High-quality large scene reconstruction method based on 3D Gaussian sputtering

Abstract

The application provides a high-quality large scene reconstruction method based on 3D Gaussian sputtering, comprising the following steps: dividing a large scene into multiple grid units; using a point cloud adaptive interpolation module to refine point cloud points within the boundary of each grid unit; using a global-local Gaussian decoder to calculate 3D Gaussian parameters within and outside the boundary of each grid unit; using a Gaussian refinement module to perform refinement processing on the 3D Gaussian distribution; and using a weighted average method to combine the 3D Gaussian distributions of all grid units into a complete large scene 3D Gaussian distribution, and finally generating a complete large scene three-dimensional model. The high-quality large scene reconstruction method can improve the consistency and quality of large scene reconstruction.

Description

Technical Field

[0001] This invention relates to the field of 3D Gaussian reconstruction, and in particular to a high-quality large-scene reconstruction method based on 3D Gaussian sputtering. Background Technology

[0002] Large-scale 3D scene reconstruction is crucial for fields such as autonomous driving, virtual reality, environmental monitoring, and aerial surveying. 3D Gaussian sputtering (3DGS) has gained attention for its high reconstruction quality and fast rendering speed. With the rapid development of 3DGS, several methods have been implemented for large-scale 3D scene reconstruction. These methods typically employ a divide-and-conquer strategy, dividing the large scene into multiple independent blocks, processing each individual block, and finally merging the processed blocks. While these methods effectively solve the problem of large-scale 3D scene reconstruction, processing each block independently ignores the relationships between them, leading to inconsistencies at block boundaries.

[0003] Therefore, proposing a high-quality large-scene reconstruction method based on 3D Gaussian sputtering to improve the consistency and quality of large-scene reconstruction has become an urgent problem to be solved. Summary of the Invention

[0004] In view of this, the present invention provides a high-quality large-scene reconstruction method based on 3D Gaussian sputtering to solve the problems existing in the prior art.

[0005] This invention provides a high-quality large-scene reconstruction method based on 3D Gaussian sputtering, comprising:

[0006] Multi-view large scene images are acquired and preprocessed to obtain preprocessed multi-view large scene images, in which there is overlap between images from adjacent views;

[0007] COLMAP is used to obtain the camera pose data and point cloud data corresponding to the preprocessed multi-view large scene images.

[0008] The scene's ground plane is divided into multiple grid cells, each with clearly defined boundaries and coordinate ranges.

[0009] Determine the contributing cameras for each grid cell, wherein, for any given grid cell, select the group of cameras that contribute the most to it according to the visibility criterion as the contributing cameras for that grid cell;

[0010] Extend the point cloud points within the boundary of each grid cell outward to obtain the extended point cloud data outside the boundary of each grid cell;

[0011] The point cloud adaptive interpolation module is used to refine the point cloud points within the boundary of each grid cell, resulting in a more refined point cloud representation.

[0012] The refined point cloud data within the boundary of each grid cell and the extended point cloud data outside the boundary of each grid cell are segmented using a pre-trained point cloud segmentation model to obtain segmented point cloud data.

[0013] Based on the segmented point cloud data, the 3D Gaussian parameters inside and outside the boundary of each grid cell are calculated using a global-local Gaussian decoder, resulting in the 3D Gaussian distribution inside and outside the boundary of each grid cell.

[0014] The Gaussian refinement module is used to dynamically adjust the 3D Gaussian distribution within the boundary of each mesh cell. The Gaussian refinement module is used to dynamically allocate the number of Gaussians according to the local geometric complexity, allocate more Gaussians to areas with complex geometric details, and delete redundant Gaussians.

[0015] Remove the 3D Gaussian distribution outside the boundary of each grid cell to obtain the optimized 3D Gaussian distribution for each grid cell;

[0016] The weights of all adjacent grid cells are determined based on the contribution of each grid cell camera, and a weighted average is performed on the weights of all adjacent grid cells and the optimized 3D Gaussian distribution parameters to obtain the 3D Gaussian distribution of the merged complete scene.

[0017] A 3D model is generated based on the 3D Gaussian distribution of the complete scene using 3D Gaussian rendering technology.

[0018] Preferably, the image preprocessing includes denoising, normalization, and cropping / scaling.

