction model to obtain the geometric texture field data corresponding to the target product output by the decoder of the autoencoder model; wherein, the geometric texture field data corresponding to the target product is used to reflect the signed distance function value and texture value of each point in the unit coordinate system space where the 3D product model corresponding to the target product is located.

[0099] Step S410: Using the moving cube algorithm, based on the signed distance function value of each point in the unit coordinate system space where the 3D product model corresponding to the target product is located, determine the vertex coordinates of each triangular mesh on the surface of the 3D product model corresponding to the target product.

[0100] Step S412: Based on the texture values ​​of each point in the unit coordinate system space where the 3D product model corresponding to the target product is located, determine the texture values ​​of the vertices of each triangular mesh on the surface of the 3D product model.

[0101] Step S414: Based on the vertex coordinates of each triangular mesh on the surface of the 3D product model and the texture values ​​of the vertices of each triangular mesh, construct the 3D product model corresponding to the target product.

[0102] Step S416: Display the 3D product model corresponding to the target product to the user in the virtual space to present the display effect of the target product in the virtual space.

[0103] In summary, the 3D model construction method provided in this embodiment involves inputting an image of the target product into a target 3D model. The target 3D model can extract image feature data from the image using an image feature extraction model, then generate latent vector data corresponding to the target product using a latent vector generation model based on the image feature data. Subsequently, the decoder of an autoencoder model can generate geometric texture field data corresponding to the target product based on the latent vector data. Finally, the output layer can construct a 3D product model corresponding to the target product based on the geometric texture field data, ultimately realizing the transformation from a product image to a 3D product model. The target 3D model can quickly and accurately generate a corresponding 3D product model based on the product image during the application stage, eliminating the need for manual modeling, effectively improving the modeling efficiency of 3D models, and saving modeling costs. Furthermore, for users, the 3D product model allows for a more comprehensive and intuitive understanding of the target product, which helps improve user shopping decision-making efficiency and shopping satisfaction, and can also effectively reduce return rates.

[0104] Corresponding to the above method embodiments, this specification also provides embodiments of a three-dimensional model construction training device. Figure 5 This diagram illustrates a structural schematic of a 3D model building training device according to an embodiment of this specification. The 3D model building is used to convert an image into a 3D model. The 3D model building includes an image feature extraction model, a latent vector generation model, an autoencoder model, and an output layer. The autoencoder model includes an encoder and a decoder, such as... Figure 5 As shown, the device includes: The first acquisition module 502 is configured to acquire a first sample set; at least one training sample in the first sample set contains first geometric texture field data of a first three-dimensional model; the first geometric texture field data is used to reflect the signed distance function value and texture value of the effective sampling points of the first three-dimensional model. The first training module 504 is configured to train the autoencoder model using the first sample set to obtain the target autoencoder model. The second acquisition module 506 is configured to acquire a second sample set; at least one training sample in the second sample set contains latent vector data corresponding to a second three-dimensional model and an image corresponding to the second three-dimensional model; the latent vector data corresponding to the second three-dimensional model is obtained by processing the second geometric texture field data of the second three-dimensional model using the encoder of the target autoencoder model; the second geometric texture field data is used to reflect the signed distance function value and texture value of the effective sampling points of the second three-dimensional model; the effective sampling points are sampling points whose distance from the surface of the first three-dimensional model or the surface of the second three-dimensional model is less than a preset length; The second training module 508 is configured to train the image feature extraction model and the latent vector generation model in the 3D construction model using the second sample set, so as to obtain the trained target 3D construction model.

[0105] In an optional embodiment, obtaining the first sample set includes: For any training sample in the first sample set, obtain the model data of the target 3D model; based on the model data of the target 3D model, determine the 3D coordinates of each effective sampling point of the target 3D model in the unit coordinate system space; based on the 3D coordinates of each effective sampling point, determine the signed distance function value and the texture value of each effective sampling point; based on the signed distance function value and the texture value of each effective sampling point, construct any training sample in the first sample set corresponding to the target 3D model.

