Reconstruction method of three-dimensional reconstruction model based on two-dimensional Gaussian splashing

Abstract

The invention provides a reconstruction method of a three-dimensional reconstruction model based on two-dimensional Gaussian splashing, which comprises the following steps: S1, carrying out sparse reconstruction on an input image sequence through a multi-view stereoscopic vision algorithm to generate an initial sparse three-dimensional point cloud and a corresponding camera pose parameter; s2, inputting an improved two-dimensional Gaussian radiation field by using the sparse three-dimensional point cloud and the camera pose as information; s3, dynamically screening a visible anchor point subset based on the current view angle parameter, and generating a rendered image through a differentiable rendering pipeline; s4, calculating a loss function of the rendering image of the training track and the input image to optimize a reconstruction scene; and S5, starting a special visualization tool, and inputting a rendering result. According to the method, by introducing a trimmable anchor point parameterization framework and a multi-scale feature fusion mechanism, light-weight and high-precision three-dimensional scene modeling is achieved, and the problems that traditional 2D Gaussian sputtering is insufficient in multi-view geometric consistency, storage overhead and weak texture region reconstruction and an existing 2D Gaussian splashing method is insufficient in self-adaptive mechanism are solved.

Description

Technical Field

[0001] The present invention relates to the technical field of three-dimensional reconstruction, and in particular, to a reconstruction method for a three-dimensional reconstruction model based on two-dimensional Gaussian splatting. Background Art

[0002] The novel view synthesis task of capturing and reconstructing a three-dimensional scene from multi-view images has always been an enduring topic in the fields of computer graphics and computer vision, and its technical achievements are widely applied in fields such as virtual reality, autonomous driving, and digital cultural heritage protection.

[0003] Traditional three-dimensional scene representations have explicit and implicit dual paradigms: explicit methods (such as volumes, meshes, point clouds) rely on discrete geometric elements and use GPU parallel architectures to achieve efficient rasterization rendering and real-time interaction; implicit representations (such as NeRF [1] ) construct continuous radiation fields through multi-layer perceptrons (MLPs), and break through the view-dependent lighting modeling bottleneck with differentiable volume rendering, but are limited by computational complexity and difficult to meet the real-time interaction requirements. Three-dimensional Gaussian splatting ( [2] [2] ) innovatively combines the advantages of both: constructs parameterized ellipsoids (position, covariance, transparency) based on the initial point cloud, and drives the gradient optimization of discrete Gaussian basis elements through a differentiable rasterization mechanism, achieving sub-millimeter reconstruction accuracy (DTU CD = 0.42) while maintaining real-time rendering capabilities (> 30fps), but the multi-view ambiguity of its ellipsoid parameters still leads to the risk of surface geometry distortion. Two-dimensional Gaussian splatting ( [3] [3] ) uses a "plane" two-dimensional Gaussian disk (surfel) embedded in three-dimensional space to replace the three-dimensional Gaussian ellipsoid, and ensures multi-view geometric consistency by introducing depth constraints and normal regularization terms, significantly improving the reconstructed geometric quality, but its texture density-based adaptive sampling mechanism lacks explicit geometric prior constraints, which may weaken the geometric detail reconstruction in textureless regions.

[0004] The proposed three-dimensional Gaussian splatting technology has achieved a landmark breakthrough in the field of novel view synthesis, but the existing view-dependent artifacts problem reveals the inherent limitations of this method in view generalization ability. In the surface reconstruction task in weakly textured or uniformly textured regions, traditional two-dimensional Gaussian splatting technology is also limited by the inherent defects of geometric representations. Therefore, there is a problem of insufficient adaptive mechanism in existing two-dimensional Gaussian splatting methods.

[0005] How to construct a lightweight multi-scale feature fusion framework to simultaneously improve geometric reconstruction accuracy and computational efficiency poses a double challenge to the mathematical completeness and engineering implementation of the algorithm architecture.

