A switching LOD rendering method based on massive Gaussian splash model data

Abstract

The application relates to the field of computer graphics and three-dimensional scene rendering technology, and discloses a switching LOD rendering method based on massive Gaussian splash model data, which comprises the following steps: establishing a structure redundancy responsibility graph supporting Gaussians and dependent Gaussians in a space block, and performing prefix truncatable rearrangement on Gaussian data; removing dependent Gaussians according to a truncation ratio to generate a truncated residual tensor signature, and generating a standby truncated residual tensor signature for boundary support Gaussians; in a rendering stage, acquiring the truncated level values of adjacent blocks to perform arbitration of the dominant right, and activating the corresponding tensor signature to perform cross-border proxy compensation; and based on the activated signature, performing directional covariance and appearance color compensation on the support Gaussians and rendering and outputting. Through the topological rearrangement and tensor signature mechanism, the application realizes lightweight topography reconstruction without full data traversal, effectively fills the cross-block visual joint, reduces the memory space cost, and guarantees the high quality of a dynamic rendering picture.

Description

Technical Field

[0001] This invention relates to the field of computer graphics and 3D scene rendering technology, specifically a switching LOD rendering method based on massive Gaussian splash model data. Background Technology

[0002] 3D Gaussian splashing technology uses a large number of anisotropic Gaussian volumes to represent 3D scenes, enabling high-quality compositing of new perspectives and real-time rendering. However, when processing massive amounts of Gaussian splashing data for large-scale scenes, full rendering consumes a huge amount of video memory and incurs extremely high computational bandwidth overhead. To balance rendering performance and image quality, it is usually necessary to introduce Level of Detail (LOD) technology to dynamically cull and simplify scene data based on the camera's viewpoint.

[0003] Existing Gaussian splashing LOD schemes have significant limitations in terms of data scheduling efficiency and visual quality maintenance. At the data scheduling level, existing methods typically rely on discrete memory addressing and state marking operations when performing Gaussian volume culling. This approach disrupts the continuous arrangement of data in video memory. When the rendering pipeline reads data, non-contiguous memory access causes severe bandwidth bottlenecks, making it difficult to meet the real-time pipeline's requirement for rapid truncation of underlying data.

[0004] In terms of maintaining visual quality, directly removing a portion of the Gaussian volume causes irreversible loss of geometric features and voids in the image. To compensate for these voids, conventional solutions often involve isotropically enlarging the retained Gaussian volume. This process cannot restore the structural extension characteristics of the removed data in specific physical directions and does not consider the changes in the optical absorption cross-section after the Gaussian volume expands, easily causing distortion in the calculation of light transmittance, resulting in energy buildup and overexposure artifacts in local rendering. Furthermore, for block-based processing architectures in large-scale scenes, the difference in truncation ratios between adjacent spatial blocks at different detail levels will create obvious visual seams at the physical interface. Existing boundary stitching solutions lack global state awareness and independent permission arbitration mechanisms. Directly calling cross-block data can easily cause memory read / write conflicts in the multi-threaded rendering stage, making it difficult to guarantee the spatial continuity and temporal stability of the scene image under dynamic roaming perspectives.

[0005] Therefore, this invention proposes a switching LOD rendering method based on massive Gaussian splash model data to address the shortcomings of existing technologies. Summary of the Invention

[0006] To address the shortcomings of existing technologies, this invention provides a switching LOD rendering method based on massive Gaussian splash model data. This method solves the problems of low efficiency in removing low-level data, severe distortion in shape compensation after removal, and easy generation of visual seams and video memory read / write conflicts when rendering across spatial blocks in existing Gaussian splash multi-level detail rendering technologies for large-scale scenes.

[0007] To achieve the above objectives, the present invention provides the following technical solution: a switching LOD rendering method based on massive Gaussian splash model data, comprising:

[0008] Within the spatial block, a structural redundancy responsibility graph is established for supporting Gaussians and dependent Gaussians. Based on the structural redundancy responsibility graph, a prefix-truncation rearrangement is performed on supporting Gaussians and dependent Gaussians.

[0009] According to the truncation ratio setting, a culling operation is performed on the attached Gaussian, a truncated residual tensor signature is generated for the culled attached Gaussian, and an alternative truncated residual tensor signature facing the shared boundary surface is generated for the supporting Gaussian at the boundary position.

[0010] During the real-time rendering phase, the truncation level values ​​of the current spatial block and adjacent spatial blocks are obtained, and the dominance arbitration is performed based on the truncation level values ​​of the current spatial block and adjacent spatial blocks.

[0011] Activate the truncated residual tensor signature corresponding to the Gaussian support based on the arbitration result, or activate the alternative truncated residual tensor signature to perform cross-boundary proxy compensation;

[0012] Based on the activated truncated residual tensor signature or the alternate truncated residual tensor signature, directional covariance compensation and appearance color compensation are performed on the corresponding Support Gaussian input rendering pipeline, and the compensated Support Gaussian input rendering pipeline is used for rendering output.

[0013] Preferably, a structural redundancy responsibility graph is established within the spatial block, comprising supporting Gaussians and dependent Gaussians, and a prefix-truncation rearrangement of supporting Gaussians and dependent Gaussians is performed based on the structural redundancy responsibility graph, including:

[0014] The three-dimensional scene is divided into multiple spatial grid units based on Morton coding and a preset Morton hierarchy to construct the spatial blocks;

[0015] Through iterative screening, the Gaussian body with the highest exclusive coverage contribution value in each spatial block is established as the supporting Gaussian body, and the other Gaussian bodies in the spatial block other than the supporting Gaussian body are selected as candidate Gaussian bodies.

[0016] When the spatial envelopment parameter of the candidate Gaussian body is greater than the envelopment threshold and the intrinsic appearance difference parameter is less than the intrinsic appearance difference threshold, the candidate Gaussian body is established as a dependent Gaussian belonging to the corresponding supporting Gaussian, and the structural redundancy responsibility graph is established.

[0017] In the contiguous video memory array, the supporting Gaussian arrangement is placed at the starting position, and the corresponding dependent Gaussian arrangement is placed after the supporting Gaussian arrangement to complete the prefix truncation rearrangement.

[0018] Preferably, generating truncated residual tensor signatures for the discarded Gaussian-dependent tensors includes:

[0019] Extract the 0th-order energy residual, 2nd-order covariance residual matrix, and appearance color residual of the removed attached Gaussian relative to the corresponding supporting Gaussian;

[0020] Perform eigenvalue decomposition on the second-order covariance residual matrix to extract the dominant residual direction;

[0021] The 0th-order energy residual, appearance color residual, and residual dominant direction are packaged to generate a truncated residual tensor signature, and the truncated residual tensor signature is written into the Gaussian-supporting vertex attributes.

[0022] Preferably, the alternative truncated residual tensor signature for the Gaussian-supported generation facing the shared boundary surface at the boundary location includes:

[0023] The supporting Gaussian whose vertical distance to the physical boundary of its spatial block is less than the boundary distance threshold is identified as the supporting Gaussian at the boundary position.

[0024] The dependent Gaussian that has a spatial wrapping relationship with the supporting Gaussian at the boundary position in the adjacent spatial block is included in the residual accumulation calculation range to generate the spare truncated residual tensor signature.

[0025] The signature of the spare truncated residual tensor is written into the Gaussian-supporting vertex attribute at the boundary position by combining the boundary corresponding index.

[0026] Preferably, during the real-time rendering phase, the truncation level values ​​of the current spatial block and adjacent spatial blocks are obtained, including:

[0027] The state function value of the current spatial block is calculated based on the camera distance, viewpoint velocity, and projected area variance.

