A method and system for generating detail levels of Gaussian splashing

Abstract

This invention provides a method and system for generating Gaussian splatter detail levels, comprising: setting N levels of detail, configuring resource control parameter groups {, ,} for each level, where: is the upper limit of the number of Gaussians at the i-th level, satisfying β1<β2<...<β N For denser triggering cycles, satisfying 1>2>...> N The downsampling rate for training images must satisfy 1 < 2 < ... < ... N =1; Perform hierarchical iterative training from i=1 to N: Starting from layer 1, perform: initialize the training environment based on the current layer parameter set; control the number of Gaussians to grow to a certain value through a dynamic growth function; trigger multi-factor densification decision when the period is reached; output the current layer LOD model and iterate to the next layer. This invention constructs a hierarchical parameter control mechanism that synchronously adjusts the number of Gaussian models, the densification rhythm, and the image resolution during training iterations, enabling the progressive generation from low-precision base to high-precision details to be completed in a single training session, significantly improving LOD generation efficiency.

Description

Technical Field

[0001] This invention belongs to the field of 3D reconstruction technology, and in particular relates to a method and system for generating Gaussian splash detail levels. Background Technology

[0002] In recent years, 3D Gaussian Splatting (3DGS) has become an important method in the field of 3D reconstruction due to its high image quality and real-time rendering capabilities. 3DGS represents the scene through an explicit 3D Gaussian distribution and, combined with rasterization rendering, eliminates the need for complex volume rendering operations, greatly improving rendering efficiency and achieving remarkable results in small-scale scenes.

[0003] With the increasing demand for city-level 3D reconstruction, researchers have begun to explore extending 3DGS to large-scale scenes. Current mainstream methods for generating LOD (Level of Detail) for large-scale scenes all employ a post-processing approach, first completing full-precision scene reconstruction and then generating multi-level models through pruning (CityGaussian) or hierarchical-3DGS merging. This approach suffers from at least the following drawbacks:

[0004] 1. Separation of training and resource control: Existing methods all require a complete reconstruction before post-processing to generate the LOD hierarchy, making it difficult to precisely control resource usage during the training phase, resulting in low training efficiency and unpredictable resource overhead.

[0005] 2. The post-processing steps are complex and redundant: Pruning and merging are both performed after reconstruction, requiring independent execution of multiple stages of reconstruction → pruning → merging. A large number of Gaussian optimizations in the early stage may be discarded in post-processing, resulting in a waste of resources.

[0006] 3. LOD accuracy is inflexible and relies on heuristic pruning or aggregation: It lacks an automatic adjustment mechanism to cope with different hardware resources or application requirements;

[0007] 4. Insufficient scalability: When facing city-scale scenarios, the combination of partitioned training and hierarchical management is not mature enough, resulting in high training difficulty and high inference load. Boundary stitching needs additional optimization. Summary of the Invention

[0008] To overcome the shortcomings of existing technologies, the present invention aims to provide a method and system for generating Gaussian splash detail levels, which directly generates multi-level LOD models during the training phase, avoids post-processing, achieves coordinated control of Gaussian quantity, density, and resolution, supports partitioned parallel training and seamless stitching of city-level scenes, and solves the problems that traditional technologies cannot meet the dual requirements of resource management and efficiency for large-scale city-level scene reconstruction.

[0009] To solve the above problems, the technical solution adopted by the present invention is as follows:

[0010] In a first aspect, the present invention provides a method for generating Gaussian splash detail levels, comprising:

[0011] Set N levels of detail, and configure resource control parameter groups at each level { , , },in:

[0012] Let β1 be the upper bound of the number of Gaussians of order i, satisfying β1 < β2 < ... < β2. N ;

[0013] To achieve a denser trigger cycle, satisfying 1> 2>...> N ;

[0014] To achieve the desired image downsampling rate during training, the following conditions must be met. 1< 2<...< N =1;

[0015] Perform hierarchical iterative training from level i=1 to level N:

[0016] Starting from level 1, execute:

[0017] Initialize the training environment based on the current layer parameter set;

[0018] The Gaussian quantity is controlled to grow using a dynamic growth function. ;

[0019] When reached Periodic triggering of multi-factor dense decision-making;

[0020] Output the LOD model of the current layer and iterate to the next layer.