[0019] Further optimization involves determining the contributing camera for any grid cell A as follows:

[0020] The boundary of the grid cell A is projected onto the image plane of the camera, and the projected area of ​​the grid cell A in each camera view is calculated;

[0021] Calculate the total area of ​​the camera image based on the total number of pixels obtained from the camera resolution.

[0022] Calculate the ratio of the projected area of ​​grid cell A in each camera viewpoint to the total area of ​​the camera image;

[0023] Select the cameras with the highest proportions as contributing cameras for the grid cell A.

[0024] Further optimization involves refining the point cloud points within the boundary of each grid cell using a point cloud adaptive interpolation module, as follows:

[0025] Create two empty sets: one to store the indices of the points that have already been processed, and the other to store the points generated after interpolation.

[0026] Iterate through each point in the initial point cloud and perform interpolation. The interpolation method for any point is as follows:

[0027] Find the K nearest neighbors of the current point;

[0028] Based on the K nearest neighbors, an incremental algorithm is used to partition the 3D Voronoi polygons, and the vertices of the Voronoi polygons are the potential interpolation points.

[0029] The Wasserstein distance is used to evaluate the topological difference between the set of K nearest neighbors and the set containing the vertices of the Voronoi polygon. If the evaluation results show that these vertices do not destroy the topology, the vertices of the Voronoi polygon are added to the set as interpolation points; otherwise, 2D Voronoi interpolation is performed.

[0030] The points generated by interpolation are merged with the original point cloud to obtain a refined point cloud.

[0031] Further optimization involves determining the number of K nearest neighbors of the current point based on the sparsity of the point cloud. The higher the curvature of the region, the more nearest neighbors are selected.

[0032] Further optimization involves the following specific steps for 2D Voronoi interpolation:

[0033] Principal component analysis is used to project the original K nearest neighbor points onto a two-dimensional plane;

[0034] Voronoi partitioning is performed on a two-dimensional plane using an incremental algorithm to obtain Voronoi polygons.

[0035] Map the vertices of the 2D Voronoi polygon back to 3D space and add the mapped vertices as interpolation points to the set.

[0036] Further optimization involves obtaining the 3D Gaussian distribution inside and outside the boundary of each grid cell, with the following specific steps:

[0037] The local Gaussian decoder is used to predict the 3D Gaussian parameters of each grid cell, and the image rendered using the predicted 3D Gaussian parameters is compared with the real image to calculate the reconstruction loss. The parameters of the local Gaussian decoder are then updated based on the reconstruction loss.

[0038] Update the parameters of the global Gaussian decoder based on the local Gaussian decoders of all mesh cells after updating the parameters;

[0039] The global Gaussian decoder is used to predict the 3D Gaussian parameters of each grid cell. Then, the difference between the Gaussian parameters predicted by the local Gaussian decoder and the global Gaussian decoder for each grid cell is calculated. The parameters of the local Gaussian decoder are then updated based on the difference to make them closer to the prediction of the global Gaussian decoder.

[0040] The updated local Gaussian decoder is used to re-predict the 3D Gaussian parameters of each grid cell, resulting in the 3D Gaussian distribution inside and outside the boundary of each grid cell.

[0041] Further optimization involves using the Gaussian refinement module to dynamically adjust the 3D Gaussian distribution within the boundary of each mesh cell, as follows:

[0042] Using a keypoint scorer, a keypoint score map is calculated from image features for each input view. The keypoint score map is used to reflect the geometric complexity of different regions in the image and can indicate the importance of each region in the image.

[0043] In the key point score map, for areas with high scores, the Gaussian center is further subdivided into multiple smaller centers, and for areas with low scores, the transparency and scaling of the corresponding Gaussian center are gradually reduced until it is completely removed.

[0044] Further preferred, the weight of adjacent grid cells is determined based on the number of cameras shared between the adjacent grid cells. The more cameras shared between adjacent grid cells, the higher their weight. If camera A is included in the contributing cameras of two adjacent grid cells, then camera A is the shared camera of these two adjacent grid cells.

