A three-dimensional model construction method, device and related equipment

Abstract

This invention provides a method, apparatus, and related equipment for constructing a 3D model, relating to the field of data processing technology. The method includes: acquiring image data of a first scene and capturing pose data of the image data; training an initial 3D model based on the image data, wherein the initial 3D model is a model with a first level of precision; dividing the location region corresponding to the first scene into multiple target blocks based on the image data and the pose data; determining a first model corresponding to each target block based on the initial 3D model, and determining first data corresponding to each target block based on the image data; training the first model corresponding to each target block based on the first data to obtain a second model corresponding to each target block; and generating a target model based on the second models corresponding to the multiple target blocks. This invention can reduce the consumption of computing resources.

Description

Technical Field

[0001] This invention relates to the field of data processing technology, specifically to a method, apparatus, and related equipment for constructing three-dimensional models. Background Technology

[0002] Constructing a 3D model of a scene enables scene simulation. One related technique is 3D Gaussian Splatting (3DGS), which is used to train a high-resolution 3D model. However, because 3DGS is used to train high-resolution 3D models, the number of Gaussian units in the 3D model increases with the scene size, leading to significant computational resource consumption.

[0003] It is evident that the related technologies suffer from the problem of consuming significant computational resources for 3D model construction. Summary of the Invention

[0004] This invention provides a method, apparatus, and related equipment for constructing three-dimensional models to solve the problem that constructing three-dimensional models consumes a lot of computing resources in related technologies.

[0005] To solve the above problems, the present invention is implemented as follows: In a first aspect, embodiments of the present invention provide a three-dimensional model construction method, including: Acquire image data of the first scene and pose data of the captured image data; An initial 3D model is trained based on the image data, and the initial 3D model is a model with first-precision accuracy; Based on the pose data, the location region corresponding to the first scene is determined, as well as a first size threshold and a second size threshold, wherein the second size threshold is less than the first size threshold; The location region is divided based on the first size threshold to obtain multiple initial blocks; Based on the second size threshold, the plurality of initial blocks are divided to obtain a plurality of target blocks; Expand the region of each target block; Based on the initial 3D model, a first model corresponding to each expanded target block is determined, and based on the image data, a first data corresponding to each expanded target block is determined. The first model is the model of the first precision corresponding to the corresponding expanded target block, and the first data is the data of the region where the corresponding expanded target block is located in the image data. Based on the first data corresponding to each target block, the first model corresponding to each target block is trained to obtain the second model corresponding to each target block. The second model corresponding to each target block is a model with second precision, which is higher than the first precision. A target model is generated based on the second model corresponding to the multiple target blocks, and the target model is the second precision model corresponding to the first scene; The image data is multi-view image data, and the multiple initial blocks include a first block. The process of dividing the multiple initial blocks based on the second size threshold includes: Determine a first quantity corresponding to the first block, wherein the first quantity is used to represent the number of image data in the multi-view image data that contain the location of the first block; If the first quantity corresponding to the first block is greater than a set quantity threshold, and the size of the first block is greater than the second size threshold, the first block is divided.

[0006] In one embodiment, the image data includes a depth map, and the method further includes: Generate a first mask, which is used to cover a preset object in the image data; Depth distance is determined based on the depth map, and a second mask is generated based on the depth map. The second mask is used to cover areas where the depth distance is greater than a first set distance. The rendering depth of the target model is determined, and a third mask is generated based on the rendering depth. The third mask is used to cover the region where the rendering depth is greater than a second set distance. The target model is adjusted based on the first mask, the second mask, and the third mask.

[0007] In one embodiment, the plurality of target blocks includes a second block, and the method further includes: Back-projecting the depth map yields multiple initial point clouds corresponding to the second block, wherein the depth map is the depth map included in the image data or the depth map obtained by 3DGS rendering; By filtering the multiple initial point clouds, at least one first point cloud is obtained; The second model corresponding to the second block is updated based on the at least one first point cloud; The filtering of the plurality of initial point clouds includes at least one of the following: Delete the initial point clouds that are masked in the plurality of initial point clouds; Delete the initial point cloud whose depth distance is greater than a third preset distance from the plurality of initial point clouds; Random sampling is performed on the multiple initial point clouds; Delete the initial point clouds whose distance from the center of the Gaussian point is greater than a fourth predetermined distance. The center of the Gaussian point is the center of the multiple first Gaussian points included in the second model corresponding to the second block.

[0008] In one embodiment, the method further includes: Based on the image data, a score is calculated for each of the plurality of first Gaussian points, and the score is used to characterize the visibility of the corresponding first Gaussian point from multiple perspectives. Multiple second Gaussian points are selected from the plurality of first Gaussian points, wherein the scores corresponding to the plurality of second Gaussian points are greater than the scores corresponding to other first Gaussian points, and the other first Gaussian points are higher than the plurality of first Gaussian points excluding the plurality of second Gaussian points; Multiple third Gaussian points are selected from the plurality of second Gaussian points, wherein the scores corresponding to the plurality of third Gaussian points are greater than the scores corresponding to other second Gaussian points, and the other second Gaussian points are higher than the plurality of second Gaussian points other than the plurality of third Gaussian points; Copy the multiple third Gaussian points; The second model corresponding to the second block is updated based on the plurality of second Gaussian points and the plurality of copied third Gaussian points.

[0009] In one embodiment, the method includes: Construct conditional constraints corresponding to each target block, wherein the conditional constraints include at least one of planar consistency constraints, lighting constraints, and geometric constraints; The step of training a first model corresponding to each target block based on the first data corresponding to each target block to obtain a second model corresponding to each target block includes: Based on the first data and condition constraints corresponding to each target block, the first model corresponding to each target block is trained to obtain the second model corresponding to each target block.

[0010] In one embodiment, training the first model corresponding to each target block based on the first data corresponding to each target block includes: Based on a preset strategy, the first model corresponding to each target block is trained using the first data corresponding to each target block. The preset strategy includes at least one of the following: The caching acceleration strategy is to read only the model parameters when rendering the depth map and Gaussian encryption process, and to read only the image data and pose data when deleting point clouds and training with constraints. A multi-level resolution training strategy, wherein the multi-level resolution training strategy is a strategy of sequentially increasing the resolution of the image data used to train the model; The decoupling strategy is a strategy of training different models for corresponding areas on the ground and / or in the sky.