[0006] The significance of solving the above technical problems is as follows: It provides a lightweight solution for high-quality 3D reconstruction of miniaturized devices. Summary of the Invention

[0007] The present invention provides a reconstruction method for a 3D reconstruction model based on two-dimensional Gaussian splashing. By introducing a prunable anchor point parameterization framework and a multi-scale feature fusion mechanism, lightweight and high-precision 3D scene modeling is achieved to solve the deficiencies of traditional 2D Gaussian splashing in multi-view geometric consistency, storage overhead, and reconstruction in weakly textured regions, as well as the problem of insufficient adaptive mechanisms in existing two-dimensional Gaussian splashing methods.

[0008] The technical solution of the present invention is as follows: The reconstruction method of the 3D reconstruction model based on two-dimensional Gaussian splashing of the present invention includes the following steps: S1. Sparse reconstruction is performed on the input image sequence through a multi-view stereo vision algorithm to generate an initial sparse 3D point cloud and corresponding camera pose parameters; S2. Using the sparse 3D point cloud and camera pose as information, input it into the improved two-dimensional Gaussian radiation field; S3. Dynamically screen the visible anchor point subset based on the current view parameters, and generate a rendered image through a differentiable rendering pipeline; S4. Calculate the loss function of the rendered image and the input image of the training trajectory to optimize the reconstructed scene; and S5. Enable a dedicated visualization tool and input the rendering result.

[0009] Optionally, in the above reconstruction method of the 3D reconstruction model based on two-dimensional Gaussian splashing, in step S1, a circular training trajectory image is provided, and preprocessing operations such as noise suppression, motion blur processing, and image screening are performed on it; the preprocessed image is input into the multi-view stereo vision algorithm to generate a sparse 3D point cloud with RGB color information and camera pose parameters in homogeneous coordinates.

[0010] Optionally, in the above reconstruction method of the 3D reconstruction model based on two-dimensional Gaussian splashing, in step S2, an efficient view adaptation mechanism is introduced on the basis of traditional two-dimensional Gaussian splashing, which mainly includes: 1) Scene voxelization M sparse point clouds As the initial point cloud is input into the radiation field, and the scene is voxelized based on it: = 1 \* GB2⑴ The geometric center of each voxel unit is defined as the initial anchor point in 3D space, where is used to eliminate duplicate points, represents the rounding operation, and the selection of the value of the initial voxel size is the median of the nearest neighbor distances of all initial points; The center of each initialized voxel is set as the anchor point , all having the characteristics of encoding the color information of the radiation field in the local area , scaling factor Control the spatial distribution scale and offset vector Characterize the local coordinate system offset; Use to represent the camera coordinates, to represent the anchor point coordinates, and the depth distance calculation formula: = 2 * GB2⑵ The line-of-sight direction calculation formula: = 3 * GB2⑶ Calculate the view-related feature weights through the differentiable attention mechanism : = 4 * GB2⑷; For the color feature Perform downsampling on it to establish a hierarchical feature library , where represents downsampling by ; Through Fuse the feature library to form the integrated anchor point feature : = 5 * GB2⑸; 2) Deduction of two-dimensional neural Gaussian On the basis of voxelization, further construct a two-dimensional neural Gaussian splash model; for the anchor point located at the spatial position , the center point of the neural Gaussian disks related to it The calculation formula is: = 6 * GB2⑹ Related attributes of the neural two-dimensional Gaussian: scaling vector , orthogonal tangent vector , opacity , color are calculated through a specific MLP: = 7 * GB2⑺ Definition of the normal of the neural two-dimensional Gaussian splash: = 8 * GB2⑻ The view direction from the camera center to each neural Gaussian center: ​= 9 * GB2⑼ Thus, the tangent plane parameters of the neural two-dimensional Gaussian disk in world space are: = 10 * GB2⑽。