[0021] In some embodiments, the dynamic growth function is configured as follows:

[0022] In the iteration number x∈[0,K] i Within the range, the number of Gaussians is controlled to approximate an accelerating growth curve. ;

[0023] The accelerated growth curve adopts a nonlinear growth model:

[0024] in, The initial Gaussian number, The maximum number of iterations for the i-th layer. It is a monotonically increasing function from [0,1] to [0,1].

[0025] In some embodiments, the To accelerate the growth function:

[0026] =

[0027] in, ∈[0,1].

[0028] In some embodiments, the rate of change of the derivative of the accelerating growth function is such that the growth rate is maximum at x=0, and x= The time is reduced to 0, achieving a balance between rapid expansion in the early stage and gradual convergence in the later stage.

[0029] In some embodiments, the multi-factor denser decision includes:

[0030] Calculate the composite importance score of Gaussian points:

[0031]

[0032] in, The number of pixels hit. To accumulate transparency, For spatial volume, The magnitude of the image gradient;

[0033] according to Fractional probability sampling of Gaussian points is used to perform densification.

[0034] In some embodiments, when the number of Gaussians is controlled by a dynamic growth function, and dense training is enabled for M iterations per layer, ;

[0035] When calculating the composite importance score of Gaussian points, the scores are used as weights for sampling. Perform a densification operation on each Gaussian point until the resource budget limit for this level is reached;

[0036] Among them, based on The calculation results set an upper limit for single-time dense sampling.

[0037] In some embodiments, a cross-level inheritance mechanism is also included:

[0038] The training of the (i+1)th layer uses the output of the previous layer as the initialization starting point;

[0039] Inheriting the previous Gaussian points and based on β i+1 Perform incremental densification.

[0040] In some embodiments, the method supports partitioned collaborative training:

[0041] Divide the scene into P spatial sub-regions;

[0042] Each spatial sub-region independently performs the described hierarchical iterative training;

[0043] Generate a set of LOD models with hierarchical labels {LOD} i^p}, i∈[1,N], p∈[1,P].

[0044] In some embodiments, the boundary regions of adjacent spatial sub-regions are defined, and a shared Gaussian point set is established for the boundary regions;

[0045] The importance of the Gaussian points assigned at the boundaries of adjacent partitions is assessed.

[0046] The associated partition metadata is updated synchronously when performing the densification operation of the boundary Gaussian points.

[0047] Secondly, the present invention provides a Gaussian splash detail level generation system, which applies the Gaussian splash detail level generation method described above, including:

[0048] The hierarchy setting module is used to set N levels of detail, with each level configuring resource control parameter groups { , , },in:

[0049] Let β1 be the upper bound of the number of Gaussians of order i, satisfying β1 < β2 < ... < β2. N ;

[0050] To achieve a denser trigger cycle, satisfying 1> 2>...> N ;

[0051] To achieve the desired image downsampling rate during training, the following conditions must be met. 1< 2<...< N =1;

[0052] The hierarchical iterative training module is used to perform hierarchical iterative training from level i=1 to level N:

[0053] Starting from level 1, execute:

[0054] Initialize the training environment based on the current layer parameter set;

[0055] The Gaussian quantity is controlled to grow using a dynamic growth function. ;

[0056] When reached Periodic triggering of multi-factor dense decision-making;

[0057] The LOD model output module is used to output the current layer's LOD model and iterate to the next layer.

[0058] Compared with the prior art, the present invention has at least the following beneficial effects:

[0059] By embedding the level-of-detail (LOD) generation process into the training framework, this invention fundamentally changes the traditional 3D reconstruction approach of first creating full-precision models and then post-processing to reduce precision. The core breakthrough lies in constructing a hierarchical parameter control mechanism that synchronously adjusts the number of Gaussian models, the density increase rate, and the image resolution during training iterations. This allows for the gradual generation of details from a low-precision base to high-precision details to be completed in a single training session. This integrated process eliminates the computational resource waste caused by traditional pruning or merging operations, significantly improving LOD generation efficiency.

[0060] Regarding resource control precision, the nonlinear growth model proposed in this invention solves the industry problem of unpredictable resource consumption in large-scale scene training. This model, through a unique growth curve design, rapidly approaches the preset upper limit of the Gaussian number in the early stages of training to accelerate scene structure construction, while automatically slowing down the growth rate in the later stages to achieve smooth convergence, resulting in a smaller deviation between actual resource consumption and the preset budget. In particular, this mechanism ensures that models at different precision levels maintain structural consistency while strictly meeting the memory limitations of heterogeneous platforms such as mobile terminals and desktop workstations.