[0011] Secondly, embodiments of the present invention provide a three-dimensional model construction apparatus, comprising: The acquisition module is used to acquire image data of the first scene and pose data of the captured image data; The first training module is used to train an initial three-dimensional model based on the image data, wherein the initial three-dimensional model is a model with a first precision. The first determining module is used to determine the location region corresponding to the first scene based on the pose data, as well as a first size threshold and a second size threshold, wherein the second size threshold is less than the first size threshold. The first partitioning module is used to partition the location region based on the first size threshold to obtain multiple initial blocks; The second partitioning module is used to partition the plurality of initial blocks based on the second size threshold to obtain a plurality of target blocks; An extension module is used to expand the region of each target block; The second determining module is used to determine the first model corresponding to each of the expanded target blocks based on the initial three-dimensional model, and to determine the first data corresponding to each of the expanded target blocks based on the image data. The first model is the model of the first precision corresponding to the corresponding expanded target block, and the first data is the data of the region where the corresponding expanded target block is located in the image data. The second training module is used to train the first model corresponding to each target block based on the first data corresponding to each target block, so as to obtain the second model corresponding to each target block. The second model corresponding to each target block is a model with a second precision, which is higher than the first precision. The first generation module is used to generate a target model based on the second model corresponding to the multiple target blocks, wherein the target model is the model with the second precision corresponding to the first scene.

[0012] Thirdly, embodiments of the present invention provide an electronic device, including: a processor, a memory, and a program stored in the memory and executable on the processor, wherein when the program is executed by the processor, it implements the steps of the three-dimensional model construction method described in the first aspect.

[0013] Fourthly, embodiments of the present invention provide a computer-readable storage medium storing a computer program, which, when executed by a processor, implements the steps of the three-dimensional model construction method described in the first aspect.

[0014] Fifthly, the present invention also provides a computer program product, including computer instructions, which, when executed by a processor, implement the steps of the three-dimensional model construction method described in the first aspect.

[0015] In this embodiment of the invention, image data of a first scene and pose data of the captured image data are acquired; an initial 3D model is trained based on the image data, the initial 3D model being a model with a first precision; a location region corresponding to the first scene is determined based on the pose data, along with a first size threshold and a second size threshold, the second size threshold being smaller than the first size threshold; the location region is divided based on the first size threshold to obtain multiple initial blocks; the multiple initial blocks are further divided based on the second size threshold to obtain multiple target blocks; the region of each target block is expanded; a first model corresponding to each expanded target block is determined based on the initial 3D model, and first data corresponding to each expanded target block is determined based on the image data, the first model being the model with the first precision corresponding to the corresponding expanded target block, and the first data being the location of the corresponding expanded target block in the image data. The data is distributed as follows: First model corresponding to each target block is trained based on the first data corresponding to each target block to obtain a second model corresponding to each target block. The second model corresponding to each target block is a model with a second precision, which is higher than the first precision. A target model is generated based on the second models corresponding to the multiple target blocks. The target model is a model with the second precision corresponding to the first scene. The image data is multi-view image data. The multiple initial blocks include a first block. Dividing the multiple initial blocks based on a second size threshold includes: determining a first quantity corresponding to the first block, where the first quantity is the number of image data in the multi-view image data where the first block is located; dividing the first block when the first quantity corresponding to the first block is greater than a set quantity threshold and the size of the first block is greater than the second size threshold. In this way, by dividing the location region corresponding to the first scene into multiple target blocks, and then determining the first data and first model corresponding to each target block, the first model corresponding to each target block can be trained based on the first data corresponding to each target block to obtain the second model corresponding to each target block. This allows the second models of different target blocks to be trained in batches, reducing the consumption of computing resources at the same time. Then, a target model is generated based on the second models corresponding to the multiple target blocks, so that the second-precision target model corresponding to the first scene is obtained through training, effectively reducing the consumption of computing resources. Attached Figure Description

[0016] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the description of the embodiments of the present invention will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0017] Figure 1 This is a flowchart of a three-dimensional model construction method provided in an embodiment of the present invention; Figure 2 This is an overall schematic diagram of the three-dimensional model construction provided in the embodiments of the present invention; Figure 3 This is a structural diagram of a three-dimensional model building device provided in an embodiment of the present invention; Figure 4 This is a structural diagram of an electronic device provided in an embodiment of the present invention. Detailed Implementation

[0018] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0019] Please see Figure 1 , Figure 1 This is a flowchart of a three-dimensional model construction method provided in an embodiment of the present invention, such as... Figure 1 As shown, it includes the following steps: Step 101: Obtain image data of the first scene and pose data of the captured image data.

[0020] The first scenario mentioned above is a scenario that requires the construction of a 3D model, such as an indoor scene (such as a shopping mall or indoor sports hall) or an outdoor scene (such as a landmark building or scenic spot).

[0021] The image data mentioned above is image data collected in the first scene, which can be image data from one viewpoint or image data from multiple viewpoints.

[0022] The pose data mentioned above is the pose data of the captured image data. By using the pose data and image data, the overall situation of the first scene can be determined, thereby realizing the construction of a three-dimensional model of the first scene.

[0023] In some implementations, acquiring image data of the first scene and pose data of the captured image data can be achieved by deploying one or more shooting devices at the location of the first scene and acquiring image data of the first scene through one or more shooting devices.

[0024] In some implementations, the image data includes multi-view image data and its corresponding camera parameters, sparse point clouds, depth maps, normal maps, and mask maps. The camera parameters and sparse point clouds can be generated using a structured bundle method; the depth maps, normal maps, and mask maps can be obtained through inference from a pre-trained deep learning network to supplement geometric and semantic prior information, thereby constructing a unified multimodal input dataset.

[0025] Step 102: Train an initial 3D model based on the image data. The initial 3D model is a model with first-precision accuracy.

[0026] The aforementioned initial 3D model is a model with first-precision resolution. First-precision resolution means a relatively low resolution. Training an initial 3D model with first-precision resolution using image data can achieve the training of an initial 3D model that can represent the entire first scene with less computing resources, so as to facilitate subsequent training at higher resolution.

[0027] In some implementations, an initial 3D model can be trained by performing 3D Gaussian splash training at a global scale based on a first precision.