[0028] The state function value is mapped to a truncation level value, and the truncation level value is written into the global state memory area.

[0029] When the absolute value of the difference between the variance of the projected area of ​​the current frame and the previous frame is greater than the fluctuation threshold, the occlusion freeze gating mechanism is triggered to force the truncation level value of the previous frame to remain unchanged.

[0030] Preferably, the dominance arbitration is performed based on the truncation level values ​​of the current spatial block and adjacent spatial blocks, including:

[0031] The truncation level values ​​of adjacent spatial blocks are extracted by reading the global state memory area;

[0032] When the truncation level value of the current spatial block is less than the truncation level value of the adjacent spatial block, it is determined that the current spatial block has the dominant right to express details.

[0033] When the truncation level value of the current spatial block is equal to the truncation level value of the adjacent spatial block, the ownership of the detailed expression dominance is determined according to a preset unique rule.

[0034] Preferably, activating the truncated residual tensor signature corresponding to the Gaussian support based on the arbitration result, or activating the alternative truncated residual tensor signature to perform cross-boundary proxy compensation, includes:

[0035] When the current spatial block has the dominant control over the detailed representation, activate the truncated residual tensor signature corresponding to the Gaussian support at the boundary position within the current spatial block;

[0036] When the adjacent spatial block has the dominant control over the detailed representation, the local residual compensation of the current spatial block is suppressed, and the backup truncated residual tensor signature corresponding to the Gaussian support at the boundary position in the adjacent spatial block is activated to perform cross-boundary proxy compensation.

[0037] Preferably, based on the activated truncated residual tensor signature or the alternate truncated residual tensor signature, directional covariance compensation and appearance color compensation are performed on the corresponding Gaussian-supporting tensor, including:

[0038] Extract the 0th-order energy residual and residual dominance direction from the activated truncated residual tensor signature or the alternative truncated residual tensor signature;

[0039] The 0th-order energy residual is converted into an expansion coefficient. Using the expansion coefficient and the largest eigenvalue of the supporting Gaussian 3D covariance matrix, a covariance tensor outer product stretching operation is performed on the supporting Gaussian 3D covariance matrix along the dominant direction of the residual to obtain the updated 3D covariance matrix.

[0040] Preferably, performing directional covariance compensation and appearance color compensation on the corresponding Gaussian support further includes:

[0041] The equivalent volume expansion coefficient is obtained by taking the square root of the ratio between the determinant of the updated three-dimensional covariance matrix and the determinant of the three-dimensional covariance matrix supporting Gaussian initialization.

[0042] Based on the equivalent optical thickness constraint, the opacity supporting Gaussian is nonlinearly corrected using the equivalent volume expansion coefficient.

[0043] Extract the appearance color residual from the activated truncated residual tensor signature or the backup truncated residual tensor signature, and superimpose the appearance color residual onto the 0th-order spherical harmonic coefficients that support Gaussians to obtain the updated 0th-order spherical harmonic coefficients.

[0044] Preferably, after performing directional covariance compensation and appearance color compensation on the corresponding Gaussian support, the method further includes:

[0045] The supported Gaussian projection, which includes the updated 3D covariance matrix, the nonlinearly corrected opacity, and the updated 0th-order spherical harmonic coefficients, is projected onto the 2D screen space.

[0046] The image is then subjected to Gaussian rasterization blending shading processing and output after being projected onto the two-dimensional screen space.

[0047] This invention provides a switching LOD rendering method based on massive Gaussian splash model data. It has the following beneficial effects:

[0048] 1. This invention performs prefix-based truncation rearrangement of Gaussian data by establishing a structural redundancy responsibility graph. Combined with the extracted truncation residual tensor signature and a backup signature oriented towards shared boundaries, the tensor signature is dynamically activated during the rendering phase based on the dominance arbitration result to perform targeted compensation. This scheme deeply coordinates data culling and lightweight compensation, achieving geometric shape and color reconstruction after multi-level detail truncation without loading the full amount of underlying vertex data. Furthermore, it effectively bridges cross-boundary visual seams through an out-of-boundary proxy mechanism, thereby maintaining high-quality scene visuals while reducing memory usage and computational overhead.

[0049] 2. This invention classifies supporting Gaussian and dependent Gaussian structures by comprehensively evaluating exclusive coverage contribution, spatial envelopment parameters, and intrinsic appearance difference parameters. This mechanism combines the geometric overlap characteristics of Gaussian volumes with diffuse color attributes, effectively avoiding topological connection errors caused by clustering based solely on spatial distance, and ensuring the accuracy of prefix-truncated data arrangement at both the physical distribution and visual representation levels.

[0050] 3. This invention utilizes the residual-dominant direction to perform tensor outer product stretching on the covariance matrix supporting the Gaussian, and performs nonlinear correction on its opacity based on the equivalent volume expansion coefficient. This not only accurately compensates for the spatial shape loss in the most significant direction after the Gaussian volume is removed, but also dynamically adjusts the optical absorptivity of the expanded Gaussian volume according to the physical transmission law, avoiding visual energy overexposure or abnormal accumulation in local rendering.

[0051] 4. This invention introduces a dominance arbitration mechanism based on truncation level comparison and occlusion freeze gating logic during the rendering stage. A one-way handover of cross-block compensation permissions is established through global state queries, avoiding memory read / write conflicts caused by boundary overlap compensation. Simultaneously, projection area variance monitoring limits frequent truncation level jumps under sudden occlusion conditions, ensuring visual stability during dynamic viewpoint roaming. Attached Figure Description

[0052] Figure 1 This is a schematic diagram of the system architecture according to an embodiment of the present invention;

[0053] Figure 2 This is a schematic diagram of the method flow according to an embodiment of the present invention;

[0054] Figure 3 This is a schematic diagram illustrating the specific workflow of constructing the structural redundancy responsibility diagram and the descending order reordering of video memory in an embodiment of the present invention;

[0055] Figure 4 This is a schematic diagram illustrating the specific workflow of generating and encapsulating truncated residual tensor signatures according to an embodiment of the present invention.

[0056] Figure 5 This is a schematic diagram illustrating the specific workflow of execution state-aware boundary arbitration and signature activation determination in an embodiment of the present invention.

[0057] Figure 6 This is a schematic diagram illustrating the specific workflow of performing directional covariance stretching compensation and rendering output in an embodiment of the present invention;

[0058] Figure 7 This is a scatter plot of three-dimensional distribution of the Gaussian topological relationship of urban blocks and the dominant direction of residuals in an embodiment of the present invention.

[0059] Figure 8 This is a schematic diagram showing the dual-axis comparison curves of rendering performance and video memory usage for different viewpoint roaming trajectories in an embodiment of the present invention.

[0060] Figure 9 This is a clustered columnar diagram of the objective evaluation index of the rendered image quality of multiple urban blocks in an embodiment of the present invention; wherein, (a) is a comparison of the peak signal-to-noise ratio of different models in a typical scene, and (b) is a comparison of the structural similarity of different models in a typical scene.

[0061] Among them, 10 is the redundancy assessment module; 20 is the residual coding module; 30 is the state arbitration module; and 40 is the directional compensation module. Detailed Implementation

[0062] The technical solutions in 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 embodiments of the present invention, and not all embodiments. 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.