[0011] Optionally, in the above reconstruction method of the three-dimensional reconstruction model based on two-dimensional Gaussian splashing, in step S3, only the anchor points located within the current camera frustum are activated and rendered, including: 1) Frustum dynamic anchor point screening: Based on the intrinsic matrix and extrinsic matrix of the current camera, calculate the six plane equations of the frustum , each plane is represented as four-dimensional homogeneous coordinates ; for each anchor point coordinate in the world coordinate system, project it to the clip space and calculate its homogeneous coordinate: v = 11 * GB2⑾ where P is the view projection matrix, related to the internal and external parameters of the current frustum camera; if it satisfies = 12 * GB2⑿ then the anchor point is within the frustum (w is the fourth component of the homogeneous coordinate) and is activated; 2) Two-dimensional Gaussian differentiable ray-splashing rendering: For each pixel in the image plane, integrate along its ray ; parameterize the camera ray as the intersection line of two orthogonal planes, that is and ; transform the two orthogonal planes to the local coordinate system to obtain: , = 13 * GB2⒀ Calculate whether the local coordinates of the neural two-dimensional Gaussian fall on the intersection line; for the center point of the neural two-dimensional Gaussian to be rendered, perform depth value sorting to ensure alpha blending from front to back: = 14 * GB2⒁。

[0012] Optionally, in the above reconstruction method of the three-dimensional reconstruction model based on two-dimensional Gaussian splashing, in step S4, the anchor point parameters and the feature library weights are synchronously optimized through the backpropagation algorithm, and finally a compact scene representation model is output, where the loss function of the traditional two-dimensional Gaussian splashing: = 15 * GB2⒂ where, is the RGB color reconstruction loss function, that is and are combined to obtain and are the depth and normal regular terms respectively. On this basis, the voxel regular term is added to this solution: = 16 * GB2⒃ Among them, represents the number of neural Gaussians in the scene, is the neural two-dimensional Gaussian scaling vector product operation; the voxel regular term controls the neural two-dimensional Gaussian in the minimum overlap area on the screen; the final loss function formula is: = 17 * GB2⒄.

[0013] Optionally, in the above-mentioned reconstruction method of the three-dimensional reconstruction model based on two-dimensional Gaussian splashing, in step S5, the rendering result is imported into a dedicated visualization Viewer based on two-dimensional Gaussian rendering technology, and the effect is presented through a three-dimensional global perspective.

[0014] According to the technical solution of the present invention, the beneficial effects are: The reconstruction method of the three-dimensional reconstruction model based on two-dimensional Gaussian splashing of the present invention enhances the geometric prior representation of adaptive sampling, reduces the storage overhead while improving the reconstruction accuracy. The three-dimensional reconstruction model of two-dimensional Gaussian splashing reconstructed by the method of the present invention adds a parameterized anchor pruning criterion to achieve adaptive regulation of feature granularity, providing a lightweight solution for high-quality three-dimensional reconstruction of small devices. Specifically, the present invention adds an adaptive radiance field optimization framework with deletable anchor views. By constructing a mapping network between the MLP and the prunable anchor parameters, a parameterized association between the anchor system and the two-dimensional Gaussian disk attributes is established, improving the accuracy of the geometric reconstruction perspective while reducing the storage cost, and solving the Gaussian distribution ambiguity problem in the textureless area.

[0015] In order to better understand and illustrate the concept, working principle and invention effect of the present invention, the following will combine the drawings and specifically describe the present invention through specific embodiments as follows: Description of the Drawings

[0016] In order to more clearly illustrate the specific implementation manners of the present invention or the technical solutions in the prior art, the following will briefly introduce the drawings required for the description of the specific implementation manners or the prior art.