[0028] Step 103: Determine the location region corresponding to the first scene based on the pose data, as well as a first size threshold and a second size threshold, wherein the second size threshold is less than the first size threshold; Step 104: Divide the location region based on the first size threshold to obtain multiple initial blocks; Step 105: Divide the multiple initial blocks according to the second size threshold to obtain multiple target blocks; The aforementioned location region is the region corresponding to the first scene, which can be a two-dimensional region or a three-dimensional region. The first scene is located within the location region. The aforementioned multiple target blocks are blocks obtained by dividing the location region. Multiple target blocks together form a complete location region.

[0029] It should be noted that multiple target blocks can be blocks of the same size or blocks of different sizes.

[0030] In some implementations, the size parameters of the target blocks can be pre-configured, and then the location region can be divided according to the size parameters of the target blocks to obtain multiple target blocks.

[0031] Step 106: Expand the region of each target block; Step 107: Determine the first model corresponding to each expanded target block based on the initial 3D model, and determine the first data corresponding to each expanded target block based on the image data. The first model is the model of the first precision corresponding to the corresponding expanded target block, and the first data is the data of the region where the corresponding expanded target block is located in the image data.

[0032] The first model mentioned above is a model with the first precision. Each target block corresponds to a first model. The first model can be the model of the initial 3D model in the target block. The first model can be used to render the corresponding target block with the first precision.

[0033] In some implementations, the target block location can be rendered based on the initial 3D model to obtain the corresponding first model, or the model parameters of the initial 3D model can be truncated to obtain the first model corresponding to the target block.

[0034] The first data mentioned above is the data corresponding to the target block. Since the target block is obtained by segmenting the first scene, there is also corresponding image data (i.e., the first data) for the target block. The model corresponding to the target block can be trained through the first data to obtain a 3D model that can render the target block.

[0035] In some implementations, the first data corresponding to each target block can be obtained by segmenting the image data of the first scene.

[0036] The above method expands the region of each target block, making the expanded region larger than the original region. The expanded portion actually lies within other adjacent blocks. By expanding the region of the target block, partial images of adjacent blocks can be rendered during rendering, ensuring continuity with the rendered images of adjacent blocks and improving rendering quality.

[0037] In some implementations, the region of each target block can be expanded by pre-configuring the expansion length parameter, and then expanding the region of the target block from different directions by configuring the expansion length parameter.

[0038] In other implementations, the region of each target block can be expanded by calculating expansion parameters based on the area of ​​the target block, and then expanding the region of the target block from different directions using the expansion parameters.

[0039] Furthermore, after expanding the target block region, since the target block region changes at this time, it is necessary to determine the first model and the first data in sequence according to the expanded target block, so that the first model can be trained using the first data to obtain the second model for each expanded target block.

[0040] In this embodiment, the region of each target block is expanded; a first model corresponding to each expanded target block is determined based on the initial 3D model, and first data corresponding to each expanded target block is determined based on the image data. Thus, by expanding the region of the target blocks, the rendering range of the second model of the target blocks can be expanded. This allows for the rendering of edge regions of different target blocks by using the expanded portions of the blocks when generating target models based on the second models of multiple target blocks, resulting in continuous rendering results and effectively improving the rendering effect.

[0041] Step 108: Train the first model corresponding to each target block based on the first data corresponding to each target block to obtain the second model corresponding to each target block. The second model corresponding to each target block is a model with second precision, which is higher than the first precision.

[0042] The second precision is higher than the first precision. The second precision is a higher resolution. The first model corresponding to the target block is trained with the first data to obtain the second model, thus achieving a higher resolution second model for each target block.

[0043] It should be noted that because the second precision is higher than the first precision, training the second-precision model will consume more computational resources. However, since the second model is designed for a single target block, rather than training a second-precision model for the entire first scene, it can be trained in batches for different target blocks, reducing the computational resource consumption at any given time. Thus, by employing a multi-level resolution training strategy, low-resolution data is used in the early stages of training for rapid global structure modeling, and the input resolution is gradually increased as training progresses to recover scene details layer by layer. During training, unified multi-level resolution joint optimization is performed on multimodal data such as images, depth, normals, and masks, thereby reducing computational overhead while improving model convergence speed and reconstruction accuracy.

[0044] In some implementations, the second model can be trained by three-dimensional Gaussian splashing of the first model corresponding to each target block at the block scale based on the second precision.

[0045] Step 109: Generate a target model based on the second model corresponding to the multiple target blocks, wherein the target model is the model with the second precision corresponding to the first scene.

[0046] It should be understood that since the second model is a second-precision model for each target block, by generating the target model based on the second model corresponding to the multiple target blocks, the target model can be a second-precision model corresponding to the first scene, which can perform high-resolution rendering of the first scene.

[0047] In some implementations, generating a target model based on the second models corresponding to the multiple target blocks can be achieved by concatenating the second models of the multiple target blocks to obtain the target model; or, it can be achieved by extracting the model parameters of the second model of each target block, and then generating the model parameters of the target model based on the model parameters of the multiple target blocks, thereby obtaining the target model.

[0048] In this embodiment of the invention, image data of a first scene and pose data of the captured image data are acquired; an initial 3D model is trained based on the image data, the initial 3D model being a model with a first precision; a location region corresponding to the first scene is determined based on the pose data, along with a first size threshold and a second size threshold, the second size threshold being smaller than the first size threshold; the location region is divided based on the first size threshold to obtain multiple initial blocks; the multiple initial blocks are further divided based on the second size threshold to obtain multiple target blocks; the region of each target block is expanded; a first model corresponding to each expanded target block is determined based on the initial 3D model, and first data corresponding to each expanded target block is determined based on the image data, the first model being the model with the first precision corresponding to the corresponding expanded target block, and the first data being the location of the corresponding expanded target block in the image data. The data is distributed as follows: First model corresponding to each target block is trained based on the first data corresponding to each target block to obtain a second model corresponding to each target block. The second model corresponding to each target block is a model with a second precision, which is higher than the first precision. A target model is generated based on the second models corresponding to the multiple target blocks. The target model is a model with the second precision corresponding to the first scene. The image data is multi-view image data. The multiple initial blocks include a first block. Dividing the multiple initial blocks based on a second size threshold includes: determining a first quantity corresponding to the first block, where the first quantity is the number of image data in the multi-view image data where the first block is located; dividing the first block when the first quantity corresponding to the first block is greater than a set quantity threshold and the size of the first block is greater than the second size threshold. In this way, by dividing the location region corresponding to the first scene into multiple target blocks, and then determining the first data and first model corresponding to each target block, the first model corresponding to each target block can be trained based on the first data corresponding to each target block to obtain the second model corresponding to each target block. This allows the second models of different target blocks to be trained in batches, reducing the consumption of computing resources at the same time. Then, a target model is generated based on the second models corresponding to the multiple target blocks, so that the second-precision target model corresponding to the first scene is obtained through training, effectively reducing the consumption of computing resources.