[0063] See attached document Figure 1 This invention provides a switching LOD rendering system based on massive Gaussian splash model data to execute a switching LOD rendering method based on massive Gaussian splash model data; the system includes:

[0064] The redundancy assessment module 10 is used to divide the 3D scene into uniform spatial blocks. Within each spatial block, the redundancy assessment module 10 calculates the spatial enclosure degree and intrinsic appearance difference parameters between Gaussian volumes, constructs a structural redundancy responsibility map under typical view set constraints using the calculated parameters, and performs a descending order rearrangement operation on the Gaussian data in the video memory buffer based on the structural redundancy responsibility map.

[0065] The residual encoding module 20 is used to calculate the difference parameters between the removed Gaussian body and the supporting Gaussian body, and generate a truncated residual tensor signature based on the 0th-order energy residual, centroid offset vector, covariance difference matrix, and appearance color residual. The residual encoding module 20 also calculates alternative truncated residual tensor signatures facing the adjacent block boundaries for the Gaussian body at the boundary position, and encapsulates the calculated signature data into the Gaussian body vertex attributes.

[0066] The state arbitration module 30 is used to obtain the variance of the projected area of ​​the camera parameters and spatial blocks during the rendering stage, calculate the truncation level value of each spatial block, and store it in the global state memory area. The state arbitration module 30 reads the truncation level values ​​of adjacent spatial blocks for comparison, and outputs an arbitration command to determine the activation state of the truncation residual tensor signature based on the comparison result.

[0067] The directional compensation module 40 is used to extract the dominant direction in the truncated residual tensor signature and stretch the Gaussian-supporting 3D covariance matrix along this direction. The directional compensation module 40 calculates the equivalent volume footprint ratio before and after stretching, performs nonlinear correction on the opacity according to the optical thickness attenuation rule, and superimposes the appearance color residual onto the base color parameters, inputting the updated data into the rendering pipeline.

[0068] See attached document Figure 2 This invention provides a method for switching LOD rendering based on massive Gaussian splash model data, including the following steps:

[0069] S100 divides the 3D scene into multiple spatial blocks, calculates the spatial enclosure degree and intrinsic appearance difference parameters of the Gaussian volume within the spatial block to construct a structural redundancy responsibility map, and reorders the Gaussian data in descending order based on the structural redundancy responsibility map.

[0070] S200 calculates the difference parameters between the culled Gaussian volume and the supporting Gaussian volume, generates the truncated residual tensor signature and the boundary-oriented alternative truncated residual tensor signature, and writes them as additional vertex attributes into the data structure.

[0071] S300 calculates the truncation level value of each spatial block during the real-time rendering stage and writes it into the global state memory area. It determines the activation ownership of the truncation residual tensor signature by comparing the truncation level values ​​of adjacent spatial blocks.

[0072] S400 performs directional stretching on the Gaussian-supporting covariance matrix based on the dominant direction of the truncated residual tensor signature, performs non-linear correction on the opacity in combination with the equivalent volume footprint ratio, and superimposes the appearance color residual before sending it into the rendering pipeline for blending shading.

[0073] To further clarify the implementation of each technical aspect of the present invention, the following will provide a detailed description of the implementation of each functional module involved above and its internal processing flow.

[0074] Reference Figure 3 In this embodiment, in order to establish local dependency candidate relationships based on the physical overlap between Gaussian volumes, and to establish local responsibility mapping relationships based on exclusive coverage contributions under the constraints of a typical view set, forming a data memory topology that supports seamless pruning, the above step S100 may specifically include the following sub-steps:

[0075] S110: Acquire initial Gaussian splash data and divide it into independent spatial blocks. In this embodiment, the redundancy evaluation module 10 acquires the initial Gaussian splash data and extracts the three-dimensional center coordinates of all Gaussian bodies. To achieve fast retrieval and block division of spatially adjacent data, the system performs Morton coding on the three-dimensional center coordinates to generate corresponding one-dimensional Morton code sequences.

[0076] Specifically, Morton coding maps multidimensional spatial data into one-dimensional integers by interleaving the binary bits of three-dimensional coordinates, thereby preserving the spatial locality of the data while reducing dimensionality. As a preferred method, the redundancy evaluation module 10 sorts all Gaussian bodies according to the numerical value of the one-dimensional Morton code sequence, and assigns each Gaussian body to the corresponding independent spatial block according to the spatial grid unit corresponding to the preset Morton level.

[0077] S120, calculate the spatial enclosure degree and intrinsic appearance difference parameters within the spatial block. After completing the division of the spatial blocks, in order to accurately quantify the mutual occlusion and color similarity between Gaussian bodies, the redundancy evaluation module 10 calculates the spatial enclosure degree and intrinsic appearance difference parameters between Gaussian bodies within each spatial block.

[0078] As a specific implementation, the redundancy assessment module 10 pre-establishes a typical view set. The typical view set refers to the set of camera pose parameters uniformly distributed at fixed spherical angle intervals around the periphery of the 3D scene. The redundancy assessment module 10 estimates the physical volume ratio of the overlapping two Gaussian bodies based on the geometric relationship of their 3D covariance ellipsoids; the typical view set is used for subsequent evaluation of exclusive coverage contribution. The redundancy assessment module 10 obtains the spatial enclosure parameter of the first Gaussian body relative to the second Gaussian body according to the spatial enclosure calculation formula. The spatial enclosure calculation formula is:

[0079] ;

[0080] In the formula, This represents the spatial envelopment parameter of the first Gaussian body on the second Gaussian body; This represents the approximate volume of the overlapping portion of the three-dimensional covariance ellipsoids of the first and second Gaussian bodies; This represents the total volume of the second Gaussian body's three-dimensional covariance ellipsoid; To prevent extremely small constants with a denominator of 0, a value of 10 is usually taken. -6 .

[0081] Introducing this minimal constant effectively avoids computational overflow anomalies that occur when the volume of the second Gaussian body degenerates to near zero. The technical purpose of this formula is to measure, from a purely physical geometric perspective, the degree to which the second Gaussian body is contained within the first Gaussian body.

[0082] Subsequently, the redundancy evaluation module 10 extracts the 0th-order spherical harmonic coefficients from the Gaussian volume vertex attributes. Considering that simple geometric wrapping is insufficient to completely determine the rendered appearance, the redundancy evaluation module 10 obtains the intrinsic appearance difference parameters by calculating the Euclidean distance between the corresponding 0th-order spherical harmonic coefficients of the two Gaussian volumes in the base color space. The formula for calculating the intrinsic appearance difference parameters is:

[0083] ;

[0084] In the formula, Indicates intrinsic appearance difference parameters; This represents the vector of 0th-order spherical harmonic coefficients of the first Gaussian body, which contains 3 color channels. This represents the vector of 0th-order spherical harmonic coefficients of the second Gaussian body, which contains three color channels. This represents the 2-norm operation of a vector. This computational operation aims to quantify the differences between adjacent Gaussian volumes at the diffuse base color level.

[0085] S130, Evaluate the exclusive coverage contribution and construct a structural redundancy responsibility map. Based on the aforementioned basic physical characteristics, the redundancy evaluation module 10 evaluates the exclusive coverage contribution of each Gaussian body within a spatial block based on a typical view set, and constructs a structural redundancy responsibility map based on the exclusive coverage contribution. The exclusive coverage contribution refers to the total effective pixel area of ​​a single Gaussian body on the screen projection plane that is not occluded by other Gaussian bodies within the typical view set.

[0086] In this embodiment, the specific calculation method involves performing depth sorting on multiple camera viewpoints within a typical field of view, counting the number of pixels in the most visible segment of the target Gaussian body under each viewpoint, and accumulating the results. The redundancy evaluation module 10 filters Gaussian bodies within the spatial block that have not yet established a subordinate relationship based on the magnitude of their exclusive coverage contribution, and establishes the Gaussian body with the highest exclusive coverage contribution value as the supporting Gaussian.