[0045] This invention provides a high-quality large-scene reconstruction method based on 3D Gaussian sputtering, which can improve the consistency and quality of large-scene reconstruction. The method first divides the large scene into multiple grid cells. Then, an adaptive point cloud interpolation module is used to refine the point cloud points within the boundary of each grid cell. This module addresses the issue of sparse point clouds in low-curvature regions while maintaining the scene's topology. A global-local Gaussian decoder is used to calculate the 3D Gaussian parameters within and outside the boundary of each grid cell. This ensures consistent Gaussian parameter predictions across different grid cells, effectively solving the consistency problem between individual grid cells. A Gaussian refinement module further refines the 3D Gaussian distribution. Dynamically adjusting the Gaussian distribution based on the complexity of the region improves the quality of the 3D Gaussian distribution while reducing redundant information. Finally, a weighted average is used to merge the 3D Gaussian distributions of all grid cells into a complete large-scene 3D Gaussian distribution, ultimately generating a complete 3D model of the large scene. Attached Figure Description

[0046] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying 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.

[0047] Figure 1 This is a flowchart of the high-quality large-scene reconstruction method based on 3D Gaussian sputtering provided by the present invention. Detailed Implementation

[0048] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be described in detail below with reference to the accompanying drawings and specific embodiments. It should be noted that, to avoid obscuring the invention with unnecessary details, only processing steps closely related to the solution of this invention are shown in the drawings, while other details not closely related to this invention are omitted.

[0049] like Figure 1 As shown, this invention provides a high-quality large-scene reconstruction method based on 3D Gaussian sputtering, comprising the following steps:

[0050] S1: Acquire large-scene images from multiple perspectives and perform image preprocessing to obtain preprocessed large-scene images from multiple perspectives, where there is overlap between images from adjacent perspectives;

[0051] Among them, large-scene images from multiple perspectives can be acquired by drone cameras. Specifically, the drone is controlled to fly around at a constant speed above the center of the large scene, and the drone camera is controlled to capture scene images from multiple angles, ensuring sufficient overlap between adjacent perspective images.

[0052] The image preprocessing includes denoising, normalization, and cropping / scaling;

[0053] The denoising process is as follows: First, select an appropriate denoising algorithm (e.g., Gaussian filtering, median filtering, bilateral filtering) based on the type of noise in the image; then, set the parameters of the denoising algorithm (e.g., filter size, number of iterations) based on the noise level and resolution of the image; finally, use an image processing library (e.g., OpenCV or Pillow) to denoise the image.

[0054] The normalization process is as follows: The pixel values ​​of the denoised image are normalized to the range of [0,1] using an image processing library;

[0055] The cropping and scaling steps are as follows: Use an image processing library to crop and scale the normalized image to a fixed size, such as 256×256 pixels;

[0056] S2: Use COLMAP to obtain the camera pose data and point cloud data corresponding to the preprocessed multi-view large scene image;

[0057] S3: Divide the scene's ground plane into multiple grid units, for example, using m×n grids, where each grid unit has a defined boundary and coordinate range;

[0058] The specific steps for S3 are as follows:

[0059] S31: Select an appropriate grid partitioning scheme based on the size and complexity of the scene to ensure that the size of each grid cell is moderate, neither too large to affect the calculation nor too small to affect the efficiency;

[0060] S32: Divide the scene's ground plane into multiple grid units according to the selected grid division scheme. Each grid unit has a clear boundary and coordinate range.

[0061] S4: Determine the contributing cameras for each grid cell, wherein, for any given grid cell, select the group of cameras that contribute the most to it according to the visibility criterion as the contributing cameras for that grid cell.

[0062] The steps for determining the contributing camera for any grid cell A are as follows:

[0063] S41: Project the boundary of the grid cell A onto the image plane of the camera, and calculate the projected area of ​​the grid cell A in each camera view.

[0064] S42: Obtain the total number of pixels based on the camera resolution, and calculate the total area of ​​the camera image;

[0065] S43: Calculate the ratio of the projected area of ​​grid cell A to the total area of ​​the camera image in each camera viewpoint;

[0066] S44: Select the cameras with the highest proportion (the number can be adjusted according to the size of the scene) as the contributing cameras of the grid unit A;

[0067] S5: Expand the point cloud points within the boundary of each grid cell outwards by a certain proportion, such as 20%, to obtain the expanded point cloud data outside the boundary of each grid cell;

[0068] S6: Use the point cloud adaptive interpolation module to refine the point cloud points within the boundary of each grid cell to obtain a more refined point cloud representation. The point cloud adaptive interpolation module can solve the problem of sparse point clouds in low curvature areas while keeping the topology of the scene unchanged.