[0061] This multi-factor dense decision-making architecture is the first to integrate real-time feedback data from the rendering pipeline with deep learning gradient information to comprehensively evaluate the importance of Gaussian points. Specifically, the pixel hit rate and cumulative transparency values ​​generated during rasterization effectively ensure the integrity of the basic structure, while the image spatial gradients obtained through backpropagation precisely enhance the detail restoration of edge contours. This multi-source data collaborative decision-making mode, combined with a spatially aware probabilistic sampling strategy, enables the reconstruction model to significantly improve visual quality while maintaining low resource consumption.

[0062] The present invention will now be described in further detail with reference to the accompanying drawings and specific embodiments. Attached Figure Description

[0063] The present invention will be further described with reference to the accompanying drawings, but the embodiments in the drawings do not constitute any limitation on the present invention. For those skilled in the art, other drawings can be obtained based on the following drawings without creative effort.

[0064] Figure 1 This is a simplified flowchart illustrating a method for generating Gaussian splash detail levels under one implementation.

[0065] Figure 2 This is a schematic diagram illustrating the specific process logic of a Gaussian splash detail level generation method under one implementation method.

[0066] Figure 3 This is a schematic diagram of the framework of a Gaussian splash detail level generation system under one implementation method. Detailed Implementation

[0067] The technical solution of the present invention will now be clearly and completely described with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0068] In the description of this invention, it should be noted that the terms "center," "upper," "lower," "left," "right," "vertical," "horizontal," "inner," and "outer," etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. They are used only for the convenience of describing the invention and for simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on the invention. Furthermore, the terms "first," "second," and "third" are used for descriptive purposes only and should not be construed as indicating or implying relative importance.

[0069] In the description of this invention, when a specific device is described as being located between a first device and a second device, an intermediary device may or may not be present between the specific device and the first or second device. When a specific device is described as being connected to other devices, the specific device may be directly connected to the other devices without an intermediary device, or it may not be directly connected to the other devices but may have an intermediary device.

[0070] Techniques, methods, and equipment known to those skilled in the art may not be discussed in detail, but where appropriate, such techniques, methods, and equipment should be considered part of the specification.

[0071] With the increasing demand for city-level 3D reconstruction, researchers have begun to explore extending 3DGS to large-scale scenes. Representative works include:

[0072] VastGaussian is the first real-time modeling method for large scenes based on 3DGS. Its innovation lies in proposing a progressive spatial partitioning and training strategy based on visibility and coverage, dividing the scene into multiple sub-units, each trained independently, and then merged after optimization. Combined with decoupled appearance modeling, it effectively reduces appearance floating objects during reconstruction. Although VastGaussian improves the reconstruction efficiency of large-scale scenes, it still outputs a single high-precision model and does not achieve multi-level resource-controllable LOD output during training, nor does it possess multi-precision hierarchical training control capabilities.

[0073] On Scaling Up 3D Gaussian Splatting Training (Grendel-GS): Focusing on multi-GPU training efficiency, it proposes a hybrid parallel training strategy and an automatic learning rate scaling mechanism, significantly improving the training throughput in large scenes. However, its core objective is to optimize training efficiency, lacking support for model accuracy control and hierarchical output. Different accuracy models still require manual adjustment or forced pruning after reconstruction.

[0074] CityGaussian: After training, the model undergoes pruning to varying degrees to generate multiple Level-of-Detail (LOD) models with different levels of precision, allowing for the selection of the appropriate model based on view distance during real-time rendering. However, its LOD generation relies entirely on post-processing, with a fixed pruning ratio, lacking precise control over precision and resources.

[0075] Hierarchical-3DGS constructs a hierarchical model from coarse to fine by merging Gaussian nodes into a tree structure, and selects appropriate nodes based on the projection size during the rendering stage. Although this method supports hierarchical selection, the training process is independent of the hierarchy, still requiring post-training merging and optimization. The training phase has no control over resources, making the overall process complex.

[0076] The shortcomings of existing technologies:

[0077] 1. Separation of training and resource control: Existing methods all require a complete reconstruction before post-processing to generate the LOD hierarchy, making it difficult to precisely control resource usage during the training phase, resulting in low training efficiency and unpredictable resource overhead.

