Complex scene-oriented three-dimensional Gaussian splash consistency modeling method and system

Abstract

The invention discloses a three-dimensional Gaussian splash consistency modeling method and system for a complex scene, belongs to the technical field of computer vision, and solves the problems that a complex area is not rebuilt and the number of Gaussian primitives in a flat area is too large in the existing 3DGS training process, so that the Gaussian primitives are reasonably distributed, and the overall rendering efficiency and the visual quality are improved. The problems that in the existing 3DGS training process, a complex area is not rebuilt, and the number of Gaussian primitives in a flat area is too large are solved, so that the Gaussian primitives are reasonably distributed, and the overall rendering efficiency and visual consistency are improved. The method comprises the following steps: establishing a three-dimensional Gaussian representation of a scene; extracting high-frequency detail information from the training view; calculating the neighborhood distribution density of each primitive; judging the matching degree of the complexity and the density; dynamically densifying or sparsifying the primitives according to a consistency result; and fusing a loss driving mechanism and a consistency driving mechanism, and comprehensively determining densification and pruning operation of the primitives. The method is suitable for three-dimensional reconstruction of complex scenes.

Description

Technical Field

[0001] This invention belongs to the field of computer vision technology, specifically relating to a three-dimensional reconstruction technology. Background Technology

[0002] 3D reconstruction, a long-standing research topic in the computer vision and computer graphics communities, aims to construct 3D representations of scenes from a set of pose-indicated input images. These representations can be used for novel perspective synthesis, geometric reconstruction, semantic understanding, and other downstream tasks. 3D Gaussian Splash (3DGS) has become the mainstream paradigm for this task in recent years, attracting significant attention due to its realistic rendering quality and real-time rendering speed. Because it explicitly uses a set of Gaussian primitives to model the scene, 3DGS is naturally compatible with graphics rasterization rendering pipelines, making it far more efficient than Neural Radiation Field (NeRF) in novel perspective synthesis. Recent research indicates that the allocation of Gaussian primitives is a key factor determining the quality of 3DGS representations, including densification and pruning operations on Gaussian primitives during training.

[0003] Training in 3DGS begins with a sparse set of Gaussian primitives, which are then iteratively optimized, densed, and pruned to gradually form a reasonably distributed and compact 3D representation. Vanilla 3DGS uses the spatial gradient of each Gaussian primitive as its denser score and marks coarse primitives with scores above a denser threshold as objects requiring denserening. While some recent works have proposed denser metrics beyond positional gradients, they still largely rely on rendering loss as the driving force. However, this loss-based denserening scheme often falls short in representing highly complex visual details. In textured regions, the rendering loss is easily smoothed out, resulting in the inability to effectively identify coarse Gaussian primitives for denserening. Although this problem can be mitigated by lowering the denser threshold to cover more primitives, this leads to an explosive increase in the number of Gaussian primitives in flat regions. These redundant primitives contribute little to improving rendering quality and instead slow down rendering efficiency. Summary of the Invention

[0004] This invention provides a 3D Gaussian splash consistency modeling method and system for complex scenes. Its purpose is to achieve higher rendering quality while maintaining the number of Gaussian primitives, and to solve the problems of under-reconstruction of complex regions and excessive number of Gaussian primitives in flat regions during the training process of existing 3DGS, thereby rationally allocating Gaussian primitives and improving overall rendering efficiency and visual quality.

[0005] Firstly, the purpose of this invention is to provide a three-dimensional Gaussian splash consistency modeling method for complex scenes, comprising: S1: Acquire training views, decompose training view images through discrete wavelet transform, obtain high-frequency components and generate a complexity energy map to obtain visual complexity index, where high-frequency components include horizontal high-frequency sub-bands, vertical high-frequency sub-bands and diagonal high-frequency sub-bands; S2: For each Gaussian primitive in the scene, obtain the spatial center point of other primitives in its neighborhood, calculate the geometric mean of the distance between the primitive and the center point of the neighboring primitives, and take the reciprocal as the density index. S3: For each Gaussian unit, normalize its visual complexity index and density index, and calculate the consistency value, where the consistency value is the result of the product of visual complexity and density normalized by the mean and standard deviation. S4: Adaptive allocation of high-level primitives based on consistency results: When the consistency value is less than zero and the visual complexity is higher than the preset threshold, it is determined to be an under-reconstructed region, and a primitive densification operation is performed. When the consistency value is less than zero and the visual complexity is lower than the preset threshold, it is determined to be an over-reconstructed region, and primitive pruning operation is performed. S5: In each training iteration, the primitive set selected by the rendering loss gradient and the primitive set selected by the consistency value are merged to jointly control the densification and pruning of primitives.