[0017] Figure 1 is the flowchart of the reconstruction method of the three-dimensional reconstruction model based on two-dimensional Gaussian splashing of the present invention; Figure 2 is the schematic diagram of two-dimensional neural Gaussian deduction; and Figure 3 It is a two-dimensional Gaussian schematic diagram of a single neuron; Figure 4 The scheme of the present invention and other schemes are compared in an experiment on the Shiny dataset scene of Ref-Nerf on the A600 server. DETAILED DESCRIPTION

[0018] In order to make the purpose, technical method and advantages of the present invention clearer, the present invention is further described in detail below in conjunction with the accompanying drawings and specific examples. These examples are only illustrative and not limiting of the present invention.

[0019] like Figure 1 As shown, the reconstruction method of the three-dimensional reconstruction model based on two-dimensional Gaussian splashing of the present invention comprises the following steps: S1. Sparsely reconstruct the input image sequence through a multi-view stereo vision algorithm (such as COLMAP) to generate an initial sparse 3D point cloud and corresponding camera pose parameters.

[0020] In this step, a circular training trajectory image is provided, and preprocessing operations such as noise suppression, motion blur processing, and image screening are performed on it. The preprocessed image is input into a multi-view stereo vision algorithm, such as COLMAP (applicable to small-scale high-precision reconstruction), MVSNet (applicable to real-time reconstruction of video streams), and AliceVision (applicable to large-scale urban modeling), to generate a sparse point cloud with RGB color information and camera pose parameters in a homogeneous coordinate system.

[0021] S2. Using sparse 3D point cloud and camera pose as information, the improved 2D Gaussian radiation field is input. Learnable anchor points are used to replace the traditional Gaussian disk. Each anchor point is associated with a dynamically adjusted set of geometric attribute parameters (position, covariance, opacity), and its optimization process is constrained by geometric saliency.

[0022] In step S2, an efficient view adaptation mechanism is introduced based on the traditional two-dimensional Gaussian splashing, which mainly includes two key steps: scene voxelization and two-dimensional neural Gaussian deduction: 1) Scene voxelization M sparse point clouds Input the radiance field as an initial point cloud and voxelize the scene based on it: = 1 \* GB2⑴ Each voxel unit The geometric center of is defined as the initial anchor point in three-dimensional space, where To eliminate duplicate points, Represents a rounding operation, the initial voxel size The value is chosen as the median of the nearest neighbor distances of all initial points.

[0023] The center of each initialized voxel is set as the anchor point and all have the characteristics of encoding the color information of the radiation field in the local area , scaling factor control the spatial distribution scale and offset vector characterize the offset of the local coordinate system

[0024] To enhance the radiation field resolution and view correlation, use to represent the camera coordinates to represent the anchor point coordinates

[0025] Depth distance calculation formula = 2 * GB2⑵ Line-of-sight direction calculation formula = 3 * GB2⑶ Calculate the view-related feature weights through the differentiable attention mechanism : = 4 * GB2⑷ For the color feature downsample it to establish a hierarchical feature library . Among them represents downsampling by . Through fuse the feature library to form an integrated anchor point feature : = 5 * GB2⑸ 2) Deduction of two-dimensional neural Gaussian The relevant process refers to Figure 2 . On the basis of voxelization, further construct a two-dimensional neural Gaussian splash model. For the anchor point at the spatial position , the center points of the neural Gaussian disks related to it are calculated by the formula = 6 * GB2⑹ Related attributes of the neural two-dimensional Gaussian: scaling vector , orthogonal tangent vector , opacity , color are calculated through a specific MLP = 7 * GB2⑺ Normal definition for Neural 2D Gaussian splatting: = 8 \* GB2⑻ The viewing direction from the camera center to the center of each neural Gaussian: = 9 \* GB2⑼ See the calculation diagram of normal and other parameters for reference. Figure 3 Therefore, the tangent plane parameters of the neural two-dimensional Gaussian disk in world space are: = 10 \* GB2⑽ S3. Dynamically select a subset of visible anchor points based on the current view parameters and generate a rendered image through a differentiable rendering pipeline. The properties of the Gaussian disk are controlled in real time by the visibility state of the anchor points to achieve view-adaptive radiation field expression.