[0078] 2. The post-processing steps are complex and redundant: Pruning and merging are both performed after reconstruction, requiring independent execution of multiple stages of reconstruction → pruning → merging. A large number of Gaussian optimizations in the early stage may be discarded in post-processing, resulting in a waste of resources.

[0079] 3. LOD accuracy is inflexible and relies on heuristic pruning or aggregation: It lacks an automatic adjustment mechanism to cope with different hardware resources or application requirements;

[0080] 4. Insufficient scalability: When facing city-scale scenarios, the combination of partitioned training and hierarchical management is not mature enough, resulting in high training difficulty and high inference load. Boundary stitching needs additional optimization.

[0081] In summary, existing LOD (Level of Detail) generation methods all employ post-processing strategies (such as pruning and merging) after reconstruction to generate multi-precision models. This approach has significant technical bottlenecks and struggles to meet the dual requirements of resource management and efficiency for large-scale city-level scene reconstruction.

[0082] In view of this, the applicant proposes a Gaussian splash detail level generation method and system, which controls the number of Gaussians and the optimization cycle of each LOD model end-to-end during the training phase, realizes synchronous management of model complexity and rendering performance, and effectively solves the technical problem that existing technologies cannot accurately generate multi-precision models during the training phase.

[0083] Specifically, refer to Figure 1 and 2 This embodiment 1 provides a method for generating Gaussian splash detail levels, including:

[0084] Step S100: Set N levels of detail, and configure resource control parameter groups for each level { , , },in:

[0085] Let β1 be the upper bound of the number of Gaussians of order i, satisfying β1 < β2 < ... < β2. N , Used to control the maximum resource overhead at each level of detail;

[0086] To achieve a denser trigger cycle, satisfying 1> 2>...> N , Used to control the model densification frequency;

[0087] To achieve the desired image downsampling rate during training, the following conditions must be met. 1< 2<...< N =1, Used to control image resolution during hierarchical training;

[0088] The above three parameters enable each layer to gradually improve from coarse to fine and from low resources to high quality, which is beneficial for controlling resource consumption and gradient optimization direction;

[0089] Step S200: Perform hierarchical iterative training from level i=1 to level N:

[0090] Starting from level 1, execute:

[0091] Initialize the training environment based on the current layer parameter set;

[0092] The Gaussian quantity is controlled to grow using a dynamic growth function. ;

[0093] When reached Periodic triggering of multi-factor dense decision-making;

[0094] Step S300: Output the current layer LOD model and iterate to the next layer.

[0095] It should be noted that during the training phase, by precisely controlling the number of Gaussian points, the densification period, and the resolution of the training images during the densification process, multi-level and resource-controllable model generation is achieved. This strategy eliminates the need for post-reconstruction pruning or merging, directly producing multi-level models for dynamic LOD switching, making it suitable for efficient reconstruction and real-time rendering of city-level scenes.

[0096] Specifically, this embodiment forms a progressive training ladder by configuring dedicated Gaussian quantity limits, densification periods, and image resolution parameters for each level of detail. Lower-level levels use low resolution and sparse Gaussian distribution to quickly construct the basic outline of the scene, while higher-level levels recover detailed features through high-frequency densification and full resolution. During the execution of the training framework, the system loads the parameter configuration files for different levels sequentially. The initialization phase resets the rendering pipeline and memory management module to ensure resource isolation between layers. After completing the preset number of iterations for the current layer, the level switching protocol is automatically triggered.

[0097] This embodiment overcomes the problem of uneven resource allocation in traditional global optimization processes, achieving a smooth transition from low to high levels. Especially when building large-scale scenes, it avoids the deficiency of insufficient detail reconstruction capabilities in later stages due to excessive resource consumption in the early stages.

[0098] As one implementation, the dynamic growth function is configured as follows:

[0099] In the iteration number x∈[0,K] i Within the range, the number of Gaussians is controlled to approximate an accelerating growth curve. ;

[0100] The accelerated growth curve adopts a nonlinear growth model:

[0101] in, The initial Gaussian number, The maximum number of iterations for the i-th layer. It is a monotonically increasing function from [0,1] to [0,1].