[0006] Furthermore, a preferred solution is provided: In step S1, obtaining the visual complexity index includes the following steps: decomposing the training view image using two-dimensional discrete wavelet transform to obtain low-frequency sub-bands and multiple high-frequency sub-bands, reconstructing the complexity energy map through the high-frequency sub-bands, mapping the energy map to the Gaussian primitives corresponding to each pixel during the rendering process to obtain the weighted weight of each primitive in the corresponding pixel, calculating the complexity response of the pixel to the primitive based on the weighted weight, and normalizing it, traversing all training views and pixels, taking the maximum value of the complexity response of each primitive, and obtaining the visual complexity index.

[0007] Furthermore, a preferred embodiment is provided: the formula for calculating the density index is: , in, Represents the primitive Several nearest neighbor sets, Indicates the center point of space. Indicates the center point of space. Indicates the target primitive index. This indicates a nearest-neighbor primitive index.

[0008] Furthermore, a preferred embodiment is provided: the formula for calculating the consistency value is: , in, Indicators of visual complexity Indicates density index, This represents the average value of the visual complexity index. This represents the average value of the density index. The standard deviation of the visual complexity index is represented by... This represents the standard deviation of the density index.

[0009] Furthermore, a preferred solution is provided: S4 further includes complexity-aware threshold adjustment: the densification threshold is dynamically set according to the primitive complexity, using the following formula: , in, The lower limit of the threshold. The upper limit of the threshold, For the Sigmoid function, To adjust the parameters, This represents the visual complexity of the primitives.

[0010] Furthermore, a preferred embodiment is provided: S5 includes: In each optimization iteration, a random sample is taken from the set of primitives selected by the consensus value, not exceeding [number missing]. The under-reconstruction primitives are densed, among which For the density scaling factor, Indicates the total number of primitives; At the same time, select no more than [number] primitives from the set of primitives selected by the rendering loss gradient. Each over-reconstructed primitive performs a sparsity operation, where This is the pruning ratio factor.

[0011] Furthermore, a preferred solution is provided: the primitive set selected by the rendering loss gradient is obtained based on the 3DGS method, specifically: when the gradient magnitude of a certain primitive exceeds a preset threshold, a primitive densification operation is performed to obtain the primitive set selected by the rendering loss gradient.

[0012] Secondly, the purpose of this invention is to propose a three-dimensional Gaussian splash consistency modeling system for complex scenes. The system is implemented based on a three-dimensional Gaussian splash consistency modeling method for complex scenes as described in any one or more of the above-mentioned schemes. The system includes: Complexity extraction module: used to acquire training views, decompose training view images through discrete wavelet transform, obtain high-frequency components and generate a complexity energy map to obtain visual complexity index, where high-frequency components include horizontal high-frequency sub-bands, vertical high-frequency sub-bands and diagonal high-frequency sub-bands; Density calculation module: For each Gaussian primitive in the scene, it obtains the spatial center point of other primitives in its neighborhood, calculates the geometric mean of the distance between the primitive and the center point of the neighboring primitives, and takes the reciprocal as the density index. Consistency assessment module: For each Gaussian unit, the visual complexity index and density index are normalized, and the consistency value is calculated. The consistency value is the result of the product of visual complexity and density being normalized by the mean and standard deviation. Adaptive allocation module: Used for adaptive allocation of high-level primitives based on consistency results. When the consistency value is less than zero and the visual complexity is higher than the preset threshold, it is determined to be an under-reconstructed region, and a primitive densification operation is performed. When the consistency value is less than zero and the visual complexity is lower than the preset threshold, it is determined to be an over-reconstructed region, and primitive pruning operation is performed. The joint allocation module of loss-driven and consistency-driven methods is used to merge the primitive set selected by the rendering loss gradient and the primitive set selected by the consistency value in each training iteration, and jointly control the densification and pruning of primitives.

[0013] Thirdly, the present invention aims to provide a computer device, the computer device including a memory and a processor, the memory storing a computer program, and when the processor runs the computer program stored in the memory, the processor executes a three-dimensional Gaussian splash consistency modeling method for complex scenes according to any one or more of the above-described schemes.

