3D GS three-dimensional reconstruction method and system based on ADC optimization
By using an ADC-optimized 3DGS 3D reconstruction method, the shortcomings of traditional 3D reconstruction technology in terms of efficiency, quality, and editability are solved, achieving efficient and accurate 3D reconstruction results, which are suitable for rendering and editing complex scenes.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- 北京国星创图科技有限公司
- Filing Date
- 2025-12-26
- Publication Date
- 2026-04-17
AI Technical Summary
Traditional 3D reconstruction techniques have shortcomings in terms of reconstruction efficiency, rendering quality, and editability. In particular, they suffer from problems such as large computational load, large data volume, difficulty in editing, difficulty in reconstructing dynamic scenes, and complexity in editing geometric textures when reconstructing large-scale complex scenes.
The 3D reconstruction method based on adaptive density control (ADC) optimization is adopted. It generates an initial sparse point cloud, initializes a 3D Gaussian distribution, and optimizes the quantity and spatial distribution of the Gaussian distribution through dynamic gradient thresholding, scene range correction and 3DGS pruning strategy. Finally, a new perspective image is synthesized through a differentiable renderer.
It significantly improves rendering quality, reduces the number of 3DGS, increases training speed and reconstruction quality, avoids overfitting and loss of detail, and adapts to complex scene reconstruction.
Smart Images
Figure CN121883709A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of three-dimensional reconstruction technology, specifically to a 3DGS three-dimensional reconstruction method and system based on ADC optimization. Background Technology
[0002] 3D reconstruction technology is a crucial support for building virtual digital worlds. Currently, traditional 3D reconstruction techniques, such as MVS, have shortcomings in reconstruction efficiency, rendering quality, and editability. Implicit representation-based 3D reconstruction methods, such as NeRF, are difficult to directly extract and apply to the construction of 3D virtual digital worlds. Traditional 3DGS uses explicit representation, which is suitable for the extraction and editing of 3D information, but its training speed and rendering quality need further improvement.
[0003] MVS has several shortcomings in 3D reconstruction. In terms of efficiency, it requires a large amount of computation when handling large-scale scenes, resulting in slow training and rendering speeds. Regarding representation, it is based on point clouds or meshes, leading to large data volumes and difficulties in editing, making it hard to directly obtain detailed geometric information. In terms of applicability, it struggles with dynamic scene reconstruction and is easily limited in handling complex scenes. In terms of editability, the geometry and texture editing process is complex and less convenient than 3DGS.
[0004] It is difficult to quickly generate images from new perspectives. In terms of representation, it uses implicit representation, making it difficult to directly obtain and manipulate 3D geometric information, which is inconvenient for subsequent editing and simulation. Regarding applicability, dynamic scene reconstruction is difficult, and large-scale complex scene reconstruction is prone to detail loss or inaccuracy. In terms of editability, geometric and texture editing is difficult, requiring complex conversions and processing, and the results are not as ideal as 3DGS.
[0005] Traditional 3DGS also faces challenges in 3D reconstruction, primarily in managing the number and distribution of Gaussian distributions. Specifically, in rough and smooth regions, a small number of large Gaussian distributions are sufficient, while in fine and complex regions, a large number of small Gaussian distributions are required. Although 3DGS addresses this issue through an Adaptive Density Control (ADC) algorithm—which clones or splits Gaussian distributions if they change frequently at relevant locations during optimization and deletes them if their opacity falls below a certain threshold—this mechanism can lead to quality issues in many cases. For example, in simple regions, the ADC algorithm may generate too many Gaussian distributions, resulting in overfitting; while in highly complex regions, such as grasslands, the reconstruction may be inadequate. Therefore, further optimization of the ADC is needed to improve reconstruction quality while reducing the number of 3DGS distributions required. Summary of the Invention
[0006] To achieve the objectives of this invention, this application provides a 3DGS three-dimensional reconstruction method based on ADC optimization, comprising: Step S1: Generate an initial sparse point cloud based on the motion recovery structure from the multi-view images; Step S2: Initialize a 3D Gaussian distribution based on the initial sparse point cloud. Each 3D Gaussian distribution includes position, covariance matrix, opacity, and color feature parameters. Step S3: Improve the adaptive density control algorithm by using a dynamic gradient threshold strategy, a scene range correction strategy, and a 3DGS pruning strategy to dynamically optimize the quantity and spatial distribution of the 3D Gaussian distribution; Step S4: Use a differentiable renderer to synthesize the optimized 3D Gaussian distribution from a new perspective and output the 3D reconstruction result.