[0087] For the remaining candidate Gaussian bodies within the current spatial block, the system further compares the spatial enclosure parameters and intrinsic appearance difference parameters between the candidate Gaussian bodies and the supporting Gaussian bodies.

[0088] When a candidate Gaussian body does not meet the criteria for determining a dependent Gaussian body, the candidate Gaussian body is retained as a candidate Gaussian body without a subordinate relationship, and the support Gaussian screening is performed again in subsequent rounds until all Gaussian bodies in the spatial block are established as support Gaussian bodies or dependent Gaussian bodies.

[0089] When the spatial enclosure parameter of a candidate Gaussian volume is determined to be greater than a set enclosure threshold, and the intrinsic appearance difference parameter is less than a set intrinsic appearance difference threshold, the redundancy evaluation module 10 establishes the candidate Gaussian volume as a dependent Gaussian belonging to the supporting Gaussian. In this embodiment, the enclosure threshold is generally set between 0.75 and 0.90; the above threshold is a preset threshold, which can be determined based on the pre-rendering results of typical scenes combined with a preset visual error tolerance.

[0090] By combining multi-dimensional judgment logic based on physical packaging and color differences, the system can effectively avoid erroneous clustering caused by relying solely on a single overlapping feature. Finally, the redundancy assessment module 10 generates a directed structural redundancy responsibility graph with the supporting Gaussian as the starting node and the dependent Gaussian as the ending node, based on the established subordinate relationships.

[0091] S140, execute a responsibility-based topology-driven prefix-truncate data layout. In order to transform the established logical topology into a data format that can be used by the underlying rendering pipeline, the redundancy assessment module 10 executes a responsibility-based topology-driven prefix-truncate data layout on the Gaussian data within the spatial block according to the connection relationship of the structural redundancy responsibility graph.

[0092] As a preferred approach, the redundancy evaluation module 10 first arranges the Gaussian support structures in descending order of their exclusive coverage contribution in a continuous video memory array buffer, and then arranges the Gaussian support data structures at the starting address position of the address space corresponding to the space block.

[0093] The redundancy assessment module 10 obtains the spatial enclosure parameters of all dependent Gaussians associated with the supporting Gaussian. Subsequently, the system arranges the data structures of dependent Gaussians sequentially after the supporting Gaussian in descending order of the spatial enclosure parameters.

[0094] With this responsibility-driven topology-driven continuous memory data arrangement structure, when the rendering pipeline executes the instantiation drawing call instruction, it only needs to adjust the instance prefix index length value passed to the graphics application interface to directly remove dependent Gaussian data with less redundancy from the tail to the head of the array.

[0095] Compared to discrete addressing memory deletion operations, this arrangement rule helps to avoid performing a full traversal search in the chain of responsibility truncation operation, providing basic data structure support for subsequent detailed level culling.

[0096] Reference Figure 4 In this embodiment, for the excluded Gaussian body subset, in order to extract its volume and energy residuals and construct a master-slave signature data structure to support cross-boundary transfers, the above step S200 may specifically include the following sub-steps:

[0097] S210, truncate the subset and extract multi-order basic residuals. In this embodiment, the residual encoding module 20 obtains the set of dependent Gaussians that have undergone memory rearrangement and identifies the dependent Gaussians to be removed based on the set truncation ratio parameter. The truncation ratio parameter determines the percentage of data discarded at different levels of detail. To achieve a smooth transition between levels of detail, the residual encoding module 20 identifies the dependent Gaussians located at the end of the contiguous memory array and within the truncation ratio range as the subset of Gaussians to be removed.

[0098] To quantify the visual and geometric information carried by these removed Gaussian volumes, the residual encoding module 20 calculates multi-order fundamental residuals for each supporting Gaussian volume. Specifically, the residual encoding module 20 accumulates the opacity-weighted volume values ​​of the removed Gaussian volume subset to obtain the 0th-order energy residual, and calculates the weighted offset vector of the 3D center coordinates of the removed Gaussian volume relative to the 3D center coordinates of the supporting Gaussian volume to obtain the 1st-order centroid residual.

[0099] To characterize the loss in spatial shape, the residual coding module 20 calculates the sum of the differences in three-dimensional covariance between the discarded Gaussian subset and the supporting Gaussian, thereby obtaining the second-order covariance residual matrix. The formula for calculating the second-order covariance residual matrix is ​​as follows:

[0100] ;

[0101] In the formula, Represents the second-order covariance residual matrix; This represents the total number of Gaussian bodies removed from the Gaussian body subset. Indicates the first The weighting coefficients of the eliminated Gaussian body are dimensionless parameters; Indicates the first The three-dimensional covariance matrix of the eliminated Gaussian body; This represents the three-dimensional covariance matrix that supports Gaussianism.

[0102] As a preferred method, in calculating the weighting coefficients The system then adds the sum of the opacities of all Gaussian bodies within the subset to a minimal anti-overflow constant to obtain a correction benchmark value. Finally, it divides the opacity of each removed Gaussian body by this correction benchmark value to obtain the final weight value. In this embodiment, the anti-overflow constant can be 10. -6 The introduction of this minimal constant aims to avoid division anomalies caused by the denominator approaching 0 when all Gaussian volumes within a subset are completely transparent, thus ensuring the completeness of the weight calculation logic.

[0103] The technical purpose of this formula is to aggregate the geometric deviations of multiple discrete small Gaussian bodies into the covariance space of the supporting Gaussian. After capturing the geometric difference, the residual encoding module 20 calculates the difference between the discarded Gaussian subset and the supporting Gaussian at the 0th order spherical harmonic coefficients to obtain the appearance color residual, which is used to compensate for the lost base diffuse color during rendering.

[0104] S220 performs dominant eigenvalue decomposition of the second-order covariance residuals. To compress the complex three-dimensional covariance difference matrix into a compact format suitable for vertex attribute storage, the residual encoding module 20 performs eigenvalue decomposition on the second-order covariance residual matrix. For the matrix diagonalization of the eigenvalue decomposition, those skilled in the art can refer to standard linear algebra algorithms; its matrix diagonalization is a well-known technique in the field and will not be elaborated upon here.

[0105] Through eigenvalue decomposition, the residual coding module 20 obtains three mutually orthogonal eigenvectors and their corresponding eigenvalues. The magnitude of the eigenvalue reflects the variation in volume variance along the direction of the corresponding eigenvector. The residual coding module 20 extracts the eigenvector corresponding to the eigenvalue with the largest value and establishes this eigenvector as the dominant direction of the residual.

[0106] To avoid uncertainty in the dominant direction of the residuals, when the second-order covariance residual matrix is ​​a zero matrix or the largest eigenvalue approaches 0 (e.g., less than 10), -5 When matrix singularity issues arise, the system defaults to setting the dominant residual direction as a preset unit vector in the global coordinate system (e.g., [1,0,0]). Extracting the dominant residual direction achieves mathematical dimensionality reduction of the residual matrix, enabling the system to use only one three-dimensional vector to indicate the stretching direction in which the volume loss of the removed Gaussian volume is most significant in space.

[0107] S230, Generate a backup truncated residual tensor signature based on cross-boundary transfer. In this embodiment, the system introduces a cross-boundary transfer mechanism when processing data at the edge of a spatial block. The residual encoding module 20 calculates the vertical distance from the 3D center coordinates of each supporting Gaussian to the physical boundary of its spatial block. When it is determined that the vertical distance is less than a set boundary distance threshold, the residual encoding module 20 marks the supporting Gaussian as a boundary Gaussian.