[0026] To improve rendering efficiency, only the anchor points within the current camera's view frustum are activated and rendered. The specific operations are divided into view frustum dynamic anchor point screening and neural 2D Gaussian differentiable ray-splatter rendering.

[0027] 1. View cone dynamic anchor point screening Based on the intrinsic and extrinsic matrices of the current camera, calculate the six plane equations of the viewing cone , each plane is represented as a four-dimensional homogeneous coordinate For each anchor point coordinate in the world coordinate system , project it into the clipping space and calculate its homogeneous coordinates: v = 11 \* GB2⑾ Where P is the view projection matrix, which is related to the camera internal and external parameters of the current view cone. = 12 \* GB2⑿ Then the anchor point is within the viewing cone (w is the fourth component of the homogeneous coordinates) and is activated.

[0028] 2. 2D Gaussian Differentiable Rays - Splash Rendering For each pixel in the image plane , along its rays Integrate. Parameterize the camera light as the intersection of two orthogonal planes, that is and . Transforming the two orthogonal planes into the local coordinate system, we obtain: , = 13 \* GB2⒀ Calculate whether the local coordinates of the computational neural two-dimensional Gaussian fall on the intersection line. For the center point of the neural two-dimensional Gaussian to be rendered Perform depth value sorting to ensure alpha blending from front to back: = 14 * GB2⒁ S4. Calculate the loss function of the rendered image of the training trajectory and the input image to optimize the reconstructed scene. Synchronously optimize the anchor point parameters and the weights of the feature library through the backpropagation algorithm, and finally output a compact scene representation model. Among them, the rendered image of the training trajectory is an image with the same camera view as the input image.

[0029] Among them, the loss function of traditional two-dimensional Gaussian splashing: = 15 * GB2⒂ Among them, is the RGB color reconstruction loss function, that is and Combined to get. and are the depth and normal regularization terms respectively. On this basis, this scheme adds a voxel regularization term: = 16 * GB2⒃ Among them, represents the number of neural Gaussians in the scene, is the neural two-dimensional Gaussian scaling vector product operation. The voxel regularization term controls the minimum overlap area of the neural two-dimensional Gaussians on the screen. Therefore, the final loss function formula adopted by this scheme is: = 17 * GB2⒄ S5. Calculate the loss function of the rendered image of the training trajectory and the input image to optimize the reconstructed scene, which is convenient for users to view the three-dimensional global effect. Import the rendering result into a dedicated visualization Viewer based on two-dimensional Gaussian rendering technology, and present the effect through a three-dimensional global perspective. This scheme also provides a standardized data interface to ensure seamless docking between the Viewer module and mainstream editing tools.

[0030] Aiming at the pain points of the weak perspective generalization ability and large storage overhead of traditional three-dimensional reconstruction methods, the advantages of the method of the present invention are: 1) Adopt two-dimensional Gaussian splashing (Gaussian Splatting) with stronger structural consistency compared to three-dimensional Gaussian splashing (Gaussian Splatting). While maintaining the surface reconstruction accuracy, retain the editing flexibility of the two-dimensional parameterization system, and support physical-based material property adjustment and topological structure optimization; Adopt a dynamic adjustable anchor point radiation field framework to compress the independently optimized Gaussian parameters in two-dimensional Gaussian Splatting into low-dimensional anchor point features, achieving a significant reduction in the model storage overhead.

[0031] 2) Adopt a multi-scale feature library and view-related weight fusion to significantly suppress the distribution ambiguity in textureless regions, effectively solving problems such as parameter ambiguity of 2D Gaussian Splatting in weakly textured regions; 3) The present invention also provides visualization software to support the visualization and quantitative evaluation of the reconstruction results.