[0007] In some specific embodiments, in step S2, the covariance matrix of the 3D Gaussian distribution is determined by parameterization based on the four-dimensional quaternion vector and the three-dimensional scale vector.
[0008] In some specific embodiments, step S3, the dynamic gradient threshold strategy includes: adjusting the gradient threshold according to an exponential increase during training, using a low threshold in the initial stage to quickly generate a Gaussian distribution, and gradually increasing the threshold in the later stage to suppress overfitting.
[0009] In some specific embodiments, the gradient threshold is determined according to the following formula: In the formula, i This represents the current iteration number. This represents the total number of iterations. As the initial value, This is the final value of the gradient threshold.
[0010] In some specific embodiments, in step S3, the scene range correction strategy includes: combining point cloud data based on motion recovery structure with camera position to determine the actual scene range and optimizing the cloning and splitting decisions of Gaussian distribution.
[0011] In some specific embodiments, the scene range correction strategy further includes: when there are dynamic objects in the scene, abnormal point cloud data is removed through temporal consistency detection before determining the actual scene range.
[0012] In some specific embodiments, the actual range of the scenario is determined according to the following formula: In the formula, N SfM The number of point clouds generated by the structure-of-motion (SFM) method. p represents the average position of the camera in the formula. i Here are the coordinates of the point cloud.
[0013] In some specific embodiments, in step S3, the 3DGS pruning strategy includes: pruning distributions that contribute little to scene fidelity based on the opacity of the Gaussian distribution and the cumulative value of its alpha blending coefficients across all views.
[0014] In some specific embodiments, the cumulative value of the alpha mixing coefficient is determined according to the following formula: In the formula, The k-th Gaussian distribution contributes weights to the color of pixel p. This represents the number of pixels in the image.
[0015] To achieve the same inventive objective, this application also provides a 3DGS three-dimensional reconstruction system based on ADC optimization, comprising: Image acquisition module: used to generate an initial sparse point cloud based on multi-view images through motion reconstruction structure; Initialization module: used to initialize a 3D Gaussian distribution based on the initial sparse point cloud. Each 3D Gaussian distribution includes position, covariance matrix, opacity and color feature parameters. Dynamic optimization module: used to improve the adaptive density control algorithm through dynamic gradient threshold strategy, scene range correction strategy and 3DGS pruning strategy, so as to dynamically optimize the quantity and spatial distribution of the 3D Gaussian distribution; The results output module is used to synthesize the optimized 3D Gaussian distribution from a new perspective using a differentiable renderer and output the 3D reconstruction results.
[0016] The beneficial effects of the above technical solution are as follows: (1) Improved Rendering Quality: The improved Adaptive Density Control (ADC) mechanism significantly improves rendering quality without increasing the number of 3DGS. By introducing scene extent calculation correction, this method not only considers the camera position but also combines view depth information to more accurately distribute Gaussian points, effectively solving the problem of unreasonable Gaussian point distribution caused by inaccurate scene extent estimation in traditional methods. In complex areas, such as grass or fine textured surfaces, the new ADC mechanism can generate enough small Gaussian points to capture details, while avoiding excessively dense Gaussian points in simple areas, thereby reducing background under-reconstruction and foreground overfitting artifacts. Experimental results show that compared with the original 3DGS, the improved rendering quality is improved on multiple datasets, especially when dealing with highly complex scenes, its advantages are more obvious.