[0106] In an optional embodiment, determining the three-dimensional coordinates of each effective sampling point of the target three-dimensional model in the unit coordinate system space based on the model data of the target three-dimensional model includes: Based on the model data of the target 3D model, the target 3D model is normalized to a unit coordinate system space; using the preset length as a reference, the unit coordinate system space is voxelized to obtain a first number of voxels; the voxels are cubes; the side length of the voxels is equal to the preset length; the first number of voxels are intersected with the surface of the target 3D model to obtain a second number of intersecting voxels; the vertices of each intersecting voxel are determined as the effective sampling points to obtain the 3D coordinates of each effective sampling point of the target 3D model in the unit coordinate system space.

[0107] In an optional embodiment, determining the signed distance function value and the texture value of each valid sampling point based on their three-dimensional coordinates includes: Based on the three-dimensional coordinates of each valid sampling point, the projection distance from each valid sampling point to the surface of the target three-dimensional model and the centroid coordinates of the projection points of each valid sampling point on the surface of the target three-dimensional model are determined; based on the projection distance from each valid sampling point to the surface of the target three-dimensional model, the signed distance function value of each valid sampling point is determined; based on the centroid coordinates of the projection points of each valid sampling point on the surface of the target three-dimensional model, the texture value of each valid sampling point is determined.

[0108] In an optional embodiment, determining the signed distance function value of each valid sampling point based on the projected distance from each valid sampling point to the surface of the target 3D model includes: For any valid sampling point, the sign of the sign distance function of the valid sampling point is determined using a flooding fill algorithm; the projection distance from the valid sampling point to the surface of the target 3D model is determined as the value of the sign distance function of the valid sampling point; the sign of the sign distance function of the valid sampling point and the value of the sign distance function of the valid sampling point are combined to obtain the sign distance function value of the valid sampling point.

[0109] In an optional embodiment, the model data of the target 3D model includes a texture map; determining the texture value of each valid sampling point based on the centroid coordinates of the projection points of each valid sampling point on the surface of the target 3D model includes: For any valid sampling point, obtain the texture coordinates of each vertex of the target triangular mesh corresponding to the valid sampling point; determine the texture coordinates of the projection point based on the centroid coordinates of the projection point of the valid sampling point on the surface of the target 3D model and the texture coordinates of each vertex of the target triangular mesh; determine the texture value of the projection point using the texture map and the texture coordinates of the projection point; and determine the texture value of the projection point as the texture value of the valid sampling point.

[0110] In an optional embodiment, training the autoencoder model using the first sample set to obtain the target autoencoder model includes: For any training sample in the first sample set, the first geometric texture field data of a first 3D model contained in the training sample is input into the encoder of the autoencoder model to obtain the latent vector data corresponding to the first 3D model output by the encoder; the latent vector data is input into the decoder of the autoencoder model to obtain the third geometric texture field data of the first 3D model reconstructed by the decoder; the third geometric texture field data is used to reflect the symbolic distance function value and texture value of each point in the unit coordinate system space where the first 3D model reconstructed by the decoder is located; based on the third geometric texture field data of the first 3D model, the symbolic distance function value and texture value of each effective sampling point of the first 3D model are determined to obtain the fourth geometric texture field data; the fourth geometric texture field data is used to reflect the symbolic distance function value and texture value of each effective sampling point of the first 3D model reconstructed by the decoder; with the goal of reducing the difference between the first geometric texture field data and the fourth geometric texture field data, the model parameters of the autoencoder model are adjusted to obtain the trained target autoencoder model.

[0111] In an optional embodiment, the output layer of the target 3D model is used to construct the target 3D model based on the geometric texture field data of the target 3D model reconstructed by the decoder; the geometric texture field data is used to reflect the signed distance function value and texture value of each point in the unit coordinate system space where the target 3D model reconstructed by the decoder is located.