[0108] As a preferred approach, the boundary distance threshold is typically set to 1.5 to 2.0 times the maximum bounding box radius of the Gaussian volume. This range is based on empirical values ​​of the truncation attenuation caused by the camera's view frustum at the tail of the Gaussian distribution, ensuring coverage of the Gaussian volume range where cross-block visual intersections may occur.

[0109] For supporting Gaussians marked as boundary Gaussians, the residual encoding module 20 generates, in addition to generating truncated residual tensor signatures containing only truncated data within the current spatial block, alternative truncated residual tensor signatures for each shared boundary surface. When calculating the alternative truncated residual tensor signatures, the residual encoding module 20 includes dependent Gaussians in adjacent spatial blocks that have a spatial wrapping relationship with the boundary Gaussian in the cumulative calculation of the multi-order basic residuals. Specifically, the system directly reuses the aforementioned spatial wrapping degree calculation formula under typical field-of-view constraints to confirm the determination of spatial wrapping relationships.

[0110] The generation of the alternative truncated residual tensor signature differs from that of the truncated residual tensor signature in terms of computational data source. The alternative truncated residual tensor signature serves as offline preparation data for the proxy representation of adjacent blocks. Its technical purpose is that, during the rendering phase, when adjacent spatial blocks undergo high-level truncation to remove their internal dependent Gaussians, the boundary Gaussian of the current spatial block can activate the alternative truncated residual tensor signature, thereby offsetting the seams and visual energy holes generated at the boundaries of adjacent blocks through its own morphological compensation.

[0111] S240, encapsulate the tensor signature data and attach it to the vertex attributes. To meet the reading specifications of the graphics application programming interface, the residual encoding module 20 performs structured encapsulation of the generated signature data. The residual encoding module 20 packages the calculated 0th-order energy residual, residual dominant direction, and appearance color residual according to the set floating-point bit width to form a compressed storage data block of truncated residual tensor signature; wherein, the centroid offset vector is used to determine the shared boundary surface corresponding to the spare truncated residual tensor signature, and is not a required field in the real-time compensation stage of this embodiment.

[0112] As a preferred approach, the residual encoding module 20 writes the truncated residual tensor signature, the alternative truncated residual tensor signature, and the corresponding adjacent spatial block identifier, shared boundary surface identifier, and boundary index as additional vertex attributes into the vertex attribute cache containing the Gaussian support. During the write operation, the residual encoding module 20 follows the memory alignment rules of four-dimensional floating-point vectors to ensure that the starting address of the tensor signature data meets the 16-byte alignment requirement.

[0113] This memory alignment method can avoid unaligned memory access penalties when the compute shader reads global video memory, ensuring the efficiency of data reading during the real-time rendering stage.

[0114] Reference Figure 5 In this embodiment, during the real-time rendering stage, in order to rely on global state reading and comparison logic to achieve lock-free transfer of cross-block residual compensation rights and avoid memory access conflicts, the above step S300 may specifically include the following sub-steps:

[0115] S310, calculate the joint state machine output and convert it into a truncation level. In this embodiment, the state arbitration module 30 obtains the current viewpoint parameters and spatial block states. In order to dynamically determine the data culling degree of each independent spatial block and establish an adaptive allocation mechanism for rendering resources, the state arbitration module 30 calculates the state function value of the spatial block.

[0116] The state function calculation incorporates parameters such as camera distance, viewpoint velocity, and projected area variance. To ensure dimensional consistency of the various physical parameters during weighted summation, the system performs normalization on each variable. Specifically, the state function calculation formula is as follows:

[0117] ;

[0118] In the formula, The state function value of the spatial block is represented by a dimensionless scalar. This represents the Euclidean distance from the current camera center to the geometric center of the spatial block, and the unit can be meters. Indicates the maximum visible depth of the scene, which can be expressed in meters (m). To prevent extremely small constants with a denominator of 0, for example, a value of 10 is used. -5 ; This indicates the current camera viewpoint movement speed, which can be expressed in m / s. This indicates the theoretical maximum movement speed of the camera during roaming in the current scene, and the unit can be m / s; This represents the variance of the projected area of ​​the Gaussian volume within the spatial block in the current frame, obtained according to the current camera parameters. The variance of the projected area in the previous frame is used for occlusion freeze gate comparison. The unit is the square of pixels. This represents the variance of the theoretical maximum projected area based on the monitor resolution, expressed in squared pixels. , , These represent the distance weighting coefficient, velocity weighting coefficient, and variance weighting coefficient, respectively, all of which are dimensionless parameters.

[0119] In this embodiment, a minimum constant is introduced. This aims to avoid division operations crashing when the normalized denominator abnormally approaches 0, thus ensuring the integrity of the algorithm logic. As a preferred approach, the sum of the above three weighting coefficients is limited to 1, and the specific values ​​can be pre-allocated according to the computing power limitations and rendering performance requirements of the target hardware platform.

[0120] The technical purpose of this computational logic is to quantify the geometric position and dynamic changes of a spatial block relative to the camera into a single evaluation index, and to ensure that each sub-item is converted into a dimensionless percentage value in the range of 0 to 1, so as to avoid numerical scale imbalance caused by different physical dimensions.

[0121] In order to map continuous state function values ​​to discrete control parameters, the state arbitration module 30 obtains the state function values ​​of the spatial block and then discretizes and maps them according to the set interval segmentation threshold to obtain the cutoff level value of the current spatial block.

[0122] As a preferred approach, the interval segmentation threshold is obtained by equally dividing the continuous state space from 0 to 1 according to the total number of layers of the target detail level. The truncation level value determines the truncation ratio of the Gaussian subset to be removed within the aforementioned spatial block.

[0123] S320 performs occlusion freeze gating and global state memory updates. When scene objects move rapidly or occlusion relationships change abruptly, relying solely on state function values ​​may cause frequent jumps in truncation levels between adjacent frames, leading to screen flickering. To maintain visual stability, the state arbitration module 30 introduces occlusion freeze gating logic.

[0124] In practice, the state arbitration module 30 compares the variance of the projected area between the current frame and the previous frame. When the absolute value of the difference in the variance of the projected area between the two frames is determined to be greater than the set fluctuation threshold, the system triggers the occlusion freeze gating mechanism.

[0125] As a preferred approach, the fluctuation threshold is set to 0.5 to 0.8 times the variance of the projected area in the previous frame. This range is derived from the statistical patterns of projected area fluctuations under typical abrupt occlusion scenarios.

[0126] When the gating mechanism is triggered, the state arbitration module 30 forcibly maintains the truncation level value of the current spatial block in the previous frame, suppressing level jumps caused by sudden occlusion.

[0127] After completing the truncation level determination, the state arbitration module 30 writes the truncation level values ​​of each spatial block obtained from the current calculation into a pre-allocated global state memory area. After the truncation level values ​​for this frame are written, this global state memory area is configured as a state query data source for subsequent shader read-only access.

[0128] S330, Perform local state comparison without cross-block addressing. In this embodiment, during the computational shading stage before rasterization, the computational shader needs to handle residual compensation across boundaries. To obtain the state information of adjacent blocks, the system avoids the operation of directly accessing the large vertex data buffer of adjacent spatial blocks.

[0129] When processing the boundary Gaussian of the current spatial block, the computation shader extracts the truncation level values ​​of the adjacent spatial blocks that are physically adjacent to the current spatial block by reading data from the global state memory area.