[0017] (2) Accelerating Training Speed: The introduction of an exponentially rising gradient threshold is a key innovation for accelerating training speed. In the early stages of training, a lower gradient threshold allows a large number of 3DGS to be generated and adjusted, thus quickly filling the scene and capturing the main structure. As training progresses, the gradient threshold gradually increases exponentially, eventually reaching a high value, which ensures that only those 3DGS with significant positional gradient changes are further split or cloned. This dynamic adjustment strategy avoids the convergence difficulties caused by insufficient 3DGS in the early stages of training, as well as the overfitting and computational redundancy caused by too many 3DGS in the later stages. Experimental data show that after adopting an exponentially rising gradient threshold, the training convergence speed is significantly accelerated, the training time is shortened by more than half, greatly improving work efficiency and reducing computational resource consumption.
[0018] (3) Maintaining High Reconstruction Quality: The saliency-aware pruning strategy excels in reducing unnecessary 3DGS while successfully maintaining high reconstruction quality. This strategy innovatively combines the opacity and cumulative alpha blending coefficient of 3DGS to evaluate their contribution to scene fidelity. 3DGS with low opacity but high cumulative alpha blending coefficient—those that significantly contribute to pixel color across multiple views—are retained; while those with smaller contributions are pruned. This method ensures that the number of 3DGS in the scene is optimized, scene compactness is improved, and detail loss and reconstruction quality degradation due to excessive pruning are avoided. Experimental results demonstrate that the scene after saliency-aware pruning maintains high reconstruction quality while significantly reducing the number of 3DGS, improving rendering efficiency, and without exhibiting obvious artifacts or detail loss issues. Attached Figure Description
[0019] To more clearly illustrate the technical solutions in the embodiments of this application, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.
[0020] Figure 1 A flowchart illustrating an ADC-optimized 3DGS three-dimensional reconstruction method according to an embodiment of the present invention is provided. Figure 2 A schematic diagram of the structure of a 3DGS three-dimensional reconstruction system based on ADC optimization is provided for one embodiment of the present invention; Figure 3 The following is a pseudocode diagram of an ADC-based 3DGS 3D reconstruction algorithm, which is provided as an embodiment of the present invention. Detailed Implementation
[0021] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments.
[0022] Examples of the embodiments are shown in the accompanying drawings, wherein the same or similar symbols denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and intended to explain the invention, and should not be construed as limiting the invention.
[0023] Example 1 One embodiment of the present invention provides a 3DGS three-dimensional reconstruction method based on ADC optimization, referring to... Figure 1 As shown, it includes: Step S1: Generate an initial sparse point cloud based on the motion recovery structure from the multi-view images; Step S2: Initialize a 3D Gaussian distribution based on the initial sparse point cloud. Each 3D Gaussian distribution includes position, covariance matrix, opacity, and color feature parameters. In a specific embodiment of the present invention, in step S2, the covariance matrix of the 3D Gaussian distribution is determined by parameterization based on the four-dimensional quaternion vector and the three-dimensional scale vector.
[0024] Reference Figure 3 As shown, a 3D Gaussian distribution is used. To represent a 3D scene. The 3D Gaussian function is composed of a 3×3 covariance matrix. and three-dimensional average vector Parameterization. Covariance matrix This represents the range of the Gaussian function in three-dimensional space, while Indicates its location or center. Covariance A four-dimensional quaternion vector describing rotation The kernel K of the multivariate Gaussian distribution is defined as follows: (The three-dimensional scale vector s represents the distribution.) To obtain the Gaussian meta-element, the 3D Gaussian function also extends two additional properties: opacity. and eigenvectors Opacity describes the transparency of a primitive, while The display color of the primitives was modulated.
[0025] 3DGS still requires an initial point cloud as input. Regardless of whether the initial point cloud is obtained through the Structure for Motion (SfM) method or other means, it often exhibits uneven density distribution in Gaussian space. Training is then performed after inputting a sparse initial point cloud.