[0112] In an optional embodiment, training the image feature extraction model and the latent vector generation model in the 3D construction model using the second sample set to obtain the trained target 3D construction model includes: For any training sample in the second sample set, the image feature extraction model with frozen image input parameters corresponding to the second 3D model contained in the training sample is used to obtain image feature data output by the image feature extraction model; the image feature data is input into the latent vector generation model to obtain latent vector data output by the latent vector generation model; the model parameters of the latent vector generation model are adjusted with the goal of reducing the difference between the latent vector data output by the latent vector generation model and the latent vector data corresponding to the second 3D model contained in the training sample, to obtain the trained target 3D construction model.

[0113] The 3D model building training device provided in this embodiment, in order to realize the conversion of an object's image into a 3D model of the object during the model application stage, can pre-train a 3D model for converting images into 3D models. During the model training process, a first sample set can be obtained firstly. At least one training sample in the first sample set contains first geometric texture field data of a first 3D model. This first geometric texture field data reflects the signed distance function value and texture value of the effective sampling points of the first 3D model. The effective sampling points are sampling points whose distance from the surface of the 3D model is less than a preset length. After obtaining the first sample set, an autoencoder model in the 3D model can be trained using the first sample set. The autoencoder model includes an encoder and a decoder. During training, the encoder of the autoencoder model can be trained to convert the geometric texture field data of the 3D model into latent vectors, and the decoder of the autoencoder model can be trained to reconstruct the geometric texture field of the 3D model based on the latent vectors. After training the autoencoder model, a target autoencoder model can be obtained. Then, a second sample set can be acquired. At least one training sample in the second sample set contains latent vector data corresponding to a second 3D model and an image corresponding to that model. The latent vector data corresponding to the second 3D model is obtained by processing the second geometric texture field data of the second 3D model using the encoder of the target autoencoder model. This second geometric texture field data reflects the signed distance function value and texture value of the effective sampling points of the second 3D model. After obtaining the second sample set, the image feature extraction model and latent vector generation model in the 3D model can be trained using this set. During training, the latent vector generation model can be trained to generate the latent vector data corresponding to the 3D model based on the image feature data extracted from the image corresponding to the 3D model by the image feature extraction model. Therefore, after training, in the application phase of the target 3D model, the image of the object is input into the target 3D model. The target 3D model can extract image feature data from the image using an image feature extraction model, and then use a latent vector generation model to generate latent vector data corresponding to the object based on the image feature data. Next, the decoder of the autoencoder model can generate geometric texture field data corresponding to the object based on the object's latent vector data. Finally, the output layer can construct the corresponding 3D model of the object based on the object's geometric texture field data, ultimately realizing the transformation from image to 3D model. The target 3D model can quickly and accurately generate the corresponding 3D model based on the object's image in the application phase, eliminating the need for manual modeling, effectively improving the modeling efficiency of 3D models, and saving modeling costs.

[0114] The above is a schematic scheme of a three-dimensional construction model training device according to this embodiment. It should be noted that the technical solution of this three-dimensional construction model training device and the technical solution of the three-dimensional construction model training method described above belong to the same concept. For details not described in detail in the technical solution of the three-dimensional construction model training device, please refer to the description of the technical solution of the three-dimensional construction model training method described above.

[0115] Corresponding to the above method embodiments, this specification also provides embodiments of a three-dimensional model construction apparatus. Figure 6 A schematic diagram of a three-dimensional model building apparatus according to one embodiment of this specification is shown. Figure 6 As shown, the device includes: The acquisition module 602 is configured to acquire the target image of the target object; The first input module 604 is configured to input the target image into the image feature extraction model in the target 3D construction model to obtain the image feature data corresponding to the target object output by the image feature extraction model; wherein, the target 3D construction model is trained by the above method; The second input module 606 is configured to input the image feature data into the latent vector generation model in the target 3D construction model to obtain the latent vector data corresponding to the target object output by the latent vector generation model. The third input module 608 is configured to input the latent vector data into the decoder of the autoencoder model in the target 3D construction model to obtain the geometric texture field data corresponding to the target object output by the decoder; the geometric texture field data is used to reflect the signed distance function value and texture value of each point in the unit coordinate system space where the 3D model corresponding to the target object is located; The construction module 610 is configured to construct a three-dimensional model of the target object based on the geometric texture field data corresponding to the target object.