[0014] Fourthly, the present invention aims to provide a computer-readable storage medium for storing a computer program that executes a three-dimensional Gaussian splash consistency modeling method for complex scenarios as described in any one or more of the above-described schemes.

[0015] Compared with the prior art, the advantages of the present invention are: This invention proposes a novel Gaussian primitive allocation method that guides 3DGS to use dense primitives in complex regions and sparse primitives in flat regions based on loss-independent visual prior information, thereby achieving intelligent allocation. This invention designs the CDC-GS method, which combines wavelet frequency prior and primitive density to identify primitives in under-reconstruction and over-reconstruction regions for densification and cropping. The complexity-density consistent prior proposed in this invention can serve as an effective supplement to existing loss-driven allocation strategies, exhibiting orthogonality and complementarity.

[0016] The CDC-GS method described in this invention achieves higher 3D reconstruction quality with the same primitive budget. By introducing the complexity-density consistency principle, this invention significantly improves detail modeling capabilities in texture-rich regions and reduces redundant primitives in smooth regions, thus optimizing rendering speed and memory usage. Compared to traditional methods that rely solely on rendering losses, this invention can more reliably identify under-reconstructed regions, thereby improving the overall scene fidelity and visual quality.

[0017] This invention is applicable to 3D reconstruction of complex scenes. Attached Figure Description

[0018] Figure 1 This is a flowchart illustrating a three-dimensional Gaussian splash consistency modeling method for complex scenarios, as described in a specific embodiment of the present invention. Detailed Implementation

[0019] In the following description, specific details such as particular system architectures and techniques are set forth for illustrative purposes and not for limitation, in order to provide a thorough understanding of the embodiments of this application. However, those skilled in the art will understand that this application can also be implemented in other embodiments without these specific details. In other instances, detailed descriptions of well-known systems, apparatuses, circuits, and methods are omitted so as not to obscure the description of this application with unnecessary detail.

[0020] The technical solutions in the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of them. Based on the embodiments in this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0021] Many specific details are set forth in the following description in order to provide a full understanding of this application. However, this application may also be implemented in other ways different from those described herein. Those skilled in the art can make similar extensions without departing from the spirit of this application. Therefore, this application is not limited to the specific embodiments disclosed below.

[0022] Implementation Method 1 The scene consists of several three-dimensional Gaussian elements. This means that each primitive contains: a spatial center point. covariance matrix Opacity and spherical harmonic coefficient Its orientation depends on the color distribution.

[0023] The spatial influence of the primitive can be calculated using the Gaussian function: in Let be any three-dimensional point. This formula is used to determine the range of contribution of a primitive to a pixel in three-dimensional space.

[0024] During rendering, each primitive is projected as a two-dimensional Gaussian. Pixel color Transparency-weighted composition: in, This indicates the number of primitives covering that pixel. This represents the 2D opacity of the primitives that make up a pixel. Indicates the 3D opacity of the primitive. Indicates the color of the primitive. This represents the 2D opacity of the primitive. This cumulative opacity model allows for continuous and noise-resistant rendering.

[0025] A method for consistent 3D Gaussian splash modeling in complex scenarios includes the following steps: S1: Acquire training views, decompose training view images through discrete wavelet transform, obtain high-frequency components and generate a complexity energy map to obtain visual complexity index, where high-frequency components include horizontal high-frequency sub-bands, vertical high-frequency sub-bands and diagonal high-frequency sub-bands; Specifically, to evaluate the visual complexity of each region in the image, this implementation employs two-dimensional discrete wavelet transform (DWT). DWT uses a pair of low-pass filters. With high-pass filter Input image Decomposed into four sub-bands: low-frequency component With high frequency components , , .

[0026] Image complexity energy graph Obtained from high-frequency subband reconstruction: in This reflects the high-frequency energy of texture and edge regions. This energy map is then mapped back to the corresponding Gaussian primitives to measure the complexity of the region in which each primitive resides. High-energy regions indicate that the primitive is located in a complex region with rich texture details.

[0027] The system then maps this energy map to the Gaussian units corresponding to each pixel during the rendering process. For any pixel... If it is rendered by several Gaussian primitives, then each primitive The weighted weight of this pixel is defined as: in, The opacity weighting term for this primitive, This is the cumulative transmission term.

[0028] Based on the above weights, the system calculates the pixel pair primitives. Complexity response: , That is, the pixel complexity value is weighted by the primitive contribution and then normalized.