[0130] By using this local state reading logic, the system can obtain the truncated state of adjacent blocks without performing cross-block underlying vertex data addressing, thus reducing the bandwidth overhead and latency cost of video memory access.

[0131] S340, Execute Dominance Arbitration and Activate Backup Truncated Residual Tensor Signature Proxy. After obtaining the truncation level values ​​of the current spatial block and adjacent spatial blocks, the state arbitration module 30 executes the state-aware boundary arbitration mechanism based on the numerical relationship between the two. Considering that a spatial block may have multiple interface surfaces, the system independently performs local state comparison and dominance arbitration for each shared boundary surface of the current spatial block.

[0132] When the truncation level value of the current spatial block is determined to be less than that of the adjacent spatial block, it indicates that the current spatial block retains more detailed data. The state arbitration module 30 thus determines that the current spatial block has the dominant right to represent details. When the truncation level value of the current spatial block is equal to that of the adjacent spatial block, the dominant right to represent details is determined according to a preset unique rule: the side with the smaller spatial block identifier value on both sides of the shared boundary has the dominant right to represent details.

[0133] Under this control branch, the system activates the truncated residual tensor signature carried by the Gaussian of the current spatial block boundary and performs residual compensation calculation locally.

[0134] Conversely, when the truncation level value of the current spatial block is greater than that of the adjacent spatial block, it indicates that the adjacent spatial block is at a more refined level of detail. The state arbitration module 30 determines that the adjacent spatial block has the dominant right to express detail.

[0135] Under this condition, the system suppresses the local residual compensation calculation process of the Gaussian boundary of the current spatial block. At the same time, the state arbitration module 30 locates and triggers the backup truncated residual tensor signature of the corresponding boundary Gaussian in the adjacent spatial block based on the pre-stored adjacent spatial block identifier, shared boundary surface identifier, and boundary corresponding index, and executes the proxy compensation logic through the backup truncated residual tensor signature.

[0136] This activation mechanism based on truncation level comparison ensures a one-way and unique transfer of cross-block residual compensation rights, avoiding memory read / write conflicts caused by adjacent spatial blocks simultaneously performing overlapping compensation on the boundary area.

[0137] Reference Figure 6 In this embodiment, before entering the rasterization stage, in order to utilize the activated tensor signature data to perform morphological reconstruction and equivalent transmittance correction supporting Gaussianism, the above step S400 may specifically include the following sub-steps:

[0138] S410, perform covariance tensor outer product stretching based on the dominant direction. In this embodiment, before entering the rasterization stage, the directional compensation module 40 performs morphological reconstruction on Gaussian-supporting data using activated truncated residual tensor signatures or backup truncated residual tensor signature data. The directional compensation module 40 extracts the 0th-order energy residual from the truncated residual tensor signatures or backup truncated residual tensor signatures and uses this as a reference to extract the expansion coefficient.

[0139] Specifically, the directional compensation module 40 multiplies the 0th-order energy residual by a set dilation scaling constant to obtain the dilation coefficient. As a preferred approach, this dilation scaling constant is typically between 0.5 and 1.0, and its specific value is preset through offline visual evaluation of typical scenes.

[0140] To ensure the consistency of physical dimensions in subsequent tensor matrix addition operations, the directional compensation module 40 extracts the largest eigenvalue of the Gaussian initial three-dimensional covariance matrix as the spatial scale reference factor.

[0141] After obtaining the expansion coefficient and spatial scale reference factor, the directional compensation module 40 performs a tensor outer product stretching operation on the Gaussian-supporting 3D covariance matrix along the residual dominance direction in the truncated residual tensor signature or the alternative truncated residual tensor signature. The formula for covariance tensor outer product stretching is:

[0142] ;

[0143] In the formula, This represents the updated three-dimensional covariance matrix, whose dimension is the square of the length; This represents the three-dimensional covariance matrix supporting a Gaussian initialization, with dimensions equal to the square of its length. This represents the calculated expansion coefficient, which is a dimensionless parameter greater than or equal to 0. This represents the largest eigenvalue of the three-dimensional covariance matrix that supports the Gaussian initialization, with the dimension being the square of the length; This represents the column vector of the dominant direction of the three-dimensional residual extracted from the truncated residual tensor signature or the alternative truncated residual tensor signature. It is a dimensionless unit vector. This indicates the transpose of the column vector; The tensor outer product of the column vectors representing the dominant direction of the residuals forms a dimensionless matrix of 3 rows and 3 columns.

[0144] The technical objective of this operation is to use an incremental matrix in a single direction to stretch the volume supporting the Gaussian in physical three-dimensional space by a specific proportion along the direction in which the loss of the removed Gaussian subset is most significant, thereby achieving directional compensation of the geometric shape.

[0145] By introducing the largest eigenvalue as the multiplier term, the system eliminates the dimensional mismatch problem and ensures the positive semidefiniteness of the increment matrix, thus avoiding the violation of the mathematical validity of the covariance matrix.

[0146] S420, extract the equivalent volume footprint ratio. After completing the stretching update of the covariance matrix, in order to quantify the impact of morphological changes on light transmission characteristics, the directional compensation module 40 calculates the volume change rate supporting Gaussian before and after stretching.

[0147] Considering that the determinant of the covariance matrix mathematically represents the product of the variances of each axis, the orientation compensation module 40 calculates the determinant of the updated three-dimensional covariance matrix and the determinant of the three-dimensional covariance matrix that supports Gaussian initialization.

[0148] In the calculation execution logic, the directional compensation module 40 adds the determinant of the initial three-dimensional covariance matrix to the minimum anti-overflow constant to obtain the baseline determinant value.

[0149] Subsequently, the directional compensation module 40 divides the determinant of the updated three-dimensional covariance matrix by the base determinant value, and performs a square root operation on the result of the division operation to obtain the equivalent volume expansion coefficient.

[0150] Introduce a very small overflow prevention constant (e.g., a value of 10). -6 This aims to avoid division errors caused by the determinant approaching zero when the initial Gaussian volume degenerates into an extremely thin plane or line segment, thus ensuring the integrity of the algorithm logic. The equivalent volume expansion coefficient characterizes the relative linear magnification factor of the three-dimensional volume occupied by the supporting Gaussian after directional compensation.

[0151] S430, perform nonlinear transparency correction based on equivalent optical thickness constraints. In this embodiment, the volume expansion of a Gaussian volume in three-dimensional space leads to a sparser distribution of matter within it. Under these conditions, if the system directly maintains the original opacity of the Gaussian volume, it will cause abnormal accumulation of local visual energy in the rendered image, resulting in local overexposure. Therefore, the linear conversion of opacity caused by volume expansion will fail in the physically based volumetric ray casting model.

[0152] To maintain consistent visual coverage energy, the directional compensation module 40 performs a nonlinear correction to the Gaussian-supporting opacity based on the equivalent volume expansion coefficient. The directional compensation module 40 constructs an exponential dilution formula based on the equivalent optical thickness constraint to obtain the corrected opacity. The exponential dilution formula is:

[0153] ;

[0154] In the formula, This represents the corrected opacity, which is a dimensionless parameter in a closed interval between 0 and 1. This indicates the opacity supporting a Gaussian initial value, which is a dimensionless parameter on a closed interval between 0 and 1. This represents the calculated equivalent volume expansion coefficient, which is a dimensionless value greater than or equal to 1.

[0155] The technical purpose of this exponential dilution formula is to proportionally reduce the optical absorption cross-section per unit length in space, while adhering to the Beer-Lambert optical absorption law, under the condition that the Gaussian volume increases and the desired path of light passing through the Gaussian volume increases.