[0116] In an optional embodiment, constructing a three-dimensional model of the target object based on the geometric texture field data corresponding to the target object includes: Using the moving cube algorithm, based on the signed distance function values ​​of each point in the unit coordinate system space where the 3D model corresponding to the target object is located, the vertex coordinates of each triangular mesh on the surface of the 3D model corresponding to the target object are determined; based on the texture values ​​of each point in the unit coordinate system space where the 3D model corresponding to the target object is located, the texture values ​​of the vertices of each triangular mesh are determined; based on the vertex coordinates of each triangular mesh on the surface of the 3D model and the texture values ​​of the vertices of each triangular mesh, the 3D model corresponding to the target object is constructed.

[0117] In an optional embodiment, the target object includes a target product, and after constructing a 3D model of the target object based on the geometric texture field data corresponding to the target object, the method further includes: A three-dimensional product model corresponding to the target product is displayed to the user in a virtual space to present the display effect of the target product in the virtual space.

[0118] The above is a schematic scheme of a three-dimensional model building device according to this embodiment. It should be noted that the technical solution of this three-dimensional model building device and the technical solution of the three-dimensional model building method described above belong to the same concept. For details not described in detail in the technical solution of the three-dimensional model building device, please refer to the description of the technical solution of the three-dimensional model building method described above.

[0119] Figure 7 A structural block diagram of a computing device 700 according to one embodiment of this specification is shown. The components of the computing device 700 include, but are not limited to, a memory 710 and a processor 720. The processor 720 is connected to the memory 710 via a bus 730, and a database 750 is used to store data.

[0120] The computing device 700 also includes an access device 740, which enables the computing device 700 to communicate via one or more networks 760. Examples of these networks include Public Switched Telephone Network (PSTN), Local Area Network (LAN), Wide Area Network (WAN), Personal Area Network (PAN), or combinations of communication networks such as the Internet. The access device 740 may include one or more of any type of wired or wireless network interface (e.g., a network interface card (NIC)), such as an IEEE 802.11 Wireless Local Area Network (WLAN) wireless interface, a Wi-MAX (Worldwide Interoperability for Microwave Access) interface, an Ethernet interface, a Universal Serial Bus (USB) interface, a cellular network interface, a Bluetooth interface, or a Near Field Communication (NFC) interface.

[0121] In one embodiment of this specification, the above-described components of the computing device 700 and Figure 7Other components, not shown, can also be connected to each other, for example, via a bus. It should be understood that... Figure 7 The block diagram of the computing device shown is for illustrative purposes only and is not intended to limit the scope of this specification. Those skilled in the art can add or replace other components as needed.

[0122] The computing device 700 can be any type of stationary or mobile computing device, including mobile computers or mobile computing devices (e.g., tablet computers, personal digital assistants, laptop computers, notebook computers, netbooks, etc.), mobile phones (e.g., smartphones), wearable computing devices (e.g., smartwatches, smart glasses, etc.) or other types of mobile devices, or stationary computing devices such as desktop computers or personal computers (PCs). The computing device 700 can also be a mobile or stationary server.

[0123] The processor 720 is used to execute the following computer-executable instructions, which, when executed by the processor, implement the steps of the above-mentioned three-dimensional model training method or three-dimensional model construction method.

[0124] The above is an illustrative scheme of a computing device according to this embodiment. It should be noted that the technical solution of this computing device belongs to the same concept as the technical solution of the above-described 3D model training method or 3D model construction method. For details not described in detail in the technical solution of the computing device, please refer to the description of the technical solution of the above-described 3D model training method or 3D model construction method.