[0049] Since the aforementioned pose data is pose data from captured image data, the overall range (i.e., location region) of the first scene can be calculated using the pose data. In some embodiments, determining the location region corresponding to the first scene based on the pose data can be achieved by calculating the location region using a preset algorithm (such as back-projecting the image data based on the pose data). In other embodiments, determining the location region corresponding to the first scene based on the pose data can also be achieved by using a pre-trained model, inputting the pose data into the model, and outputting the location region through the model.

[0050] The aforementioned first and second size thresholds are used to divide the location region into blocks. The first size threshold represents the maximum size value of a block, and the second size threshold represents the minimum size value of a block; alternatively, the second size threshold represents the threshold for dividing the initial block. Dividing the location region using the first and second size thresholds ensures that the storage of the resulting multiple target blocks is limited to a reasonable range.

[0051] In some implementations, the first and second size thresholds can be obtained directly from pose data. For example, if a position region is calculated from pose data, the first size threshold can be the product of the position region and a first ratio, and the second size threshold can be the product of the position region and a second ratio, where the first ratio is greater than the second ratio.

[0052] The above-mentioned division of the location region based on the first size threshold yields multiple initial blocks, each of which is based on the first size threshold; while the division of the multiple initial blocks based on the second size threshold can be further divided into blocks based on the second size threshold, thereby further limiting the size of the blocks.

[0053] In some implementations, the plurality of initial blocks are divided based on a second size threshold. This can be achieved by randomly selecting an initial block from the plurality of initial blocks and then further dividing the randomly selected initial block using the second size threshold.

[0054] For example, the overall scale and spatial distribution characteristics of the scene are estimated based on the sparse point cloud range and pose data, and the initial block size is adaptively determined by combining the number of images. At the same time, maximum and minimum block thresholds (i.e., the first size threshold and the second size threshold) are set. On this basis, a primary spatial partitioning is performed: regions exceeding the preset threshold are recursively subdivided, and blocks are formed in the order of priority of the first principal axis and second principal axis; regions below the threshold are left unsubdivided.

[0055] In this embodiment of the invention, the location region corresponding to the first scene is determined based on the pose data; the location region is divided based on a first size threshold to obtain multiple initial blocks; the multiple initial blocks are divided based on a second size threshold to obtain multiple target blocks, wherein the second size threshold is smaller than the first size threshold.

[0056] In one embodiment, the image data is multi-view image data, the plurality of initial blocks include a first block, and the division of the plurality of initial blocks based on a second size threshold includes: Determine a first quantity corresponding to the first block, wherein the first quantity is used to represent the number of image data in the multi-view image data that contain the location of the first block; If the first quantity corresponding to the first block is greater than a set quantity threshold, and the size of the first block is greater than the second size threshold, the first block is divided.

[0057] The aforementioned first quantity is used to characterize the number of image data in the multi-view image data where the first block is located. The larger the first quantity, the easier the initial block is to be observed, the higher its importance, and the more it needs to be further divided to improve the rendering accuracy of the model. Conversely, the smaller the first quantity, the less likely the initial block is to be observed, the lower its importance, and the less it needs to be divided.

[0058] After the first block is divided, the divided blocks can be further analyzed to determine whether to divide them again, until the first number is less than the set number threshold, or the size of the block is less than the second size threshold. Then, no further division is required, and the block obtained at this time is the final target block.

[0059] Specifically, the first block is divided when the first quantity corresponding to the first block is greater than a set quantity threshold and the size of the first block is greater than a second size threshold; and the first block is not divided when the first quantity corresponding to the first block is less than a set quantity threshold or the size of the first block is less than a second size threshold.

[0060] For example, if the multi-view image data consists of images from 5 perspectives, and a threshold of 3 is set, and the first block is larger than a second size threshold, the first number determines whether further subdivision of the first block is needed. If 4 out of the 5 perspectives include the first block, then the first block is subdivided; if 2 out of the 5 perspectives include the first block, then the first block is not subdivided.

[0061] In one embodiment, the image data includes a depth map, and the method further includes: Generate a first mask, which is used to cover a preset object in the image data; Depth distance is determined based on the depth map, and a second mask is generated based on the depth map. The second mask is used to cover areas where the depth distance is greater than a first set distance. The rendering depth of the target model is determined, and a third mask is generated based on the rendering depth. The third mask is used to cover the region where the rendering depth is greater than a second set distance. The target model is adjusted based on the first mask, the second mask, and the third mask.

[0062] It should be noted that some objects may not be displayed in the trained target model, such as people or buildings that require privacy protection, as well as other areas that do not need to be displayed. Therefore, it is necessary to adjust the target model by generating a mask so that the areas that cannot be displayed can be covered up by the mask.

[0063] The aforementioned first mask is a mask used to cover a preset object in the image data. The preset object can be a person or an object requiring privacy protection, such as a building. In some implementations, the first mask can be generated directly from the image data, and the location of the first mask is the location of the preset object.

[0064] The second mask mentioned above is a mask for regions whose depth distance is greater than the first preset distance. It should be understood that when the depth distance of a region is greater than the first preset distance, its image data stability is poor. Therefore, a second mask can be generated for this region to reduce the resources consumed by model training and rendering.

[0065] The third mask mentioned above is for regions where the rendering depth is greater than the second set distance. It should be understood that when the rendering depth is greater than the second set distance, users will not pay much attention to this region when observing the rendering results of the model (such as the region where the sky is located). Therefore, a second mask can be generated for this region to reduce the resources consumed by model training and rendering.

[0066] Specifically, firstly, inverse depth normalization is performed based on the input depth map to remove unstable depth regions at long distances, resulting in a second mask. In outdoor scenes, a depth threshold is estimated based on a 3D Gaussian splash model of the sky, and combined with the rendering depth, regions exceeding the threshold are selected as candidate sky regions, resulting in a third mask. Finally, the first, second, and third masks are aligned and fused in a multi-scale space to generate a unified and effective ground observation mask. This mask can be used for depth, normal, and geometric consistency supervision, and the target model is then adjusted using the effective ground observation mask.