[0069] The specific steps for S6 are as follows:

[0070] S61: Create two empty sets, one to store the indices of the points that have been processed, and the other to store the points generated after interpolation;

[0071] S62: Traverse each point in the initial point cloud and perform interpolation. The interpolation method for any point is as follows:

[0072] S621: Find the K nearest neighbors of the current point. The selection of the K nearest neighbors needs to take into account the sparsity of the point cloud. Fewer nearest neighbors can be selected in low curvature regions, while more nearest neighbors are needed in high curvature regions.

[0073] S622: Use an incremental algorithm to partition the 3D Voronoi polygons based on the K nearest neighbors. The vertices of the Voronoi polygons are the potential interpolation points.

[0074] S623: Use Wasserstein distance to evaluate the topological difference between the set of K nearest neighbors and the set containing the vertices of the Voronoi polygon. If the evaluation results show that these vertices do not destroy the topology, add the vertices of the Voronoi polygon as interpolation points to the set; otherwise, perform 2D Voronoi interpolation.

[0075] The specific steps of 2D Voronoi interpolation are as follows:

[0076] S6231: Use Principal Component Analysis (PCA) to project the original K nearest neighbors onto a two-dimensional plane;

[0077] S6232: Use an incremental algorithm to perform Voronoi partitioning on a two-dimensional plane to obtain Voronoi polygons;

[0078] S6233: Map the vertices of a 2D Voronoi polygon back to 3D space and add the mapped vertices as interpolation points to the set;

[0079] S63: Merge the interpolated points with the original point cloud to obtain a refined point cloud;

[0080] S7: Use a pre-trained point cloud segmentation model to segment the refined point cloud data within the boundary of each grid cell and the extended point cloud data outside the boundary of each grid cell to obtain the segmented point cloud data;

[0081] S8: Based on the segmented point cloud data, the global-local Gaussian decoder is used to calculate the 3D Gaussian parameters inside and outside the boundary of each grid cell, and obtain the 3D Gaussian distribution inside and outside the boundary of each grid cell. The global-local Gaussian decoder can ensure that the Gaussian parameter predictions between different grid cells are consistent, thereby effectively solving the consistency problem between each grid cell.

[0082] The specific steps for S8 are as follows:

[0083] S81: Use a local Gaussian decoder to predict the 3D Gaussian parameters of each grid cell, and use the predicted 3D Gaussian parameters to render an image and compare it with the real image to calculate the reconstruction loss. Then update the parameters of the local Gaussian decoder based on the reconstruction loss.

[0084] S82: Update the parameters of the global Gaussian decoder based on the local Gaussian decoders of all mesh cells after parameter updates;

[0085] S83: Use a global Gaussian decoder to predict the 3D Gaussian parameters of each grid cell. Then, calculate the difference between the Gaussian parameters predicted by the local Gaussian decoder and the global Gaussian decoder for each grid cell. Update the parameters of the local Gaussian decoder according to the difference to make them closer to the prediction of the global Gaussian decoder.

[0086] S84: Re-predict the 3D Gaussian parameters of each grid cell using the updated local Gaussian decoder to obtain the 3D Gaussian distribution inside and outside the boundary of each grid cell;

[0087] S9: Use the Gaussian refinement module to dynamically adjust the 3D Gaussian distribution within the boundary of each mesh cell, wherein the Gaussian refinement module is used to dynamically allocate the number of Gaussians according to the local geometric complexity, allocate more Gaussians to areas with complex geometric details, and delete redundant Gaussians.

[0088] The specific steps for S9 are as follows:

[0089] S91: Using a keypoint scorer, calculate a keypoint score map for each input view from image features, wherein the keypoint score map is used to reflect the geometric complexity of different regions in the image and can indicate the importance of each region in the image;

[0090] S92: In the key point score map, for areas with high scores, the Gaussian center is further subdivided into multiple smaller centers to more accurately capture local geometric details and subtle changes in the image. For areas with low scores, the transparency and scaling of the corresponding Gaussian center are gradually reduced until it is completely removed. Removing redundant information from the Gaussian set can avoid Gaussian center overlap and improve the efficiency and accuracy of the model.

[0091] S10: Remove the 3D Gaussian distribution outside the boundary of each grid cell to obtain the optimized 3D Gaussian distribution of each grid cell, in order to prevent boundary artifacts and ensure the quality of the merged scene.