[0156] Through this nonlinear correction step, the system ensures that the overall energy attenuation of the Gaussian body after light penetration and update remains consistent with that before the stretching operation, thus avoiding visual artifacts caused by truncation compensation.

[0157] S440 performs appearance transfer projection and rasterization hybrid output. After completing the physical correction of geometry and transparency, the orientation compensation module 40 performs compensation superposition of the base color. The orientation compensation module 40 extracts the appearance color residual from the truncated residual tensor signature or the alternative truncated residual tensor signature, and performs corresponding vector addition operation with the 0th order spherical harmonic coefficients in the supporting Gaussian vertex attribute to realize the update operation of the base diffuse color.

[0158] To prevent out-of-bounds anomalies after color energy superposition, the directional compensation module 40 performs a numerical clamping operation on the result of the addition operation, limiting the updated 0th order spherical harmonic coefficients to the effective numerical range of the preset basic color space, thus avoiding numerical overflow in the rendering pipeline.

[0159] Subsequently, the updated 3D covariance matrix, corrected opacity, and updated 0th-order spherical harmonic coefficients are fed into the screen-space rendering pipeline. The orientation compensation module 40 projects the updated 3D covariance matrix into a 2D screen covariance matrix through an affine transformation based on the current camera's view projection matrix.

[0160] For the camera view projection transformation of Gaussian volume from three-dimensional world space to two-dimensional screen space and the final rasterization process, those skilled in the art can refer to the standard Gaussian splash rendering pipeline. Its camera matrix transformation and depth-ordered hybrid shading mechanism are well-known technologies in the field and will not be elaborated here.

[0161] Finally, the rendering pipeline performs Gaussian blending based on corrected opacity on all Gaussian-enabled pixels on the screen in front-to-back depth space order, outputting a final color image with continuous levels of detail.

[0162] The present invention also provides a computer device, including: a processor and a memory, the memory storing a computer program executable by the processor, the computer program performing the method described above when executed by the processor.

[0163] The present invention also provides a storage medium storing a computer program, which is executed by a processor to perform the method described above.

[0164] The storage medium can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as Static Random Access Memory (SRAM), Electrically Erasable Programmable Read-Only Memory (EEPROM), Erasable Programmable Read-Only Memory (EPROM), Programmable Red-Only Memory (PROM), Read-Only Memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk.

[0165] To further aid in understanding the present invention, the following specific application example of a large-scale UAV oblique photography of a city-level 3D scene is used to illustrate the actual operation performance and experimental verification of the system.

[0166] Taking a Gaussian splash reconstruction model of a downtown street as an example, the original scene contains approximately 85 million Gaussian volumes, resulting in an extremely large data scale. If full rendering is used, a high-end commercial graphics card with a standard 24GB of video memory would face the risk of memory overflow and crash.

[0167] The system first performs redundancy assessment and residual coding on the city block during the preprocessing stage. Combined with... Figure 7The 3D scatter plot shown demonstrates how the system divides the entire street into a uniform spatial grid based on Morton codes. When processing a densely tree-lined area, numerous overlapping micro-Gaussian volumes within the tree canopy are accurately identified by the redundancy evaluation module 10. The system extracts the largest Gaussian volume on the outer side of the canopy with the highest exclusive coverage contribution as the supporting Gaussian, while marking the occluded Gaussian volumes within as dependent Gaussians. After the culling decision is completed, the residual encoding module 20 performs feature decomposition on the geometric shape and color differences lost from the internal leaves, extracting the dominant residual direction along the extension direction of the canopy branches and leaves, and packaging the energy and color compensation data into the vertex attributes of the outer supporting Gaussian. During rendering stretching, the system performs nonlinear inverse compensation on the opacity of the supporting Gaussian based on the equivalent optical thickness constraint formula, ensuring that the expanded canopy fills the visual gaps left after the internal Gaussian volume culling without causing distortion due to linear transparency. Figure 7 The China-Israel mesh network visually illustrates the star-shaped responsibility topology and the dominant direction vector attached to the central support node.

[0168] During the real-time street walkthrough rendering phase, the camera dives down from a height of thousands of meters to the ground-level streets. From this high-altitude perspective, the state arbitration module 30 calculates an extremely small variance in the projected area, assigning the highest level of truncation to most street blocks. The Gaussian attachments of numerous building facades and vegetation are directly intercepted in video memory by the graphics API's prefix instance rendering instructions, eliminating the need for them to enter the subsequent rendering pipeline.

[0169] When the camera crosses the boundary between two spatial blocks, the state arbitration module 30 reads the global state memory and finds that the truncation level of the left block is higher than that of the right block. The system immediately triggers the dominance arbitration rule, suppressing the local compensation of the Gaussian at the right block boundary and forcibly activating the backup truncation residual tensor signature carried by the Gaussian at the right boundary. Based on this proxy signature, the directional compensation module 40 performs cross-boundary stretching on the covariance matrix of the right boundary along a preset dominant direction and simultaneously dilutes its opacity. This proxy operation completely stitches up the visual seam between the left and right blocks caused by the different truncation levels.

[0170] To verify the absolute computing power advantage and performance of this system in a real rendering pipeline, this solution was rigorously compared with the baseline model (Vanilla 3DGS, full rendering without LOD) and the traditional discrete LOD model (Octree 3DGS, node visibility switching based on octree) through testing.

[0171] Combination Figure 8The performance comparison graph shows that the horizontal axis represents the camera's flight path over 2000 frames in the city. Due to the massive amount of data, the baseline model consistently approached a dangerous 23GB of VRAM usage (right Y-axis), and the rendering frame rate (left Y-axis) plummeted to around 18 FPS in dense commercial areas. While the traditional discrete LOD model kept VRAM usage around 14GB, the frame rate curve exhibited severe jagged fluctuations when crossing level thresholds. After applying this solution, the system's VRAM usage was stably compressed to the 8GB-10GB range. More importantly, by avoiding complex octree traversal addressing and cross-block VRAM copying, the system's frame generation time variance was extremely small. In tests unlocking the frame rate cap, the native rendering frame rate directly exceeded 100 FPS, and under normal operating conditions, it consistently adhered to the 60 FPS vertical synchronization cap, completely eliminating any stuttering.

[0172] In terms of evaluating the quality of rendered images, combined with Figure 9 The system uses a clustered histogram to extract peak signal-to-noise ratio (PSNR) and structural similarity (SSIM) data from three typical test sets: commercial areas, residential areas, and parks / green spaces. Although this system discards nearly 65% ​​of the tiny Gaussian volumes in the foreground, thanks to the morphological orientation reconstruction dominated by residuals and the scientific constraints of equivalent optical thickness, its PSNR only decreased by less than 0.4 dB compared to the baseline full-rendering model, remaining steadily above 32 dB in the high-fidelity range; the SSIM also remained stable at around 0.92. In contrast, the traditional discrete LOD model directly deletes nodes at low levels, resulting in severe loss of high-frequency textures. In actual testing, its PSNR in residential areas plummeted to 27 dB, accompanied by visible structural voids and flickering artifacts.

[0173] The application examples and cross-comparison experiments clearly demonstrate that the topology-driven prefix truncation and cross-block residual proxy compensation mechanism designed in this solution can achieve a significant release of video memory and a leap in frame rate with extremely low image quality loss when dealing with massive Gaussian splash data, fundamentally solving the video memory bandwidth bottleneck in real-time shading of large-scale 3D scenes.