[0067] In some implementations, a single ground 3DGS representation is constructed in indoor scenes; in outdoor scenes, two types of 3D Gaussian splash models, one for the sky and one for the ground, are constructed separately to achieve decoupled representation and independent optimization of different areas.

[0068] In one embodiment, the plurality of target blocks includes a second block, and the method further includes: Back-projecting the depth map yields multiple initial point clouds corresponding to the second block, wherein the depth map is the depth map included in the image data or the depth map obtained by 3DGS rendering; By filtering the multiple initial point clouds, at least one first point cloud is obtained; The second model corresponding to the second block is updated based on the at least one first point cloud; The filtering of the plurality of initial point clouds includes at least one of the following: Delete the initial point clouds that are masked in the plurality of initial point clouds; Delete the initial point cloud whose depth distance is greater than a third preset distance from the plurality of initial point clouds; Random sampling is performed on the multiple initial point clouds; Delete the initial point clouds whose distance from the center of the Gaussian point is greater than a fourth predetermined distance. The center of the Gaussian point is the center of the multiple first Gaussian points included in the second model corresponding to the second block.

[0069] The aforementioned initial point clouds are obtained directly from depth maps. Since there are a large number of point clouds at this stage, training the model requires significant computational resources. However, in the subsequent actual rendering process, some initial point clouds will not be rendered (such as point clouds masked by a mask). To reduce computational resource consumption, it is necessary to filter the initial point clouds to maintain rendering quality while minimizing resource consumption.

[0070] Specifically, in this embodiment, the multiple initial point clouds are filtered, including at least one of the following: Delete the initial point clouds that are masked from the plurality of initial point clouds. It should be understood that since the areas masked are not rendered, deleting the initial point clouds that are masked from the plurality of initial point clouds will not affect the rendering effect, and can reduce the consumption of computing resources.

[0071] Delete the initial point clouds whose depth distance is greater than a third predetermined distance from the plurality of initial point clouds. It should be noted that, since the image data in the region with a depth distance greater than the third predetermined distance is less stable, and users will not pay much attention to this region when observing the rendering results of the model, the initial point clouds with a depth distance greater than the third predetermined distance from the plurality of initial point clouds can be deleted.

[0072] Random sampling is performed on the multiple initial point clouds. This random sampling can be based on pre-configured rules, such as selecting two initial point clouds from three, thereby reducing the number of initial point clouds in batches and thus reducing the consumption of computing resources.

[0073] The initial point clouds with a distance greater than a fourth predetermined distance from the center of the Gaussian point are deleted. The center of the Gaussian point is the center of multiple first Gaussian points included in the second model corresponding to the second block. It should be noted that the center of the Gaussian point is the area of ​​interest for the user during rendering. The farther away from the center of the Gaussian point, the lower the importance of the initial point cloud. Therefore, in this embodiment, a fourth predetermined distance is introduced to delete the initial point clouds with a distance greater than the fourth predetermined distance from the center of the Gaussian point, thereby reducing the consumption of computing resources without affecting the rendering effect.

[0074] In this embodiment of the invention, a backprojection is performed on the depth map to obtain multiple initial point clouds corresponding to the second block. The depth map is either a depth map included in the image data or a depth map rendered by 3DGS. The multiple initial point clouds are then filtered to obtain at least one first point cloud. The second model corresponding to the second block is updated based on the at least one first point cloud. In this way, by filtering the initial point clouds, the number of point clouds in the second model is reduced, thereby reducing the computational resources required to train the second model.

[0075] In one embodiment, the method further includes: Based on the image data, a score is calculated for each of the plurality of first Gaussian points, and the score is used to characterize the visibility of the corresponding first Gaussian point from multiple perspectives. Multiple second Gaussian points are selected from the plurality of first Gaussian points, wherein the scores corresponding to the plurality of second Gaussian points are greater than the scores corresponding to other first Gaussian points, and the other first Gaussian points are higher than the plurality of first Gaussian points excluding the plurality of second Gaussian points; Multiple third Gaussian points are selected from the plurality of second Gaussian points, wherein the scores corresponding to the plurality of third Gaussian points are greater than the scores corresponding to other second Gaussian points, and the other second Gaussian points are higher than the plurality of second Gaussian points other than the plurality of third Gaussian points; Copy the multiple third Gaussian points; The second model corresponding to the second block is updated based on the plurality of second Gaussian points and the plurality of copied third Gaussian points.

[0076] The aforementioned first Gaussian points are Gaussian points in the second model, used for image rendering. The scores represent the visibility of each Gaussian point from multiple viewpoints; higher scores indicate that the first Gaussian point is easily observed from different viewpoints, signifying its importance; lower scores indicate that the first Gaussian point can only be observed from a few viewpoints, signifying its importance. By calculating the scores of the first Gaussian points, it is determined whether to delete or duplicate them, further improving the model's image rendering capabilities.

[0077] In some embodiments, multiple second Gaussian points can be selected from multiple first Gaussian points, and the scores of multiple second Gaussian points are higher than the scores of other first Gaussian points. In this case, only multiple second Gaussian points are retained, so as to maintain good rendering capabilities while reducing the consumption of computing resources.

[0078] Furthermore, among the multiple second Gaussian points, there are also multiple third Gaussian points. The scores of the multiple third Gaussian points are higher, and their importance is greater. By replicating the multiple third Gaussian points, the rendering capability of the model can be improved.

[0079] In some implementations, a score can be calculated using multi-view image data. First, the visibility count and consistency parameter of each first Gaussian point are obtained. The visibility count is the number of image data in the multi-view image data where the first Gaussian point can be observed, and the consistency parameter is the consistency assessment of whether the first Gaussian point is visible in the multi-view image data. The score is calculated by weighting the visibility count and the consistency parameter.

[0080] In one embodiment, the method includes: Construct conditional constraints corresponding to each target block, wherein the conditional constraints include at least one of planar consistency constraints, lighting constraints, and geometric constraints; The step of training a first model corresponding to each target block based on the first data corresponding to each target block to obtain a second model corresponding to each target block includes: Based on the first data and condition constraints corresponding to each target block, the first model corresponding to each target block is trained to obtain the second model corresponding to each target block.