[0092] S11: Determine the weights of all adjacent grid cells based on the contributing camera of each grid cell, and perform a weighted average of the weights of all adjacent grid cells and the optimized 3D Gaussian distribution parameters to obtain the 3D Gaussian distribution of the merged complete scene.

[0093] The weight of adjacent grid cells is determined based on the number of cameras shared between the adjacent grid cells. The more cameras shared between adjacent grid cells, the higher their weight. If camera A is included in the contributing cameras of two adjacent grid cells, then camera A is the shared camera of these two adjacent grid cells.

[0094] S12: Use 3D Gaussian rendering technology to generate a three-dimensional model based on the 3D Gaussian distribution of the complete scene.

[0095] This invention provides a high-quality large-scene reconstruction method based on 3D Gaussian sputtering, which can improve the consistency and quality of large-scene reconstruction. The method first divides the large scene into multiple grid cells. Then, an adaptive point cloud interpolation module is used to refine the point cloud points within the boundary of each grid cell. This module addresses the issue of sparse point clouds in low-curvature regions while maintaining the scene's topology. A global-local Gaussian decoder is used to calculate the 3D Gaussian parameters within and outside the boundary of each grid cell. This ensures consistent Gaussian parameter predictions across different grid cells, effectively solving the consistency problem between individual grid cells. A Gaussian refinement module further refines the 3D Gaussian distribution. Dynamically adjusting the Gaussian distribution based on the complexity of the region improves the quality of the 3D Gaussian distribution while reducing redundant information. Finally, a weighted average is used to merge the 3D Gaussian distributions of all grid cells into a complete large-scene 3D Gaussian distribution, ultimately generating a complete 3D model of the large scene.

[0096] It should be noted that the purpose of disclosing the embodiments is to help further understand the present invention; however, those skilled in the art will understand that various substitutions and modifications are possible without departing from the spirit and scope of the present invention and the appended claims. Therefore, the present invention should not be limited to the content disclosed in the embodiments, and the scope of protection of the present invention is defined by the scope of the claims.

Claims

1. A high-quality large-scene reconstruction method based on 3D Gaussian sputtering, characterized in that, include: Multi-view large scene images are acquired and preprocessed to obtain preprocessed multi-view large scene images, in which there is overlap between images from adjacent views; COLMAP is used to obtain the camera pose data and point cloud data corresponding to the preprocessed multi-view large scene images. The scene's ground plane is divided into multiple grid cells, each with clearly defined boundaries and coordinate ranges. Determine the contributing cameras for each grid cell, wherein, for any given grid cell, select the group of cameras that contribute the most to it according to the visibility criterion as the contributing cameras for that grid cell; Extend the point cloud points within the boundary of each grid cell outward to obtain the extended point cloud data outside the boundary of each grid cell; The point cloud adaptive interpolation module is used to refine the point cloud points within the boundary of each grid cell, resulting in a more refined point cloud representation. The refined point cloud data within the boundary of each grid cell and the extended point cloud data outside the boundary of each grid cell are segmented using a pre-trained point cloud segmentation model to obtain segmented point cloud data. Based on the segmented point cloud data, the 3D Gaussian parameters inside and outside the boundary of each grid cell are calculated using a global-local Gaussian decoder, resulting in the 3D Gaussian distribution inside and outside the boundary of each grid cell. The Gaussian refinement module is used to dynamically adjust the 3D Gaussian distribution within the boundary of each mesh cell. The Gaussian refinement module is used to dynamically allocate the number of Gaussians according to the local geometric complexity, allocate more Gaussians to areas with complex geometric details, and delete redundant Gaussians. Remove the 3D Gaussian distribution outside the boundary of each grid cell to obtain the optimized 3D Gaussian distribution for each grid cell; The weights of all adjacent grid cells are determined based on the contribution of each grid cell camera, and a weighted average is performed on the weights of all adjacent grid cells and the optimized 3D Gaussian distribution parameters to obtain the 3D Gaussian distribution of the merged complete scene. A 3D model is generated based on the 3D Gaussian distribution of the complete scene using 3D Gaussian rendering technology.

2. The high-quality large-scene reconstruction method based on 3D Gaussian sputtering according to claim 1, characterized in that: The image preprocessing includes denoising, normalization, and cropping / scaling.