[0029] The system further iterates through all training views and pixels, taking the maximum value of the complexity response for each primitive to obtain its global complexity index: in, Characterization primitives The maximum complexity response of the region.

[0030] A higher complexity value indicates that the primitive is located in a region with rich texture details or sharp edge variations; a lower value indicates that the primitive is located in a smooth or low-frequency region. Finally, this module outputs the complexity index for each primitive. The data is then passed to the consistency assessment module, which uses it in conjunction with the density index to determine the reconstruction status of the region.

[0031] S2: For each Gaussian primitive in the scene, obtain the spatial center point of other primitives in its neighborhood, calculate the geometric mean of the distance between the primitive and the center point of the neighboring primitives, and take the reciprocal as the density index. To measure the density of primitives within a local space, this implementation calculates the distance between the center of the target primitive and the center of its neighboring primitives, and takes the reciprocal of their geometric mean as the density index: in, Represents the primitive The nearest neighbor set. The larger the value of this index, the denser the distribution of primitives in the region; the smaller the value, the sparser the primitives, which may indicate under-reconstruction.

[0032] S3: For each Gaussian unit, normalize its visual complexity index and density index, and calculate the consistency value, where the consistency value is the result of the product of visual complexity and density normalized by the mean and standard deviation. This implementation proposes a complexity-density consistency index to determine whether a region is too dense or too sparse. The consistency value is defined as the normalized product of the two: in, and These represent complexity and density, respectively. , For its average value, , The standard deviation is denoted as .

[0033] when When the complexity and density are well matched, it indicates that the complexity and density are well matched. when Furthermore, high complexity indicates insufficient reconstruction, and more primitives should be added; when Furthermore, if the complexity is low, it indicates over-reconstruction, and redundant primitives should be deleted.

[0034] S4: Adaptive allocation of high-level primitives based on consistency results: When the consistency value is less than zero and the visual complexity is higher than the preset threshold, it is determined to be an under-reconstructed region, and a primitive densification operation is performed. When the consistency value is less than zero and the visual complexity is lower than the preset threshold, it is determined to be an over-reconstructed region, and primitive pruning operation is performed. Specifically: Densification of under-reconstruction areas: for Furthermore, for primitive regions with high complexity, a certain proportion is randomly sampled to perform a densification operation, and the primitive parameters are copied and fine-tuned to improve local details.

[0035] Over-reconstruction sparsity: for Furthermore, in regions with lower complexity, some primitives are randomly deleted to reduce redundancy.

[0036] Complexity-aware adaptive threshold adjustment: This implementation further adaptively adjusts the densification threshold based on complexity. in, These are the upper and lower limits of the threshold. For the Sigmoid function, To adjust the parameters, complex regions correspond to lower thresholds, making them more likely to trigger densification; simple regions, on the other hand, have higher thresholds, thereby reducing invalid allocations.

[0037] S5: In each training iteration, the primitive set selected by the rendering loss gradient and the primitive set selected by the consistency value are merged to jointly control the densification and pruning of primitives.

[0038] Specifically: In traditional 3DGS methods, primitive densification primarily relies on the positional gradient signals generated by the rendering loss. When the gradient magnitude of a primitive exceeds a preset threshold... When the system determines that the region exhibits under-reconstruction, it triggers a primitive densification operation. This mechanism is denoted as the loss-driven set, i.e. .

[0039] However, rendering loss is primarily sensitive to low-frequency regions, easily overlooking high-frequency regions with complex textures but smooth error distribution, leading to underfitting of the model in detailed areas. Therefore, this implementation introduces a supplementary set based on the complexity-density consistency index. ,pass The sign and amplitude determine the reconstruction status of the region: when Furthermore, when the complexity is high, it is identified as an under-reconstruction region; when Furthermore, when the complexity is low, it is judged as an over-reconstructed region.

[0040] Merge the two sets: ,in, Selected by the traditional rendering loss gradient; The selection is based on the consistency mechanism proposed in this implementation.

[0041] In each optimization iteration, this implementation method sets... Random sampling in the middle does not exceed Each under-reconstructed primitive performs a densification operation. (This is the density scaling factor). At the same time, in the set Select no more than Each over-reconstructed primitive performs a sparsification operation. (This refers to the pruning ratio factor). To improve overall stability, this implementation eliminates the original 3DGS strategy of periodically pruning every 100 steps, and instead deletes only primitives with an opacity lower than 0.1 every 3000 steps.