[0081] The aforementioned planar consistency constraints are used to calculate the cosine consistency of normals based on the normal map to detect planar structures and generate a planar mask. A reliable planar region is obtained by combining this with an adaptive effective mask. Within this region, the rendered normals are monitored pixel-by-pixel, and minimum scale constraints and covariance principal axis alignment constraints are applied to the Gaussians projected onto the plane, thereby improving the geometric consistency and stability of the planar structure. Furthermore, Gaussian primitives are optimized by minimizing the pixel-level error between the rendered image and the input image data to improve the appearance consistency of the reconstruction results.

[0082] The aforementioned lighting constraints are constructed based on the appearance differences between the rendered result and the input image. The model is supervised and optimized by minimizing the reconstruction error between the rendered image and the real input image, so that it learns the relationship between viewpoint-related lighting changes and appearance mapping, thereby enhancing the model's adaptability to complex lighting conditions and improving the appearance consistency and robustness of the reconstruction results.

[0083] The aforementioned geometric constraints are used to constrain the block partitioning. Specifically, 3D Gaussian elements are jointly sampled according to importance scores and opacity, and a tetrahedral mesh structure is constructed using a spatial partitioning strategy to obtain a geometric representation with good spatial coverage. Subsequently, multi-view TSDF information is fused to apply cross-view geometric consistency constraints to the tetrahedral mesh to alleviate the uncertainty caused by single-view observation. Furthermore, a differentiable tetrahedral traversal algorithm is used to extract continuous surfaces from the geometric field, and the obtained surface structure is jointly optimized with the image-based rendering loss. Gaussian element parameters and geometric representations are updated synchronously through an end-to-end backpropagation mechanism, thereby significantly improving geometric accuracy and structural consistency while ensuring rendering quality.

[0084] In this embodiment of the invention, conditional constraints are constructed for each target block, including at least one of planar consistency constraints, lighting constraints, and geometric constraints. Based on the first data and conditional constraints corresponding to each target block, a first model corresponding to each target block is trained to obtain a second model corresponding to each target block. Thus, by constraining the training process of the second model corresponding to each target block through conditional constraints, the second model can satisfy at least one of the planar consistency constraints, lighting constraints, and geometric constraints during rendering, thereby effectively improving the rendering effect of the model.

[0085] In one embodiment, training the first model corresponding to each target block based on the first data corresponding to each target block includes: Based on a preset strategy, the first model corresponding to each target block is trained using the first data corresponding to each target block. The preset strategy includes at least one of the following: The caching acceleration strategy is to read only the model parameters when rendering the depth map and Gaussian encryption process, and to read only the image data and pose data when deleting point clouds and training with constraints. A multi-level resolution training strategy, wherein the multi-level resolution training strategy is a strategy of sequentially increasing the resolution of the image data used to train the model; The decoupling strategy is a strategy of training different models for corresponding areas on the ground and / or in the sky.

[0086] Understandably, at different training stages, the data reading and caching strategies for images, depth, normals, and model parameters can be adaptively adjusted based on computational needs and data dependencies to reduce data access overhead and improve training efficiency.

[0087] Specifically, in the Gaussian densification process based on the rendered depth map (i.e., the process of copying the third Gaussian point), only the model parameters are read and cached to reduce unnecessary image data access. In the 3DGS adaptive point deletion and densification process (i.e., the process of filtering the initial point cloud and filtering the first Gaussian point) and the geometric optimization process based on tetrahedral mesh (i.e., adjusting the blocks based on conditional constraints), only image data and camera pose parameters are used for calculation, without loading parameters such as depth maps and normal maps, thereby further reducing memory usage and data transmission overhead, and achieving efficient resource scheduling in different training stages.

[0088] The aforementioned multi-level resolution training strategy involves progressively increasing the resolution of the image data used to train the model, thereby improving the accuracy of the trained model.

[0089] Specifically, the above decoupling strategy can be to construct a single ground 3DGS representation in indoor scenes and construct two types of 3D Gaussian splash models, one for the sky and one for the ground, in outdoor scenes, to achieve decoupled representation and independent optimization of different areas.

[0090] The overall process of this invention is as follows Figure 2 As shown, firstly, image data and pose data are acquired, a first model for the first scene is trained, and multiple target blocks are obtained; then, a second model for each target block is trained, including setting data reading and caching strategies, decoupling the sky and ground 3DGS, Gaussian densification of the Gaussian points in the 3DGS, and constructing conditional constraints to obtain the second model for each target block; then, the second models of multiple target blocks are fused to obtain the final target model.

[0091] Please see Figure 3 , Figure 3This is a structural diagram of a three-dimensional model building device provided in an embodiment of the present invention, such as... Figure 3 As shown, the 3D model building device 300 includes: The acquisition module 301 is used to acquire image data of the first scene and pose data of the captured image data; The first training module 302 is used to train an initial three-dimensional model based on the image data, wherein the initial three-dimensional model is a model with a first precision. The first determining module 303 is used to determine the location region corresponding to the first scene based on the pose data, as well as a first size threshold and a second size threshold, wherein the second size threshold is less than the first size threshold. The first division module 304 is used to divide the location region based on the first size threshold to obtain multiple initial blocks; The second partitioning module 305 is used to partition the plurality of initial blocks based on the second size threshold to obtain a plurality of target blocks; Extension module 306 is used to expand the region of each target block; The second determining module 307 is used to determine the first model corresponding to each expanded target block based on the initial three-dimensional model, and to determine the first data corresponding to each expanded target block based on the image data. The first model is the model of the first precision corresponding to the corresponding expanded target block, and the first data is the data of the region where the corresponding expanded target block is located in the image data. The second training module 308 is used to train the first model corresponding to each target block based on the first data corresponding to each target block to obtain the second model corresponding to each target block. The second model corresponding to each target block is a model with a second precision, and the second precision is higher than the first precision. The first generation module 309 is used to generate a target model based on the second model corresponding to the plurality of target blocks, wherein the target model is the model of the second precision corresponding to the first scene.

[0092] In one embodiment, the image data is multi-view image data, the plurality of initial blocks include a first block, and the second segmentation module 305 includes: A determining unit is used to determine a first quantity corresponding to the first block, wherein the first quantity is the number of image data in the multi-view image data that contains the location of the first block; The partitioning unit is used to partition the first block when the first quantity corresponding to the first block is greater than a set quantity threshold and the size of the first block is greater than a second size threshold.