[0026] The goal of training is to select only those Gaussian distributions that appear in the camera view.
[0027] A 3D Gaussian distribution is projected onto a 2D image plane of the selected view. The resulting 2D Gaussian distribution... It is called splat. and It is a 2D projected view-dependent parameterization, while It is opacity. From Inherited feature vectors.
[0028] set up It is a list of splats sorted by depth, each splat covering one pixel. . This represents the Gaussian kernel. Alpha mixing can be described as follows: In the formula, pixel The final RGB color By traversing all of the list The pixels are calculated using splat. The contribution of splat to pixel color was weighed. Calculation The opacity and density (transparency) of a given splat are taken into account. ) and the transmittance of all its predecessors Therefore, the color effect of the first splatter is stronger than that of the subsequent splatters.
[0029] Step S3: Improve the adaptive density control algorithm by using dynamic gradient thresholding strategy, scene range correction strategy and 3DGS pruning strategy to dynamically optimize the quantity and spatial distribution of 3D Gaussian distribution; In a specific embodiment of the present invention, step S3, the dynamic gradient threshold strategy includes: adjusting the gradient threshold according to an exponential increase law during the training process, using a low threshold in the initial stage to quickly generate a Gaussian distribution, and gradually increasing the threshold in the later stage to suppress overfitting.
[0030] Specifically, to control the number of Gaussian points split or cloned during densification, 3DGS compares the cumulative positional gradient of each splat with a predefined threshold. Gaussian distributions with high gradients are selected for densification, indicating frequent movement during optimization. Intuitively, this approach assumes that if a Gaussian moves frequently, it effectively fills multiple locations at once. These Gaussian points are then split or cloned on two independent Gaussian distributions.
[0031] In traditional 3DGS, this threshold is set to a fixed value, which is not conducive to fast convergence. Specifically, in the early optimization steps, the scene contains only a few Gaussian distributions, which hinders convergence without generating new Gaussian distributions; conversely, at the end of training, the scene is already composed of many Gaussian distributions, where additional densification can lead to overfitting. To address these two issues, this patent proposes a gradient ascent threshold that starts with a low value, such as 0.0001, allowing many Gaussian distributions to be densified at the beginning, and ends with a high value, such as 0.0004, where only those Gaussian distributions with very high positional gradients are cloned and split. The exponential scheduling of the threshold Ti in iteration i is described as follows: In the formula, i is the current iteration number. This represents the total number of iterations. As the initial value, This is the final value of the gradient threshold.
[0032] In a specific embodiment of the present invention, step S3 includes: combining point cloud data based on motion recovery structure with camera position to determine the actual range of the scene and optimizing the cloning and splitting decisions of Gaussian distribution.
[0033] Specifically, in 3D Gaussian point rendering, the scene range This is used to determine whether to clone or split a Gaussian distribution that meets the gradient threshold condition. (Larger) Values favor cloning, while smaller values favor cloning. A value that favors splitting is also beneficial. Furthermore, This also affects size pruning because the Gaussian distribution's size is evaluated relative to the scene extent. In the default implementation of 3DGS, The value is calculated based on the camera position, using the following formula: in, N cam Indicates the number of camera views. C i Indicates the first iThe formula shows that the scene extent is proportional to the distance from the farthest camera position to the average camera position. This calculation method makes the optimization algorithm heavily dependent on the scene capture process. For example, if a camera is shooting around a small object in a large environment, the scene extent may be very small, even if the actual scene is large. Current scene extent calculations only reflect the distance between camera positions, not the actual scene volume. During training, this bias can cause some Gaussian distributions to be split or pruned, even if they are appropriately sized for such large scenes. To address this issue, this application proposes a modified scene extent that does not depend on camera positions but rather on the SfM point cloud used for 3DGS initialization. Specifically, the new scene extent formula proposed in this patent is as follows: Where, N SfM Let p be the number of SfM points, and C be the average position of the camera in the formula. In general, this correction provides a scene range of the same order of magnitude, better corresponding to the natural perception of each scene.