[0042] Through this loss-driven and consistency-driven collaborative mechanism, CDC-GS can simultaneously focus on rendering errors and visual complexity, achieving more comprehensive dynamic primitive allocation and significantly improving the algorithm's generalization and reconstruction efficiency.

[0043] Implementation Method 2 This embodiment is a further illustrative example of the three-dimensional Gaussian splash consistency modeling method for complex scenarios described in Embodiment 1.

[0044] Datasets: This implementation evaluates the method on several real-world datasets widely used in novel perspective synthesis tasks. Specifically, this implementation follows the standard settings of Vanilla 3DGS, and the preprocessing of the training data is strictly performed according to Vanilla 3DGS methods.

[0045] Evaluation Metrics: Quantitative evaluation results for three standard metrics are as follows: Peak Signal-to-Noise Ratio (PSNR); Structural Similarity Metric (SSIM); and Learned Aware Image Similarity (LPIPS). All metrics were calculated on the test view, following the original evaluation protocol. The results for each dataset are the average values ​​for all scenes within that dataset. To mitigate the impact of randomness, each experiment was repeated three times using a different random seed, and the average performance is presented.

[0046] Comparison Method: This embodiment compares the described method with a series of representative Gaussian primitive allocation methods, including: Vanilla 3DGS; Taming-3DGS (which constrains the number of primitives to the value set in this embodiment). Same); AbsGS (dense version with a densification threshold of 0.0004); Mini-Splatting (densification only, no cropping); Pixel-GS. All baseline method results were reproduced using their official implementations and default configurations.

[0047] Implementation details: This implementation method is based on the official Vanilla 3DGS code. The complexity graph of the training view is extracted using a single-layer Haar DWT; the density of Gaussian primitives is calculated using the three nearest neighbors, i.e., for all primitives... By default, the cropping and densification sampling ratios are both set to 1. , This implementation also evaluated the improvement of the densification ratio. Version; the complexity-aware densification threshold is set as follows: Other hyperparameters remained consistent with Vanilla 3DGS. All experiments were conducted on a single NVIDIA RTX 3090 graphics card.

[0048] Quantitative Comparison: Table 1 presents the quantitative comparison results between CDC-GS and various baseline methods in this embodiment. For fair comparison, methods with approximately the same number of primitives are grouped together: in the sparse primitives group, CDC-GS (… This implementation consistently outperforms all baseline methods, exhibiting higher fidelity and perceived quality while maintaining the same number of primitives. In dense primitive groups, this implementation combines the advantages of CDC-GS. The value was set to 0.02 to match the number of primitives in the baseline methods, ensuring a fair and comparable experiment. Overall, CDC-GS achieved the best rendering quality on all benchmark datasets with fewer primitives. This further validates the complexity-density consistency strategy proposed in this implementation: it models structural details more accurately while improving primitive allocation efficiency.

[0049] Table 1 Quantitative Comparison Qualitative analysis: The method of this embodiment outperforms in all aspects, with clearer textures, cleaner structural boundaries, and fewer artifacts. These results fully demonstrate the success of the method described in this embodiment in identifying and addressing "under-reconstruction regions" that are easily overlooked by baseline methods.

[0050] Ablation experiment: This implementation method uses the Tanks & Temples dataset with Vanilla 3DGS as the backbone to conduct ablation experiments, evaluating the two core modules proposed in this implementation method: Complexity-Aware Adaptive Threshold (CAAT) and Complexity-Density Consistency (CDC). The experimental results are summarized in Table 2.

[0051] Table 2 Ablation Experiment Ablation analysis of CAAT: CAAT improves reconstruction quality by dynamically adjusting its densification threshold based on the complexity of each primitive. It more aggressively promotes densification in complex regions, thus leading to a slight increase in the number of Gaussian primitives.

[0052] Ablation analysis of CDC: The CDC module achieves significantly improved rendering quality by simultaneously densifying under-reconstructed regions and pruning over-reconstructed regions. The improved visual quality indicates that CDC effectively alleviates the under-reconstruction problem in Vanilla 3DGS; simultaneously, the total number of Gaussian primitives decreases, suggesting that CDC successfully removes redundant primitives in low-complexity regions.