[0102] It should be noted that the generation of detail levels begins from the first LOD level. The corresponding upper limit for the number of levels, the densification period, and the image downsampling coefficient are also set to [specific values ​​to be inserted here]. , and In the first i The maximum number of Gaussian scalars allowed per training round in each LOD level training phase is determined by a function. The calculation shows that:

[0103]

[0104] Specifically, It is the number of training iterations when densification terminates at the i-th layer, assuming M iterations are performed for each layer with densification enabled. ; This is the initial Gaussian number during scene initialization. This function allows the Gaussian number to grow relatively quickly in the early stages of training.

[0105] Furthermore, if the rapid growth of the Gaussian number is carried out in a linear manner, it is prone to oscillatory overshoot when approaching the upper limit. The use of a nonlinear growth model stems from the accurate simulation of the training dynamics system—high gradients are needed in the early stages of optimization to facilitate space exploration, while stable convergence is required in the later stages. Specifically, in the early stages of iteration (… The Gaussian growth rate (<0.5) exhibits exponential growth characteristics, rapidly covering the main structural regions of the scene. When the number of Gaussians approaches the preset upper limit, the growth rate decays to zero in a quadratic manner. This automatic convergence mechanism replaces traditional manual threshold adjustment. It ensures the formation of a stable scene skeleton (such as the outline of a building) during the middle of training, while reserving sufficient optimization space for subsequent texture details. Compared to a fixed growth strategy, this function can significantly reduce the number of invalid iterations.

[0106] Preferably, the To accelerate the growth function:

[0107] =

[0108] in, ∈[0,1].

[0109] The rate of change of the derivative of the accelerating growth function is such that the growth rate is maximum when x=0, and x= The time is reduced to 0, achieving a balance between rapid expansion in the early stage and gradual convergence in the later stage.

[0110] The maximum growth rate is placed at the training starting point (x=0), causing Gaussian points to quickly fill the blank area; the derivative gradually decays to zero as the process progresses (x= In essence, mathematics constructs a damped dynamic system. By modifying the curvature term of the function, it can adapt to different scene complexities. In densely vegetated areas, the initial slope is increased to quickly distribute branches and leaves; in building clusters, the later decay rate is reduced to finely depict door and window structures. This forms an adaptive growth rhythm control capability, solving the problem of regional detail imbalance caused by uniform growth, and improving the reconstruction accuracy of small elements such as road signs.

[0111] As one implementation method, the multi-factor denser decision-making includes:

[0112] Calculate the composite importance score of Gaussian points:

[0113]

[0114] in, The number of pixels hit. To accumulate transparency, For spatial volume, The magnitude of the image gradient;

[0115] according to Fractional probability sampling of Gaussian points is used to perform densification.

[0116] It should be noted that whenever densification is triggered, a subset of all training views is extracted, and the densification importance score of all current Gaussians is calculated. F , specifically, It is the number of pixels hit by the Gaussian algorithm; It represents the cumulative visibility during Gaussian rendering; It is the approximate volume of Gauss, that is, the product of its three scales. This is the average gradient of the Gaussian in the image space. The first three terms can be obtained from the intermediate variables of the 3D Gaussian rasterization operation, and the last term can be calculated through backpropagation.

[0117] Preferably, when controlling the growth of the Gaussian number through a dynamic growth function, and enabling dense training for M rounds of iterations per layer, ;

[0118] When calculating the composite importance score of Gaussian points, the scores are used as weights for sampling. Perform a densification operation on each Gaussian point until the resource budget limit for this level is reached;

[0119] Among them, based on The calculation results set an upper limit for single-time dense sampling.

[0120] It should be noted that a single metric cannot fully describe the geometric and apparent value of Gaussian points. (Number of hit pixels) Characterizing spatial coverage, cumulative transparency Reflects material density and spatial volume V Indicator geometric contribution, image gradient Capture edge features. Four elements complement each other. Cached during the rasterization stage. and Obtained through the differential rendering pipeline The weighted reconstruction is performed in a GPU parallel architecture. Normalization is performed during sampling to ensure that all factors are of an additive order of magnitude. This is achieved through a method based on... The calculation results are used to set the upper limit for single-pass dense sampling, effectively preventing local over-optimization. This embodiment achieves collaborative decision-making between rendering pipeline data and geometric features, reducing edge jaggedness while avoiding distant hole problems compared to traditional volumetric clipping methods.

[0121] After training is complete for each layer, the Gaussian model of that layer is exported as the output of the LOD model for that layer. Training then proceeds to the next layer, with parameters changing from... , , Replace with β i+1 γ i+1 , This results in multiple LOD layers that can be deployed independently or switched in real time.