[0032] To verify the advantages of the present invention, the present invention's solution and other solutions conduct a comparative experiment on the Shiny [4] dataset scenario on an A600 server. As Figure 4 shown in and Table 1, by visualizing the normal maps reconstructed by traditional two-dimensional Gaussian splashing and the present method, it can be observed that the present solution will have better performance in the rendering results (especially in weakly textured regions) and training time. Table SEQ Table \* ARABIC 1

[0033] Based on the traditional two-dimensional Gaussian Splatting model, the present invention proposes a neural two-dimensional Gaussian radiation field reconstruction method based on an efficient view adaptation mechanism. By introducing a scene voxelization hierarchical feature fusion strategy and a neural Gaussian deduction mechanism, and cooperating with the loss function optimization of voxel regularization constraints, the geometric consistency of the reconstructed scene is significantly improved. It can be widely applied in the fields of computer vision, virtual reality, etc.

[0034] The above description is the best embodiment according to the concept and working principle of the invention. The above embodiments should not be construed as limiting the protection scope of the present claims. Combinations of other implementation manners and implementation modes in accordance with the concept of the present invention all fall within the protection scope of the present invention.

[0035] [1]MILDENHALL B, SRINIVASAN P P, TANCIK M, et al. NeRF: RepresentingScenes as Neural Radiance Fields for View Synthesis[A / OL]. arXiv, 2020[2025-04-01]. http: / / arxiv.org / abs / 2003.08934. DOI:10.48550 / arXiv.2003.08934. [2]KERBL B, KOPANAS G, LEIMKÜHLER T, et al. 3D Gaussian Splatting for Real-Time Radiance Field Rendering [A / OL]. arXiv, 2023 [2025-04-01]. http: / / arxiv.org / abs / 2308.04079. DOI:10.48550 / arXiv.2308.04079. [3]HUANG B, YU Z, CHEN A, et al. 2D Gaussian Splatting for Geometrically Accurate Radiance Fields [C / OL] / / Special Interest Group on Computer Graphics and Interactive Techniques Conference Conference Papers ’24. 2024: 1-11 [2025-04-01]. http: / / arxiv.org / abs / 2403.17888. DOI:10.1145 / 3641519.3657428. [4]VERBIN D, HEDMAN P, MILDENHALL B, et al. Ref-NeRF: Structured View-Dependent Appearance for Neural Radiance Fields [A / OL]. arXiv, 2021 [2025-04-23]. http: / / arxiv.org / abs / 2112.03907. DOI:10.48550 / arXiv.2112.03907.

Claims

1. A reconstruction method for a three-dimensional reconstruction model based on two-dimensional Gaussian splashing, characterized in that, Including the following steps: S1. Sparse reconstruction is performed on the input image sequence through a multi-view stereo vision algorithm to generate an initial sparse three-dimensional point cloud and corresponding camera pose parameters; S2. Using the sparse three-dimensional point cloud and camera pose information, input it into the improved two-dimensional Gaussian radiation field; S3. Dynamically screen the visible anchor point subset based on the current view parameters, and generate a rendered image through a differentiable rendering pipeline; S4. Calculate the loss function between the rendered image of the training trajectory and the input image to optimize the reconstructed scene; And S5. Enable a dedicated visualization tool and input the rendering result.

2. The reconstruction method of the three-dimensional reconstruction model based on two-dimensional Gaussian splashing according to claim 1, characterized in that, In step S1, a circular training trajectory image is provided, and preprocessing operations such as noise suppression, motion blur processing, and image screening are performed on it; the preprocessed image is input into the multi-view stereo vision algorithm to generate the sparse three-dimensional point cloud with RGB color information and the parameters of the camera pose in homogeneous coordinates.