[0125] An embodiment of this specification also provides a computer-readable storage medium storing computer-executable instructions that, when executed by a processor, implement the steps of the above-described 3D model training method or 3D model construction method.

[0126] The above is an illustrative scheme of a computer-readable storage medium according to this embodiment. It should be noted that the technical solution of this storage medium belongs to the same concept as the technical solution of the above-described 3D model training method or 3D model construction method. For details not described in detail in the technical solution of the storage medium, please refer to the description of the technical solution of the above-described 3D model training method or 3D model construction method.

[0127] An embodiment of this specification also provides a computer program product, including a computer program or instructions that, when executed by a processor, implement the steps of the above-described 3D model training method or 3D model construction method.

[0128] The above is an illustrative scheme of a computer program product according to this embodiment. It should be noted that the technical solution of this computer program product belongs to the same concept as the technical solution of the above-mentioned 3D model training method or 3D model construction method. For details not described in detail in the technical solution of the computer program product, please refer to the description of the technical solution of the above-mentioned 3D model training method or 3D model construction method.

[0129] The foregoing has described specific embodiments of this specification. Other embodiments are within the scope of the appended claims. In some cases, the actions or steps recited in the claims may be performed in a different order than that shown in the embodiments and may still achieve the desired result. Furthermore, the processes depicted in the drawings do not necessarily require the specific or sequential order shown to achieve the desired result. In some embodiments, multitasking and parallel processing are possible or may be advantageous.

[0130] The computer program / instructions include computer program code, which may be in the form of source code, object code, executable file, or certain intermediate forms. The computer-readable medium may include: any entity or device capable of carrying the computer program code, recording media, USB flash drive, portable hard drive, magnetic disk, optical disk, computer memory, read-only memory (ROM), random access memory (RAM), electrical carrier signals, telecommunication signals, and software distribution media, etc. It should be noted that the content included in the computer-readable medium may be appropriately added or removed according to the requirements of patent practice. For example, in some regions, according to patent practice, computer-readable media may not include electrical carrier signals and telecommunication signals.

[0131] It should be noted that, for the sake of simplicity, the foregoing method embodiments are all described as a series of actions. However, those skilled in the art should understand that the embodiments in this specification are not limited to the described order of actions, because according to the embodiments in this specification, some steps can be performed in other orders or simultaneously. Furthermore, those skilled in the art should also understand that the embodiments described in this specification are all preferred embodiments, and the actions and modules involved are not necessarily essential to the embodiments in this specification.

[0132] In the above embodiments, the descriptions of each embodiment have different focuses. For parts not described in detail in a certain embodiment, please refer to the relevant descriptions of other embodiments.

[0133] The preferred embodiments disclosed above are merely illustrative of this specification. The optional embodiments do not exhaustively describe all details, nor do they limit the invention to the specific implementations described. Clearly, many modifications and variations can be made based on the embodiments described herein. These embodiments are selected and specifically described in this specification to better explain the principles and practical applications of the embodiments, thereby enabling those skilled in the art to better understand and utilize this specification. This specification is limited only by the claims and their full scope and equivalents.

Claims

1. A method for training a 3D model, wherein the 3D model is used to convert an image into a 3D model, the 3D model comprising an image feature extraction model, a latent vector generation model, an autoencoder model, and an output layer, the autoencoder model comprising an encoder and a decoder, the method comprising: Obtain the first sample set; At least one training sample in the first sample set contains first geometric texture field data of a first three-dimensional model; The first geometric texture field data is used to reflect the signed distance function value and texture value of the effective sampling points of the first three-dimensional model; The autoencoder model is trained using the first sample set to obtain the target autoencoder model; A second sample set is obtained; at least one training sample in the second sample set contains latent vector data corresponding to a second 3D model and an image corresponding to the second 3D model; the latent vector data corresponding to the second 3D model is obtained by processing the second geometric texture field data of the second 3D model using the encoder of the target autoencoder model; the second geometric texture field data is used to reflect the signed distance function value and texture value of the effective sampling points of the second 3D model; the effective sampling points are sampling points whose distance from the surface of the first 3D model or the surface of the second 3D model is less than a preset length; The image feature extraction model and the latent vector generation model in the 3D construction model are trained using the second sample set to obtain the trained target 3D construction model.