[0174] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A switching LOD rendering method based on massive Gaussian splash model data, characterized in that, include: Within the spatial block, a structural redundancy responsibility graph is established for supporting Gaussians and dependent Gaussians. Based on the structural redundancy responsibility graph, a prefix-truncation rearrangement is performed on supporting Gaussians and dependent Gaussians. According to the truncation ratio setting, a culling operation is performed on the attached Gaussian, a truncated residual tensor signature is generated for the culled attached Gaussian, and an alternative truncated residual tensor signature facing the shared boundary surface is generated for the supporting Gaussian at the boundary position. During the real-time rendering phase, the truncation level values ​​of the current spatial block and adjacent spatial blocks are obtained, and the dominance arbitration is performed based on the truncation level values ​​of the current spatial block and adjacent spatial blocks. Activate the truncated residual tensor signature corresponding to the Gaussian support based on the arbitration result, or activate the alternative truncated residual tensor signature to perform cross-boundary proxy compensation; Based on the activated truncated residual tensor signature or the alternate truncated residual tensor signature, directional covariance compensation and appearance color compensation are performed on the corresponding Support Gaussian input rendering pipeline, and the compensated Support Gaussian input rendering pipeline is used for rendering output.

2. The switching LOD rendering method based on massive Gaussian splash model data according to claim 1, characterized in that, Within a spatial block, a structural redundancy responsibility graph is established for supporting Gaussians and dependent Gaussians. Based on this structural redundancy responsibility graph, a prefix-truncation rearrangement is performed on the supporting Gaussians and dependent Gaussians, including: The three-dimensional scene is divided into multiple spatial grid units based on Morton coding and a preset Morton hierarchy to construct the spatial blocks; Through iterative screening, the Gaussian body with the highest exclusive coverage contribution value in each spatial block is established as the supporting Gaussian body, and the other Gaussian bodies in the spatial block other than the supporting Gaussian body are selected as candidate Gaussian bodies. When the spatial envelopment parameter of the candidate Gaussian body is greater than the envelopment threshold and the intrinsic appearance difference parameter is less than the intrinsic appearance difference threshold, the candidate Gaussian body is established as a dependent Gaussian belonging to the corresponding supporting Gaussian, and the structural redundancy responsibility graph is established. In the contiguous video memory array, the supporting Gaussian arrangement is placed at the starting position, and the corresponding dependent Gaussian arrangement is placed after the supporting Gaussian arrangement to complete the prefix truncation rearrangement.

3. The switching LOD rendering method based on massive Gaussian splash model data according to claim 1, characterized in that, For the discarded Gaussian-dependent truncated residual tensor signatures, the following are included: Extract the 0th-order energy residual, 2nd-order covariance residual matrix, and appearance color residual of the removed attached Gaussian relative to the corresponding supporting Gaussian; Perform eigenvalue decomposition on the second-order covariance residual matrix to extract the dominant residual direction; The 0th-order energy residual, appearance color residual, and residual dominant direction are packaged to generate a truncated residual tensor signature, and the truncated residual tensor signature is written into the Gaussian-supporting vertex attributes.

4. The switching LOD rendering method based on massive Gaussian splash model data according to claim 1, characterized in that, For the alternative truncated residual tensor signature supporting Gaussian generation oriented towards the shared boundary surface at the boundary location, it includes: The supporting Gaussian whose vertical distance to the physical boundary of its spatial block is less than the boundary distance threshold is identified as the supporting Gaussian at the boundary position. The dependent Gaussian that has a spatial wrapping relationship with the supporting Gaussian at the boundary position in the adjacent spatial block is included in the residual accumulation calculation range to generate the spare truncated residual tensor signature. The signature of the spare truncated residual tensor is written into the Gaussian-supporting vertex attribute at the boundary position by combining the boundary corresponding index.

5. The switching LOD rendering method based on massive Gaussian splash model data according to claim 1, characterized in that, During the real-time rendering phase, obtain the truncation level values ​​of the current spatial chunk and adjacent spatial chunks, including: The state function value of the current spatial block is calculated based on the camera distance, viewpoint velocity, and projected area variance. The state function value is mapped to a truncation level value, and the truncation level value is written into the global state memory area. When the absolute value of the difference between the variance of the projected area of ​​the current frame and the previous frame is greater than the fluctuation threshold, the occlusion freeze gating mechanism is triggered to force the truncation level value of the previous frame to remain unchanged.

6. The switching LOD rendering method based on massive Gaussian splash model data according to claim 5, characterized in that, Based on the truncation level values ​​of the current spatial block and adjacent spatial blocks, a dominance arbitration is performed, including: The truncation level values ​​of adjacent spatial blocks are extracted by reading the global state memory area; When the truncation level value of the current spatial block is less than the truncation level value of the adjacent spatial block, it is determined that the current spatial block has the dominant right to express details. When the truncation level value of the current spatial block is equal to the truncation level value of the adjacent spatial block, the ownership of the detailed expression dominance is determined according to a preset unique rule.

7. The switching LOD rendering method based on massive Gaussian splash model data according to claim 6, characterized in that, Activating the truncated residual tensor signature corresponding to the Gaussian support based on the arbitration result, or activating the alternative truncated residual tensor signature to perform cross-boundary proxy compensation, including: When the current spatial block has the dominant control over the detailed representation, activate the truncated residual tensor signature corresponding to the Gaussian support at the boundary position within the current spatial block; When the adjacent spatial block has the dominant control over the detailed representation, the local residual compensation of the current spatial block is suppressed, and the backup truncated residual tensor signature corresponding to the Gaussian support at the boundary position in the adjacent spatial block is activated to perform cross-boundary proxy compensation.

8. The switching LOD rendering method based on massive Gaussian splash model data according to claim 1, characterized in that, Based on the activated truncated residual tensor signature or the alternate truncated residual tensor signature, perform directional covariance compensation and appearance color compensation on the corresponding Gaussian-supporting tensor, including: Extract the 0th-order energy residual and residual dominance direction from the activated truncated residual tensor signature or the alternative truncated residual tensor signature; The 0th-order energy residual is converted into an expansion coefficient. Using the expansion coefficient and the largest eigenvalue of the supporting Gaussian 3D covariance matrix, a covariance tensor outer product stretching operation is performed on the supporting Gaussian 3D covariance matrix along the dominant direction of the residual to obtain the updated 3D covariance matrix.

9. The switching LOD rendering method based on massive Gaussian splash model data according to claim 8, characterized in that, Performing directional covariance compensation and appearance color compensation on the corresponding Gaussian support also includes: The equivalent volume expansion coefficient is obtained by taking the square root of the ratio between the determinant of the updated three-dimensional covariance matrix and the determinant of the three-dimensional covariance matrix supporting Gaussian initialization. Based on the equivalent optical thickness constraint, the opacity supporting Gaussian is nonlinearly corrected using the equivalent volume expansion coefficient. Extract the appearance color residual from the activated truncated residual tensor signature or the backup truncated residual tensor signature, and superimpose the appearance color residual onto the 0th-order spherical harmonic coefficients that support Gaussians to obtain the updated 0th-order spherical harmonic coefficients.

10. The switching LOD rendering method based on massive Gaussian splash model data according to claim 9, characterized in that, After performing directional covariance compensation and appearance color compensation on the corresponding Gaussian support, it also includes: The supported Gaussian projection, which includes the updated 3D covariance matrix, the nonlinearly corrected opacity, and the updated 0th-order spherical harmonic coefficients, is projected onto the 2D screen space. The image is then subjected to Gaussian rasterization blending shading processing and output after being projected onto the two-dimensional screen space.

Images