3. The reconstruction method of the three-dimensional reconstruction model based on two-dimensional Gaussian splashing according to claim 1, wherein, In step S2, an efficient view adaptation mechanism is introduced on the basis of traditional two-dimensional Gaussian splashing, mainly including: 1) Scene voxelization M sparse point clouds Input the radiation field with the initial point cloud and voxelize the scene based on it: ⑴ Each voxel unit whose geometric center is defined as the initial anchor point in three-dimensional space, where to eliminate duplicate points, denoted as a rounding operation, the initial voxel size value is selected as the median of the nearest neighbor distances of all initial points; The center of each initialized voxel is set as the anchor point and has the characteristic of encoding the color information of the radiation field in the local area , scaling factor to control the spatial distribution scale and offset vector representing the offset of the local coordinate system; Use to represent the camera coordinates, to represent the anchor point coordinates, and the depth distance calculation formula: ⑵ Line-of-sight direction calculation formula: ⑶ Calculating view-related feature weights through a differentiable attention mechanism : ⑷； For color features Downsample it to establish a hierarchical feature library , where Indicates to Downsample ; Through Fuse the feature library to form an integrated anchor feature : ⑸； 2) Deduction of two-dimensional neural Gaussian On the basis of voxelization, a two-dimensional neural Gaussian splash model is further constructed; for the spatial position anchor point , the center points of the neural Gaussian disks related to it are calculated by the formula: ⑹ Relevant attributes of the neural two-dimensional Gaussian: scaling vector , orthogonal tangent vector , opacity , color The calculation is obtained through a specific MLP: ⑺ Normal definition of neural two-dimensional Gaussian splashing: ⑻ View direction from the camera center to each neural Gaussian center: ⑼ Thus, the tangent plane parameters of the neural two-dimensional Gaussian disk in world space are: ⑽。 4. The reconstruction method of the three-dimensional reconstruction model based on two-dimensional Gaussian splashing according to claim 1, wherein In step S3, only the anchor points located within the current camera frustum are activated and rendered, including: 1) Frustum dynamic anchor point screening: Calculate the six plane equations of the frustum based on the internal parameter matrix and external parameter matrix of the current camera , and each plane is represented as four-dimensional homogeneous coordinates ; for each anchor point coordinate in the world coordinate system , project it onto the clip space and calculate its homogeneous coordinate: v ⑾ Among them, P is the view projection matrix, which is related to the internal and external parameters of the camera of the current frustum; if it satisfies ⑿ Then the anchor point is within the frustum (w is the fourth component of the homogeneous coordinate) and is activated; 2) Two-dimensional Gaussian differentiable ray-splashing rendering: For each pixel in the image plane , integrate along its ray ; parameterize the camera ray as the intersection of two orthogonal planes, namely and ; transform the two orthogonal planes to the local coordinate system to obtain: , ⒀ Calculate whether the local coordinates of the computational neural two-dimensional Gaussian fall on the intersection line; for the center point of the neural two-dimensional Gaussian to be rendered Perform depth value sorting to ensure alpha blending from front to back: ⒁。 5. The reconstruction method of the three-dimensional reconstruction model based on two-dimensional Gaussian splash according to claim 1, wherein In step S4, the anchor point parameters and the feature library weights are synchronously optimized through the backpropagation algorithm, and finally a compact scene representation model is output. Among them, the loss function of traditional two-dimensional Gaussian splashing: ⒂ Among them, is the RGB color reconstruction loss function, that is and are combined to obtain and are the depth and normal regular terms respectively. On this basis, the voxel regular term is added to this scheme: ⒃ Among them, represents the number of neural Gaussians in the scene, is the neural two-dimensional Gaussian scaling vector product operation; the voxel regularization term controls the neural two-dimensional Gaussian in the minimum overlap region on the screen; the final loss function formula is: ⒄。 6. The reconstruction method of the three-dimensional reconstruction model based on two-dimensional Gaussian splash according to claim 1, wherein In step S5, the rendering result is imported into a dedicated visualization Viewer based on two-dimensional Gaussian rendering technology, and the effect is presented through a three-dimensional global view.

Images