[0053] Complete Method (CDC-GS): When CAAT and CDC are combined in this embodiment, they exhibit a synergistic effect: CAAT further enhances the density quality of local details; CDC adjusts the primitive distribution at the global level according to complexity, forcibly maintaining complexity-density consistency. Ultimately, the CDC-GS method in this embodiment significantly outperforms the original Vanilla 3DGS in terms of reconstruction quality, while maintaining a comparable number of primitives, verifying the effectiveness and universality of the Gaussian primitive allocation strategy proposed in this embodiment.

[0054] This invention proposes a novel Gaussian grammar allocation method—Complexity-Density Consistent Gaussian Splashing (CDC-GS)—which is essentially orthogonal to traditional loss-driven methods (such as position gradient). The core idea is to reconstruct insufficient regions using high-complexity but low-density regions, and to reconstruct excessive regions using low-complexity but high-density regions, thereby achieving more reasonable densification and pruning of Gaussian grammars. Extensive experiments in this implementation demonstrate the superiority of this method and prove its powerful complement to existing methods in Gaussian grammar allocation tasks.

[0055] Implementation Method 3: A 3D Gaussian splash consistency modeling system for complex scenes, the system being implemented based on a 3D Gaussian splash consistency modeling method for complex scenes as described in Embodiment 1, the system comprising: Complexity extraction module: used to acquire training views, decompose training view images through discrete wavelet transform, obtain high-frequency components and generate a complexity energy map to obtain visual complexity index, where high-frequency components include horizontal high-frequency sub-bands, vertical high-frequency sub-bands and diagonal high-frequency sub-bands; Density calculation module: For each Gaussian primitive in the scene, it obtains the spatial center point of other primitives in its neighborhood, calculates the geometric mean of the distance between the primitive and the center point of the neighboring primitives, and takes the reciprocal as the density index. Consistency assessment module: For each Gaussian unit, the visual complexity index and density index are normalized, and the consistency value is calculated. The consistency value is the result of the product of visual complexity and density being normalized by the mean and standard deviation. Adaptive allocation module: Used for adaptive allocation of high-level primitives based on consistency results. When the consistency value is less than zero and the visual complexity is higher than the preset threshold, it is determined to be an under-reconstructed region, and a primitive densification operation is performed. When the consistency value is less than zero and the visual complexity is lower than the preset threshold, it is determined to be an over-reconstructed region, and primitive pruning operation is performed. The joint allocation module of loss-driven and consistency-driven methods is used to merge the primitive set selected by the rendering loss gradient and the primitive set selected by the consistency value in each training iteration, and jointly control the densification and pruning of primitives.

[0056] It is understood that the present invention has been described through some embodiments, and those skilled in the art will recognize that various changes or equivalent substitutions can be made to these features and embodiments without departing from the spirit and scope of the invention. Furthermore, under the teachings of the present invention, these features and embodiments can be modified to adapt to specific situations and materials without departing from the spirit and scope of the invention. Therefore, the present invention is not limited to the specific embodiments disclosed herein, and all embodiments falling within the scope of the claims of this application are within the protection scope of the present invention.

Claims

1. A three-dimensional Gaussian splash consistency modeling method for complex scenes, characterized in that, The method includes: S1: Acquire training views, decompose training view images through discrete wavelet transform, obtain high-frequency components and generate a complexity energy map to obtain visual complexity index, where high-frequency components include horizontal high-frequency sub-bands, vertical high-frequency sub-bands and diagonal high-frequency sub-bands; S2: For each Gaussian primitive in the scene, obtain the spatial center point of other primitives in its neighborhood, calculate the geometric mean of the distance between the primitive and the center point of the neighboring primitives, and take the reciprocal as the density index. S3: For each Gaussian unit, normalize its visual complexity index and density index, and calculate the consistency value, where the consistency value is the result of the product of visual complexity and density normalized by the mean and standard deviation. S4: Adaptive allocation of high-level primitives based on consistency results: When the consistency value is less than zero and the visual complexity is higher than the preset threshold, it is determined to be an under-reconstructed region, and a primitive densification operation is performed. When the consistency value is less than zero and the visual complexity is lower than the preset threshold, it is determined to be an over-reconstructed region, and primitive pruning operation is performed. S5: In each training iteration, the primitive set selected by the rendering loss gradient and the primitive set selected by the consistency value are merged to jointly control the densification and pruning of primitives.