[0122] As one implementation method, a cross-level inheritance mechanism is also included:

[0123] The training of the (i+1)th layer uses the output of the previous layer as the initialization starting point;

[0124] Inheriting the previous Gaussian points and based on β i+1 Perform incremental densification.

[0125] It should be noted that a complete re-initialization during layer switching would disrupt the established scene topology. The cross-layer inheritance mechanism in this implementation maintains the spatiotemporal coherence of geometric elements. Specifically, the Gaussian point set output from the previous layer serves as the initialization seed for the new layer, preserving position and scaling parameters through feature mapping. Incremental densification only expands newly introduced resource quotas; original points are only fine-tuned. This implementation maintains logical consistency from coarse to fine, eliminates object position jumps between layers, and reduces continuity errors at building edges from dynamic viewing perspectives.

[0126] As one implementation method, the method supports partitioned collaborative training:

[0127] The scene is divided into P spatial sub-regions. During the division process, it is necessary to ensure that individual buildings are not cut off, important landmarks are completely contained in a single block, and a 300-500 meter overlap buffer zone is maintained between blocks. This avoids geometric misalignment caused by the road centerline cutting off buildings and provides room for operation in subsequent boundary processing.

[0128] Each spatial sub-region independently performs the layered iterative training as described above. When performing layered training independently in each sub-region, the system records three key dimensions of information. The first is to use a local coordinate system and label the block code to form spatial coordinates. The second is to dynamically mark the LOD level to which the Gaussian point belongs during the densification process to form a level label. The third is to record the number of training iterations generated by the Gaussian point, i.e., the timestamp.

[0129] Generate a set of LOD models with hierarchical labels {LOD} i^p}, i∈[1,N], p∈[1,P]. Wherein, when block training is complete, a structured storage {LOD} is generated. i^p} = {coordinate set, hierarchical matrix, time series}, where the hierarchical matrix uses bit compression technology to store the hierarchical affiliation of each Gaussian point.

[0130] Preferably, the boundary regions of adjacent spatial sub-regions are defined, and a shared Gaussian point set is established for the boundary regions. The boundary regions are not fixed physical ranges, but rather interactive zones that are dynamically generated based on scene features. First, a three-dimensional space volume extending 50 meters outward from the edge of each block is calculated. Dynamic objects such as vehicles and pedestrians running within this space volume automatically trigger boundary expansion. Topological snapping is performed on linear elements such as road markings and continuous fences that cross blocks. The boundary range is adjusted in real time as the view moves.

[0131] The importance of the Gaussian points assigned at the boundaries of adjacent partitions is assessed.

[0132] During the densification operation at the boundary Gaussian points, the metadata of the associated partitions is updated synchronously. When a densification operation occurs at a boundary Gaussian point, the system executes the update synchronously, appending a version number and timestamp to each shared point. Only the changed attributes are transmitted, not the entire dataset. After receiving the update, the associated partition returns an acknowledgment signal. If no acknowledgment is received within a preset time (usually a 3-frame cycle), a reverse transaction rollback is initiated. The entire process is implemented through distributed shared memory, resulting in low latency.

[0133] It should be noted that, due to the discontinuity of normals and lighting breaks at the boundaries caused by traditional independent training of different partitions, this implementation allows each spatial sub-region to independently and in parallel generate a Level of Detail (LOD) hierarchical model, establishing a joint optimization space for boundary Gaussian points. Visibility data from adjacent partitions are fused and calculated, making it suitable for large-scale city modeling. A shared Gaussian point set triggers a special rendering pipeline during the rasterization stage, aggregating viewpoint sampling data from multiple partitions. This enables sub-pixel alignment of cross-regional objects such as road markings at the stitching points, resolving the bottleneck problem of stitching artifacts that restricts large-scale applications.