2. The three-dimensional model training method according to claim 1, wherein obtaining the first sample set includes: For any training sample in the first sample set, obtain the model data of the target 3D model; Based on the model data of the target 3D model, determine the 3D coordinates of each effective sampling point of the target 3D model in the unit coordinate system space; Based on the three-dimensional coordinates of each valid sampling point, determine the signed distance function value and the texture value of each valid sampling point; Based on the symbolic distance function value and texture value of each effective sampling point, any training sample in the first sample set corresponding to the target 3D model is constructed.

3. The three-dimensional model training method according to claim 2, wherein determining the three-dimensional coordinates of each effective sampling point of the target three-dimensional model in the unit coordinate system space based on the model data of the target three-dimensional model includes: Based on the model data of the target 3D model, the target 3D model is normalized to a unit coordinate system space; Based on the preset length, the unit coordinate system space is voxelized to obtain a first number of voxels; the voxels are cubes. The side length of the voxel is equal to the preset length; The first number of voxels are intersected with the surface of the target 3D model to obtain a second number of intersecting voxels that intersect with the surface of the target 3D model; The vertices of each intersecting voxel are determined as the effective sampling points, and the three-dimensional coordinates of each effective sampling point of the target three-dimensional model in the unit coordinate system space are obtained.

4. The three-dimensional model training method according to claim 2, wherein determining the signed distance function value and the texture value of each effective sampling point based on the three-dimensional coordinates of each effective sampling point includes: Based on the three-dimensional coordinates of each effective sampling point, the projection distance of each effective sampling point to the surface of the target three-dimensional model and the centroid coordinates of the projection points of each effective sampling point on the surface of the target three-dimensional model are determined. Based on the projection distance from each valid sampling point to the surface of the target 3D model, the sign distance function value of each valid sampling point is determined; The texture value of each effective sampling point is determined based on the centroid coordinates of the projection points of each effective sampling point on the surface of the target 3D model.

5. The three-dimensional model training method according to claim 4, wherein determining the signed distance function value of each effective sampling point based on the projection distance from each effective sampling point to the surface of the target three-dimensional model includes: For any valid sampling point, the sign of the sign distance function of the valid sampling point is determined using the flooding fill algorithm; The projection distance from the effective sampling point to the surface of the target 3D model is determined as the value of the signed distance function of the effective sampling point; The sign distance function value of the effective sampling point is obtained by combining the positive and negative values ​​of the sign distance function of the effective sampling point with the numerical value of the sign distance function of the effective sampling point.

6. The three-dimensional model training method according to claim 4, wherein the model data of the target three-dimensional model includes texture maps; determining the texture value of each effective sampling point based on the centroid coordinates of the projection points of each effective sampling point on the surface of the target three-dimensional model includes: For any valid sampling point, obtain the texture coordinates of each vertex of the target triangular mesh corresponding to the valid sampling point; Based on the centroid coordinates of the projection points of the effective sampling points on the surface of the target 3D model and the texture coordinates of each vertex of the target triangular mesh, the texture coordinates of the projection points are determined. Using the texture map, the texture value of the projection point is determined based on the texture coordinates of the projection point; The texture value of the projection point is determined as the texture value of the effective sampling point.