3. The high-quality large-scene reconstruction method based on 3D Gaussian sputtering according to claim 1, characterized in that: The steps to determine the contributing camera for any grid cell A are as follows: The boundary of the grid cell A is projected onto the image plane of the camera, and the projected area of ​​the grid cell A in each camera view is calculated; Calculate the total area of ​​the camera image based on the total number of pixels obtained from the camera resolution. Calculate the ratio of the projected area of ​​grid cell A in each camera viewpoint to the total area of ​​the camera image; Select the cameras with the highest proportions as contributing cameras for the grid cell A.

4. The high-quality large-scene reconstruction method based on 3D Gaussian sputtering according to claim 1, characterized in that: The specific steps for refining the point cloud points within the boundary of each grid cell using the point cloud adaptive interpolation module are as follows: Create two empty sets: one to store the indices of the points that have already been processed, and the other to store the points generated after interpolation. Iterate through each point in the initial point cloud and perform interpolation. The interpolation method for any point is as follows: Find the K nearest neighbors of the current point; Based on the K nearest neighbors, an incremental algorithm is used to partition the 3D Voronoi polygons, and the vertices of the Voronoi polygons are the potential interpolation points. The Wasserstein distance is used to evaluate the topological difference between the set of K nearest neighbors and the set containing the vertices of the Voronoi polygon. If the evaluation results show that these vertices do not destroy the topology, the vertices of the Voronoi polygon are added to the set as interpolation points; otherwise, 2D Voronoi interpolation is performed. The points generated by interpolation are merged with the original point cloud to obtain a refined point cloud.

5. The high-quality large-scene reconstruction method based on 3D Gaussian sputtering according to claim 4, characterized in that: The number of K nearest neighbors of the current point is determined based on the sparsity of the point cloud. The higher the curvature of the region, the more nearest neighbors are selected.

6. The high-quality large-scene reconstruction method based on 3D Gaussian sputtering according to claim 4, characterized in that: The specific steps of 2D Voronoi interpolation are as follows: Principal component analysis is used to project the original K nearest neighbor points onto a two-dimensional plane; Voronoi partitioning is performed on a two-dimensional plane using an incremental algorithm to obtain Voronoi polygons. Map the vertices of the 2D Voronoi polygon back to 3D space and add the mapped vertices as interpolation points to the set.

7. The high-quality large-scene reconstruction method based on 3D Gaussian sputtering according to claim 1, characterized in that: The specific steps to obtain the 3D Gaussian distribution inside and outside the boundary of each grid cell are as follows: The local Gaussian decoder is used to predict the 3D Gaussian parameters of each grid cell, and the image rendered using the predicted 3D Gaussian parameters is compared with the real image to calculate the reconstruction loss. The parameters of the local Gaussian decoder are then updated based on the reconstruction loss. Update the parameters of the global Gaussian decoder based on the local Gaussian decoders of all mesh cells after updating the parameters; The global Gaussian decoder is used to predict the 3D Gaussian parameters of each grid cell. Then, the difference between the Gaussian parameters predicted by the local Gaussian decoder and the global Gaussian decoder for each grid cell is calculated. The parameters of the local Gaussian decoder are then updated based on the difference to make them closer to the prediction of the global Gaussian decoder. The updated local Gaussian decoder is used to re-predict the 3D Gaussian parameters of each grid cell, resulting in the 3D Gaussian distribution inside and outside the boundary of each grid cell.

8. The high-quality large-scene reconstruction method based on 3D Gaussian sputtering according to claim 1, characterized in that: The specific steps for dynamically adjusting the 3D Gaussian distribution within the boundary of each mesh cell using the Gaussian refinement module are as follows: Using a keypoint scorer, a keypoint score map is calculated from image features for each input view. The keypoint score map is used to reflect the geometric complexity of different regions in the image and can indicate the importance of each region in the image. In the key point score map, for areas with high scores, the Gaussian center is further subdivided into multiple smaller centers, and for areas with low scores, the transparency and scaling of the corresponding Gaussian center are gradually reduced until it is completely removed.

9. The high-quality large-scene reconstruction method based on 3D Gaussian sputtering according to claim 1, characterized in that: The weight of adjacent grid cells is determined based on the number of cameras shared between the adjacent grid cells. The more cameras shared between adjacent grid cells, the higher their weight. If camera A is included in the contributing cameras of two adjacent grid cells, then camera A is the shared camera of these two adjacent grid cells.

Images