[0134] Reference Figure 3 This embodiment 2 provides a Gaussian splash detail level generation system, which applies a Gaussian splash detail level generation method as described in embodiment 1 above, including:

[0135] The hierarchy setting module is used to set N levels of detail, with each level configuring resource control parameter groups { , , },in:

[0136] Let β1 be the upper bound of the number of Gaussians of order i, satisfying β1 < β2 < ... < β2. N ;

[0137] To achieve a denser trigger cycle, satisfying 1> 2>...> N ;

[0138] To achieve the desired image downsampling rate during training, the following conditions must be met. 1< 2<...< N =1;

[0139] The hierarchical iterative training module is used to perform hierarchical iterative training from level i=1 to level N:

[0140] Starting from level 1, execute:

[0141] Initialize the training environment based on the current layer parameter set;

[0142] The Gaussian quantity is controlled to grow using a dynamic growth function. ;

[0143] When reached Periodic triggering of multi-factor dense decision-making;

[0144] The LOD model output module is used to output the current layer's LOD model and iterate to the next layer.

[0145] This system is based on a multi-level LOD model generation mechanism during the training phase: unlike existing post-processing generation methods, this system directly outputs a multi-level LOD model by setting resource constraints and training parameters layer by layer during the training phase, forming a detailed level with consistent structure and controlled resources.

[0146] Gaussian quantity growth function control strategy: Introducing an explicit function Control the upper limit of the number of Gaussians allowed during the training process of each layer to achieve resource scheduling and rhythm control during the densification process;

[0147] A multi-factor weighted density importance assessment mechanism: A Gaussian importance assessment index is proposed that integrates factors such as the number of hit pixels, transparency, scale volume, and average gradient. It is used to guide controlled density sampling and densitying operations, effectively improving model quality and densitying efficiency;

[0148] Training strategy for coordinated control of image downsampling and densification period: Each LOD layer is assigned an independent image resolution. With the densification cycle This is to achieve refined control of resources during the training phase;

[0149] Automatic output and deployment support mechanism for multi-layer models: This invention can directly export the model after each layer is trained, with a natural LOD layered structure, which is suitable for subsequent dynamic loading, hardware adaptation, rendering engine access and other application scenarios, and has good practicality and scalability.

[0150] Partition-compatible design for city-level 3D modeling tasks: It supports integration with the spatial partitioning mechanism of large-scale city-level scenes to achieve parallel training and model stitching. Each partition can also independently generate resource-controlled LOD-level models.

[0151] The following is a detailed description of embodiment 3:

[0152] The Gaussian splash detail level generation method proposed in Embodiment 3 has been experimentally verified on multiple large-scale real-world scenarios. The experimental results fully demonstrate the effectiveness and advancement of the "controlled density-based LOD level generation strategy" proposed in this invention, as shown below:

[0153] As shown in Table 1, the LOD hierarchy generated using this invention is significantly superior to previous technologies in terms of scene reconstruction quality.

[0154] Table 1 Experimental Results

[0155]

[0156] As shown in Table 2, the number of Gaussians can always be kept within the set range during the generation of each layer, and the resource consumption is stable and controllable.

[0157] Table 2 Resource Budget Control Effectiveness

[0158]

[0159] Each layer of the model can be used independently for real-time rendering and has good dynamic loading capabilities.

[0160] In summary, compared with the prior art, the above embodiments have at least the following technical advantages:

[0161] This invention directly outputs multi-level LOD models during the training phase, eliminating the need for post-processing: Existing technologies rely on pruning or merging Gaussians after reconstruction to generate models of varying accuracy, a cumbersome process that cannot guarantee optimal training for each layer. This invention outputs LOD models layer by layer during training, avoiding resource waste and quality degradation, and greatly simplifying the modeling process.

[0162] Resource consumption is controllable and adaptable to different platforms or task requirements: The number of Gaussians in each layer is set by explicit parameters and the growth process is dynamically controlled by functions, making model resource usage predictable and schedulable, meeting the rendering needs of various environments such as embedded devices, mobile terminals or cloud platforms.

[0163] Supports end-to-end training, improving efficiency and quality consistency: all LOD levels are generated in a unified training framework, avoiding information loss or inconsistency issues caused by post-processing; there is a natural progression of details between layers, which facilitates subsequent multi-precision fusion or dynamic switching.

[0164] With its simple structure and parallel scalability, this invention is suitable for city-level scene modeling: It is designed to be compatible with the regional training mechanism, and each region can independently generate multi-layer models, enabling large-scale distributed training and rapid stitching, effectively supporting city-level and regional-level large model reconstruction tasks.

[0165] Flexible densification strategies improve the balance between training efficiency and quality: By using a multi-factor weighted "densification importance score" to sample Gaussian points for replication, it is possible to balance structure restoration and resource saving, achieving a dynamic trade-off between precision and efficiency.