[0034] In a specific embodiment of the present invention, the scene range correction strategy further includes: when there are dynamic objects in the scene, abnormal point cloud data is removed by time sequence consistency detection before determining the actual range of the scene.
[0035] In a specific embodiment of the present invention, the actual range of the scenario is determined according to the following formula: In the formula, N SfM The number of point clouds generated by the structure-of-motion (SFM) method. p represents the average position of the camera in the formula. i Here are the coordinates of the point cloud.
[0036] In a specific embodiment of the present invention, in step S3, the 3DGS pruning strategy includes: pruning distributions that contribute little to scene fidelity based on the opacity of the Gaussian distribution and the cumulative value of its alpha blending coefficient in all views.
[0037] In one specific embodiment of the present invention, the cumulative value of the alpha mixing coefficient is determined according to the following formula: In the formula, The k-th Gaussian distribution contributes weights to the color of pixel p. This represents the number of pixels in the image.
[0038] Step S4: Use a differentiable renderer to synthesize the optimized 3D Gaussian distribution from a new perspective and output the 3D reconstruction results.
[0039] By considering opacity in pruning decisions and alpha mixing coefficient This application can more accurately evaluate the contribution of Gaussians to the scene. This accurate evaluation enables reliable pruning decisions to be made throughout the scene training process. Therefore, this application more effectively optimizes the trade-off between reconstruction quality and scene compactness.
[0040] Example 2 One embodiment of the present invention provides a 3DGS three-dimensional reconstruction system based on ADC optimization, referring to... Figure 2 As shown, it includes: Image acquisition module 10: used to generate an initial sparse point cloud based on multi-view images through motion reconstruction structure; Initialization module 20: used to initialize a 3D Gaussian distribution based on the initial sparse point cloud, each 3D Gaussian distribution includes position, covariance matrix, opacity and color feature parameters; In one embodiment of the present invention, the covariance matrix of the 3D Gaussian distribution is determined by parameterization based on a four-dimensional quaternion vector and a three-dimensional scale vector.
[0041] Dynamic optimization module 30: used to improve the adaptive density control algorithm through dynamic gradient threshold strategy, scene range correction strategy and 3DGS pruning strategy, so as to dynamically optimize the quantity and spatial distribution of 3D Gaussian distribution; In one embodiment of the present invention, the dynamic gradient threshold strategy includes: adjusting the gradient threshold according to an exponential increase during training, using a low threshold in the initial stage to quickly generate a Gaussian distribution, and gradually increasing the threshold in the later stage to suppress overfitting.
[0042] In one embodiment of the present invention, the gradient threshold is determined according to the following formula: In the formula, i This represents the current iteration number. This represents the total number of iterations. As the initial value, This is the final value of the gradient threshold.
[0043] In one embodiment of the present invention, the scene range correction strategy includes: combining point cloud data based on motion recovery structure with camera position to determine the actual scene range and optimizing the cloning and splitting decisions of Gaussian distribution.
[0044] In one embodiment of the present invention, the scene range correction strategy further includes: when there are dynamic objects in the scene, abnormal point cloud data is removed by time sequence consistency detection before determining the actual range of the scene.
[0045] In one embodiment of the present invention, the actual range of the scenario is determined according to the following formula: In the formula, N SfM The number of point clouds generated by the structure-of-motion (SFM) method. p represents the average position of the camera in the formula. i Here are the coordinates of the point cloud.
[0046] In one embodiment of the present invention, in step S3, the 3DGS pruning strategy includes: pruning distributions that contribute little to scene fidelity based on the opacity of the Gaussian distribution and the cumulative value of its alpha blending coefficients in all views.