[0093] In one embodiment, the image data includes a depth map, and the 3D model building device 300 further includes: The second generation module is used to generate a first mask, which is used to cover a preset object in the image data; The second determining module is used to determine the depth distance based on the depth map and generate a second mask based on the depth map. The second mask is used to cover the area where the depth distance is greater than a first set distance. The second determining module is used to determine the rendering depth of the target model and generate a third mask based on the rendering depth. The third mask is used to cover the area where the rendering depth is greater than a second set distance. The adjustment module is used to adjust the target model based on the first mask, the second mask, and the third mask.

[0094] In one embodiment, the plurality of target blocks includes a second block, and the 3D model construction device 300 further includes: The projection module is used to back-project the depth map to obtain multiple initial point clouds corresponding to the second block. The depth map is the depth map included in the image data or the depth map obtained by 3DGS rendering. The first filtering module is used to filter the plurality of initial point clouds to obtain at least one first point cloud; The first update module is used to update the second model corresponding to the second block based on the at least one first point cloud; The filtering of the plurality of initial point clouds includes at least one of the following: Delete the initial point clouds that are masked in the plurality of initial point clouds; Delete the initial point cloud whose depth distance is greater than a third preset distance from the plurality of initial point clouds; Random sampling is performed on the multiple initial point clouds; Delete the initial point clouds whose distance from the center of the Gaussian point is greater than a fourth predetermined distance. The center of the Gaussian point is the center of the multiple first Gaussian points included in the second model corresponding to the second block.

[0095] In one embodiment, the three-dimensional model building device 300 further includes: The scoring module is used to calculate a score for each of the plurality of first Gaussian points based on the image data. The score is used to characterize the visibility of the corresponding first Gaussian point from multiple perspectives. The second filtering module is used to filter multiple second Gaussian points from the plurality of first Gaussian points, wherein the scores corresponding to the plurality of second Gaussian points are greater than the scores corresponding to other first Gaussian points, and the other first Gaussian points are higher than the plurality of first Gaussian points excluding the plurality of second Gaussian points. The third filtering module is used to filter multiple third Gaussian points from the multiple second Gaussian points, wherein the scores corresponding to the multiple third Gaussian points are greater than the scores corresponding to other second Gaussian points, and the other second Gaussian points are higher than the multiple second Gaussian points other than the multiple third Gaussian points. The copy module is used to copy the plurality of third Gaussian points; The second update module is used to update the second model corresponding to the second block based on the plurality of second Gaussian points and the replicated plurality of third Gaussian points.

[0096] In one embodiment, the three-dimensional model building device 300 includes: A construction module is used to construct the condition constraints corresponding to each target block, wherein the condition constraints include at least one of planar consistency constraints, lighting constraints, and geometric constraints; The second training module 308 includes: The training submodule is used to train the first model corresponding to each target block based on the first data and condition constraints corresponding to each target block, so as to obtain the second model corresponding to each target block.

[0097] The three-dimensional model building apparatus provided in this embodiment of the invention can realize each process of each embodiment of the above-described three-dimensional model building method. The technical features are one-to-one and can achieve the same technical effect. To avoid repetition, it will not be described again here.

[0098] It should be noted that the three-dimensional model building device in the embodiments of the present invention can be a device, or it can be a component, integrated circuit, or chip in an electronic device.

[0099] This invention also provides an electronic device, see [link to relevant documentation]. Figure 4 , Figure 4 This is a schematic diagram of the structure of an electronic device provided by an embodiment of the present invention. The electronic device includes a memory 401, a processor 402, and a program or instructions stored in the memory 401 that run on the memory. When the program or instructions are executed by the processor 402, they can achieve the following: Figure 1 The steps in the corresponding three-dimensional model construction method embodiments and the achievement of the same beneficial effects will not be elaborated here.

[0100] The processor can be a central processing unit (CPU), an application-specific integrated circuit (ASIC), a field-programmable gate array (FPGA), or a graphics processing unit (GPU).

[0101] This invention also provides a computer-readable storage medium storing a computer program that, when executed by a processor, performs the above-described functions. Figure 1 The various processes corresponding to the embodiments of the 3D model construction method achieve the same technical effect, and will not be described again here to avoid repetition. The computer-readable storage medium mentioned includes, for example, read-only memory (ROM), random access memory (RAM), magnetic disk, or optical disk.

[0102] The present invention also provides a computer program product, including computer instructions that, when executed by a processor, implement the above-described... Figure 1 The various processes corresponding to the three-dimensional model construction method embodiments can achieve the same technical effect, and will not be described again here to avoid repetition.

[0103] In the embodiments of this invention, the terms "first," "second," etc., are used to distinguish similar object parameters and are not necessarily used to describe a specific order or sequence. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion. For example, a process, method, system, product, or device that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to these processes, methods, products, or devices. Additionally, the use of "and / or" in this application indicates at least one of the connected object parameters, such as A and / or B and / or C, representing seven possibilities: A alone, B alone, C alone, both A and B present, both B and C present, both A and C present, and A, B, and C present.

[0104] It should be noted that, in this document, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes that element.

[0105] Through the above description of the embodiments, those skilled in the art can clearly understand that the methods of the above embodiments can be implemented by means of software plus necessary general-purpose hardware platforms. Of course, they can also be implemented by hardware, but in many cases the former is a better implementation method. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product is stored in a storage medium (such as ROM / RAM, magnetic disk, optical disk) and includes several instructions to cause a terminal (which may be a mobile phone, computer, server, air conditioner, etc.) to execute the methods of the various embodiments of this application.

[0106] The embodiments of this application have been described above with reference to the accompanying drawings. However, this application is not limited to the specific embodiments described above. The specific embodiments described above are merely illustrative and not restrictive. Those skilled in the art can make many other forms under the guidance of this application without departing from the spirit and scope of the claims, and all of these forms are within the protection scope of this application.