[0166] It natively supports dynamic LOD switching, enhancing the real-time rendering experience: the output multi-layered models have clear resource tags and quality levels, which can be directly used for LOD strategies such as view distance control and hardware adaptation, enabling smooth deployment and performance adaptation from mobile devices to high-end graphics platforms.

[0167] The above embodiments are merely preferred embodiments of the present invention and should not be construed as limiting the scope of protection of the present invention. Any non-substantial changes and substitutions made by those skilled in the art based on the present invention shall fall within the scope of protection claimed by the present invention.

Claims

1. A method for generating Gaussian splash detail levels, characterized by, Comprising: Set N level of detail, each level configuration resource control parameter group , , }, wherein: β1< β2<... < βN-1< βN= 1 N ; For the densification trigger period, meet 1> 2>...> N ; To train the image down-sampling rate, satisfy 1 2 N = 1; performing hierarchical iterative training from i = 1 to N levels: starting from the 1st level, performing: initializing a training environment based on the current level parameter group; controlling the number of gaussians to grow by a dynamic growth function ; When the periodicity is reached Trigger multi-factor densification decision when periodicity is reached; outputting the current level LOD model and iterating to the next level; the dynamic growth function is configured to: In the iteration number x∈[0,K i ] range, the control of the number of Gauss is approaching by the accelerated growth curve ; the accelerated growth curve adopts a nonlinear growth model: wherein, is the initial Gaussian number, is the maximum number of iterations for the i-th layer, is a monotonically increasing function from [0, 1]→[0, 1]; the multi-factor densification decision includes: calculating the Gaussian point complex importance score: wherein, is the number of hit pixels, is the accumulated transparency, is the scale space volume, is the image gradient magnitude; According to Fractional probability sampling Gaussian points perform densification.

2. A method for generating Gaussian specular detail levels as recited in claim 1, wherein, The To accelerate the growth function: = wherein ∈ [0, 1].

3. A method for generating Gaussian specular detail levels as recited in claim 2, wherein, The rate of change of the derivative of the accelerating growth function is such that the growth rate is maximum when x=0, and x= The time is reduced to 0, achieving a balance between rapid expansion in the early stage and gradual convergence in the later stage.

4. The Gaussian splash detail level generation method of claim 3, further comprising: In the case of controlling the number of Gaussians growth by dynamic growth function, each layer enables the dense training M rounds of iterations, ; In calculating the Gaussian point compound importance score, according to the score high and low, as a weight sampling A dense operation is performed on the Gaussian points until the upper limit of the resource budget of the current level is reached. wherein, based on The computation results set the upper limit of the single densification sampling.

5. The method of claim 1, wherein the Gaussian splotch detail layer is generated by: further comprising a cross-layer inheritance mechanism: ​ the i+1th level training takes the previous level output as the initialization starting point; Inherit previous layer gauss point and based on beta i+1 Perform delta densification.

6. The method of claim 1, wherein the Gaussian spatter detail layer is generated by: the method supports partition collaborative training: ​ dividing the scene into P spatial sub-regions; each spatial sub-region independently performs hierarchical iterative training; generating a set of LOD models with hierarchical labels {LOD i^p}, i e [1, N], p e [1, P].

7. The Gaussian splash detail level generation method of claim 6, further comprising: defining a boundary region of adjacent spatial sub-regions, and establishing a shared Gaussian point set for the boundary region; performing shared importance evaluation on the Gaussian point allocation at the boundary of adjacent partitions; synchronously updating the associated partition metadata when performing the densification operation of the boundary Gaussian points.

8. A Gaussian spatter detail level generation system applying a Gaussian spatter detail level generation method according to any one of claims 1 to 7, characterized by, Comprising: The hierarchical setting module is configured to set N levels of detail hierarchy, each level being configured with a set of resource control parameters } wherein:​​ is the upper limit of the number of Gaussians for the i-th level, satisfying β1< β2<... < β N ; For the densification trigger period, meet 1> 2>...> N ; To train the image down-sampling rate, satisfy 1 2 N = 1; a hierarchical iterative training module for performing hierarchical iterative training from i = 1 to N levels: starting from the 1st level, performing: initializing a training environment based on the current level parameter group; controlling the number of gaussians to grow by a dynamic growth function ; When the periodicity is reached trigger a multi-factor densification decision; an LOD model output module for outputting the current level LOD model and iterating to the next level.

Images