2. The three-dimensional Gaussian splash consistency modeling method for complex scenes according to claim 1, characterized in that, In step S1, obtaining the visual complexity index includes the following steps: decomposing the training view image using two-dimensional discrete wavelet transform to obtain low-frequency sub-bands and multiple high-frequency sub-bands, reconstructing the complexity energy map through the high-frequency sub-bands, mapping the energy map to the Gaussian primitives corresponding to each pixel during the rendering process, obtaining the weighted weight of each primitive in the corresponding pixel, calculating the complexity response of the pixel to the primitive based on the weighted weight, and normalizing it, traversing all training views and pixels, taking the maximum value of the complexity response of each primitive, and obtaining the visual complexity index.

3. The three-dimensional Gaussian splash consistency modeling method for complex scenes according to claim 1, characterized in that, The formula for calculating the density index is as follows: ， in, Represents the primitive Several nearest neighbor sets, Indicates the center point of space. Indicates the center point of space. Indicates the target primitive index. This indicates a nearest-neighbor primitive index.

4. The three-dimensional Gaussian splash consistency modeling method for complex scenes according to claim 1, characterized in that, The formula for calculating the consistency value is: ， in, Indicators of visual complexity Indicates density index, This represents the average value of the visual complexity index. This represents the average value of the density index. The standard deviation of visual complexity is represented by the following: This represents the standard deviation of the density index.

5. The three-dimensional Gaussian splash consistency modeling method for complex scenes according to claim 1, characterized in that, S4 also includes complexity-aware threshold adjustment: the densification threshold is dynamically set based on the primitive complexity, using the following formula: ， in, The lower limit of the threshold. The upper limit of the threshold, For the Sigmoid function, To adjust the parameters, This represents the visual complexity of the primitives.

6. The three-dimensional Gaussian splash consistency modeling method for complex scenes according to claim 1, characterized in that, S5 includes: In each optimization iteration, a random sample is taken from the set of primitives selected by the consensus value, not exceeding [number missing]. The under-reconstruction primitives are densed, among which For the density scaling factor, Indicates the total number of primitives; At the same time, select no more than [number] primitives from the set of primitives selected by the rendering loss gradient. Each over-reconstructed primitive performs a sparsity operation, where This is the pruning ratio factor.

7. The three-dimensional Gaussian splash consistency modeling method for complex scenes according to claim 1, characterized in that, The primitive set selected by the rendering loss gradient is obtained based on the 3DGS method. Specifically, when the gradient magnitude of a primitive exceeds a preset threshold, a primitive densification operation is performed to obtain the primitive set selected by the rendering loss gradient.

8. A three-dimensional Gaussian splash consistency modeling system for complex scenes, characterized in that, The system is implemented based on a three-dimensional Gaussian splash consistency modeling method for complex scenes as described in any one of claims 1-7, and the system includes: Complexity extraction module: used to acquire training views, decompose training view images through discrete wavelet transform, obtain high-frequency components and generate a complexity energy map to obtain visual complexity index, where high-frequency components include horizontal high-frequency sub-bands, vertical high-frequency sub-bands and diagonal high-frequency sub-bands; Density calculation module: For each Gaussian primitive in the scene, it obtains the spatial center point of other primitives in its neighborhood, calculates the geometric mean of the distance between the primitive and the center point of the neighboring primitives, and takes the reciprocal as the density index. Consistency assessment module: For each Gaussian unit, the visual complexity index and density index are normalized, and the consistency value is calculated. The consistency value is the result of the product of visual complexity and density being normalized by the mean and standard deviation. Adaptive allocation module: Used for adaptive allocation of high-level primitives based on consistency results. When the consistency value is less than zero and the visual complexity is higher than the preset threshold, it is determined to be an under-reconstructed region, and a primitive densification operation is performed. When the consistency value is less than zero and the visual complexity is lower than the preset threshold, it is determined to be an over-reconstructed region, and primitive pruning operation is performed. The joint allocation module of loss-driven and consistency-driven methods is used to merge the primitive set selected by the rendering loss gradient and the primitive set selected by the consistency value in each training iteration, and jointly control the densification and pruning of primitives.

9. A computer device, characterized in that, The computer device includes a memory and a processor. The memory stores a computer program. When the processor runs the computer program stored in the memory, the processor executes a three-dimensional Gaussian splash consistency modeling method for complex scenes according to any one of claims 1-7.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium is used to store a computer program that executes a three-dimensional Gaussian splash consistency modeling method for complex scenarios according to any one of claims 1-7.

Images