7. The three-dimensional model training method according to claim 1, wherein training the autoencoder model using the first sample set to obtain the target autoencoder model comprises: For any training sample in the first sample set, the first geometric texture field data of a first three-dimensional model contained in the training sample is input into the encoder of the autoencoder model to obtain the latent vector data corresponding to the first three-dimensional model output by the encoder. The latent vector data is input into the decoder of the autoencoder model to obtain the third geometric texture field data of the first three-dimensional model reconstructed by the decoder; The third geometric texture field data is used to reflect the symbolic distance function value and texture value of each point in the unit coordinate system space where the first three-dimensional model reconstructed by the decoder is located; Based on the third geometric texture field data of the first three-dimensional model, the symbolic distance function value and texture value of each effective sampling point of the first three-dimensional model are determined to obtain the fourth geometric texture field data; the fourth geometric texture field data is used to reflect the symbolic distance function value and texture value of each effective sampling point of the first three-dimensional model reconstructed by the decoder; With the goal of reducing the difference between the first geometric texture field data and the fourth geometric texture field data, the model parameters of the autoencoder model are adjusted to obtain the trained target autoencoder model.

8. The three-dimensional model training method according to claim 1, wherein the output layer of the target three-dimensional model is used to construct the target three-dimensional model based on the geometric texture field data of the target three-dimensional model reconstructed by the decoder; the geometric texture field data is used to reflect the signed distance function value and texture value of each point in the unit coordinate system space where the target three-dimensional model reconstructed by the decoder is located.

9. The 3D model training method according to claim 1, wherein training the image feature extraction model and the latent vector generation model in the 3D model using the second sample set to obtain the trained target 3D model includes: For any training sample in the second sample set, the image feature extraction model corresponding to the second three-dimensional model contained in the training sample is frozen to obtain the image feature data output by the image feature extraction model; The image feature data is input into the latent vector generation model to obtain the latent vector data output by the latent vector generation model. With the goal of reducing the difference between the latent vector data output by the latent vector generation model and the latent vector data corresponding to the second 3D model contained in the training samples, the model parameters of the latent vector generation model are adjusted to obtain the trained target 3D construction model.

10. A method for constructing a three-dimensional model, comprising: Acquire the target image of the target object; The target image is input into the image feature extraction model in the target 3D construction model to obtain the image feature data corresponding to the target object output by the image feature extraction model; wherein, the target 3D construction model is trained by any one of claims 1 to 9; The image feature data is input into the latent vector generation model in the target 3D construction model to obtain the latent vector data corresponding to the target object output by the latent vector generation model; The latent vector data is input into the decoder of the autoencoder model in the target 3D model to obtain the geometric texture field data corresponding to the target object output by the decoder; the geometric texture field data is used to reflect the signed distance function value and texture value of each point in the unit coordinate system space where the 3D model corresponding to the target object is located; Based on the geometric texture field data corresponding to the target object, a three-dimensional model of the target object is constructed.

11. The three-dimensional model construction method according to claim 10, wherein constructing a three-dimensional model corresponding to the target object based on the geometric texture field data corresponding to the target object comprises: Using the moving cube algorithm, based on the signed distance function value of each point in the unit coordinate system space where the 3D model corresponding to the target object is located, the vertex coordinates of each triangular mesh on the surface of the 3D model corresponding to the target object are determined; Based on the texture values ​​of each point in the unit coordinate system space where the three-dimensional model corresponding to the target object is located, the texture values ​​of the vertices of each triangular mesh are determined. Based on the vertex coordinates of each triangular mesh on the surface of the 3D model and the texture values ​​of the vertices of each triangular mesh, a 3D model corresponding to the target object is constructed.

12. The three-dimensional model construction method according to claim 11, wherein the target object includes a target product, and after constructing a three-dimensional model corresponding to the target object based on the geometric texture field data corresponding to the target object, the method further includes: A three-dimensional product model corresponding to the target product is displayed to the user in a virtual space to present the display effect of the target product in the virtual space.

13. A computing device, comprising: Memory and processor; The memory is used to store computer-executable instructions, and the processor is used to execute the computer-executable instructions, which, when executed by the processor, implement the steps of the method according to any one of claims 1 to 12.

14. A computer-readable storage medium storing computer-executable instructions that, when executed by a processor, implement the steps of the method according to any one of claims 1 to 12.

15. A computer program product comprising a computer program or instructions which, when executed by a processor, implement the steps of the method according to any one of claims 1 to 12.

Images