Claims

1. A three-dimensional model construction method characterized by, include: Acquire image data of the first scene and pose data of the captured image data; An initial 3D model is trained based on the image data, and the initial 3D model is a model with first-precision accuracy; Based on the pose data, the location region corresponding to the first scene is determined, as well as a first size threshold and a second size threshold, wherein the second size threshold is less than the first size threshold; The location region is divided based on the first size threshold to obtain multiple initial blocks; Based on the second size threshold, the plurality of initial blocks are divided to obtain a plurality of target blocks; Expand the region of each target block; Based on the initial 3D model, a first model corresponding to each expanded target block is determined, and based on the image data, a first data corresponding to each expanded target block is determined. The first model is the model of the first precision corresponding to the corresponding expanded target block, and the first data is the data of the region where the corresponding expanded target block is located in the image data. Based on the first data corresponding to each target block, the first model corresponding to each target block is trained to obtain the second model corresponding to each target block. The second model corresponding to each target block is a model with second precision, which is higher than the first precision. A target model is generated based on the second model corresponding to the multiple target blocks, and the target model is the second precision model corresponding to the first scene; The image data is multi-view image data, and the multiple initial blocks include a first block. The process of dividing the multiple initial blocks based on the second size threshold includes: Determine a first quantity corresponding to the first block, wherein the first quantity is used to represent the number of image data in the multi-view image data that contain the location of the first block; If the first quantity corresponding to the first block is greater than a set quantity threshold, and the size of the first block is greater than the second size threshold, the first block is divided.

2. The method as described in claim 1, characterized in that, The image data includes a depth map, and the method further includes: Generate a first mask, which is used to cover a preset object in the image data; Depth distance is determined based on the depth map, and a second mask is generated based on the depth map. The second mask is used to cover areas where the depth distance is greater than a first set distance. The rendering depth of the target model is determined, and a third mask is generated based on the rendering depth. The third mask is used to cover the region where the rendering depth is greater than a second set distance. The target model is adjusted based on the first mask, the second mask, and the third mask.

3. The method as described in claim 1, characterized in that, The plurality of target blocks includes a second block, and the method further includes: Back-projecting the depth map yields multiple initial point clouds corresponding to the second block, wherein the depth map is the depth map included in the image data or the depth map obtained by 3DGS rendering; By filtering the multiple initial point clouds, at least one first point cloud is obtained; The second model corresponding to the second block is updated based on the at least one first point cloud; The filtering of the plurality of initial point clouds includes at least one of the following: Delete the initial point clouds that are masked in the plurality of initial point clouds; Delete the initial point cloud whose depth distance is greater than a third preset distance from the plurality of initial point clouds; Random sampling is performed on the multiple initial point clouds; Delete the initial point clouds whose distance from the center of the Gaussian point is greater than a fourth predetermined distance. The center of the Gaussian point is the center of the multiple first Gaussian points included in the second model corresponding to the second block.

4. The method as described in claim 3, characterized in that, The method further includes: Based on the image data, a score is calculated for each of the plurality of first Gaussian points, and the score is used to characterize the visibility of the corresponding first Gaussian point from multiple perspectives. Multiple second Gaussian points are selected from the plurality of first Gaussian points, wherein the scores corresponding to the plurality of second Gaussian points are greater than the scores corresponding to other first Gaussian points, and the other first Gaussian points are higher than the plurality of first Gaussian points excluding the plurality of second Gaussian points; Multiple third Gaussian points are selected from the plurality of second Gaussian points, wherein the scores corresponding to the plurality of third Gaussian points are greater than the scores corresponding to other second Gaussian points, and the other second Gaussian points are higher than the plurality of second Gaussian points excluding the plurality of third Gaussian points; Copy the multiple third Gaussian points; The second model corresponding to the second block is updated based on the plurality of second Gaussian points and the plurality of copied third Gaussian points.

5. The method according to any one of claims 1 to 4, characterized in that, The method includes: Construct conditional constraints corresponding to each target block, wherein the conditional constraints include at least one of planar consistency constraints, lighting constraints, and geometric constraints; The step of training a first model corresponding to each target block based on the first data corresponding to each target block to obtain a second model corresponding to each target block includes: Based on the first data and condition constraints corresponding to each target block, the first model corresponding to each target block is trained to obtain the second model corresponding to each target block.

6. The method according to any one of claims 1 to 4, characterized in that, The step of training the first model corresponding to each target block based on the first data corresponding to each target block includes: Based on a preset strategy, the first model corresponding to each target block is trained using the first data corresponding to each target block. The preset strategy includes at least one of the following: The caching acceleration strategy is to read only the model parameters when rendering the depth map and Gaussian encryption process, and to read only the image data and pose data when deleting point clouds and training with constraints. A multi-level resolution training strategy, wherein the multi-level resolution training strategy is a strategy of sequentially increasing the resolution of the image data used to train the model; The decoupling strategy is a strategy of training different models for corresponding areas on the ground and / or in the sky.

7. A three-dimensional model construction device, characterized in that, include: The acquisition module is used to acquire image data of the first scene and pose data of the captured image data; The first training module is used to train an initial three-dimensional model based on the image data, wherein the initial three-dimensional model is a model with a first precision. The first determining module is used to determine the location region corresponding to the first scene based on the pose data, as well as a first size threshold and a second size threshold, wherein the second size threshold is less than the first size threshold. The first partitioning module is used to partition the location region based on the first size threshold to obtain multiple initial blocks; The second partitioning module is used to partition the plurality of initial blocks based on the second size threshold to obtain a plurality of target blocks; An extension module is used to expand the region of each target block; The second determining module is used to determine the first model corresponding to each of the expanded target blocks based on the initial three-dimensional model, and to determine the first data corresponding to each of the expanded target blocks based on the image data. The first model is the model of the first precision corresponding to the corresponding expanded target block, and the first data is the data of the region where the corresponding expanded target block is located in the image data. The second training module is used to train the first model corresponding to each target block based on the first data corresponding to each target block, so as to obtain the second model corresponding to each target block. The second model corresponding to each target block is a model with a second precision, which is higher than the first precision. The first generation module is used to generate a target model based on the second model corresponding to the multiple target blocks, wherein the target model is the second precision model corresponding to the first scene.

8. An electronic device, characterized in that, include: A processor, a memory, and a program stored in the memory and executable on the processor, wherein the program, when executed by the processor, implements the steps of the three-dimensional model construction method as described in any one of claims 1 to 6.

9. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that, when executed by a processor, implements the steps of the three-dimensional model construction method as described in any one of claims 1 to 6.

10. A computer program product, characterized in that, It includes computer instructions that, when executed by a processor, implement the steps of the three-dimensional model construction method as described in any one of claims 1 to 6.

Images