[0047] In one embodiment of the present invention, the cumulative value of the alpha mixing coefficient is determined according to the following formula: In the formula, The k-th Gaussian distribution contributes weights to the color of pixel p. This represents the number of pixels in the image.
[0048] Output module 40: Used to synthesize the optimized 3D Gaussian distribution from a new perspective using a differentiable renderer, and output the 3D reconstruction results.
[0049] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.
[0050] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on its differences from other embodiments. Similar or identical parts between embodiments can be referred to interchangeably. The embodiments of the present invention are described with reference to flowchart illustrations and / or block diagrams of methods, terminal devices (systems), and computer program products according to embodiments of the present invention. It should be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing terminal device to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing terminal device, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1The computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing terminal device to operate in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The functions specified in one or more boxes. These computer program instructions may also be loaded onto a computer or other programmable data processing terminal equipment to cause a series of operational steps to be performed on the computer or other programmable terminal equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable terminal equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the functions specified in one or more boxes. Although preferred embodiments of the invention have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments as well as all changes and modifications falling within the scope of the embodiments of the invention. Finally, it should be noted that in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Moreover, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or terminal device that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or terminal device. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or terminal device that includes said element.
[0051] The methods and apparatus provided by the present invention have been described in detail above. Specific examples have been used to illustrate the principles and implementation methods of the present invention. The descriptions of the above embodiments are only for the purpose of helping to understand the methods and core ideas of the present invention. At the same time, for those skilled in the art, there will be changes in the specific implementation methods and application scope based on the ideas of the present invention. Therefore, the content of this specification should not be construed as a limitation of the present invention.
[0052] In the description of this specification, references to terms such as "an embodiment," "some embodiments," "example," "specific example," or "a specific embodiment" or "some examples," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of the invention. In this specification, illustrative expressions of terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples.
[0053] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this application.
Claims
1. A 3DGS three-dimensional reconstruction method based on ADC optimization, characterized in that, include: Step S1: Generate an initial sparse point cloud based on the motion recovery structure from the multi-view images; Step S2: Initialize a 3D Gaussian distribution based on the initial sparse point cloud. Each 3D Gaussian distribution includes position, covariance matrix, opacity, and color feature parameters. Step S3: Improve the adaptive density control algorithm by using a dynamic gradient threshold strategy, a scene range correction strategy, and a 3DGS pruning strategy to dynamically optimize the quantity and spatial distribution of the 3D Gaussian distribution; Step S4: Use a differentiable renderer to synthesize the optimized 3D Gaussian distribution from a new perspective and output the 3D reconstruction result.
2. The 3DGS three-dimensional reconstruction method based on ADC optimization according to claim 1, characterized in that, In step S2, the covariance matrix of the 3D Gaussian distribution is determined by parameterization based on the four-dimensional quaternion vector and the three-dimensional scale vector.
3. The 3DGS three-dimensional reconstruction method based on ADC optimization according to claim 1, characterized in that, In step S3, the dynamic gradient threshold strategy includes: adjusting the gradient threshold according to an exponential increase during training, using a low threshold in the initial stage to quickly generate a Gaussian distribution, and gradually increasing the threshold in the later stage to suppress overfitting.
4. The 3DGS three-dimensional reconstruction method based on ADC optimization according to claim 3, characterized in that, The gradient threshold is determined according to the following formula: In the formula, i This represents the current iteration number. This represents the total number of iterations. As the initial value, This is the final value of the gradient threshold.
5. The 3DGS three-dimensional reconstruction method based on ADC optimization according to claim 1, characterized in that, In step S3, the scene range correction strategy includes: combining point cloud data based on motion recovery structure with camera position to determine the actual scene range and optimizing the cloning and splitting decisions of Gaussian distribution.
6. The 3DGS three-dimensional reconstruction method based on ADC optimization according to claim 5, characterized in that, The actual scope of the scenario is determined through the following steps: When there are dynamic objects in the scene, abnormal point cloud data is removed through temporal consistency detection before the actual range of the scene is determined.
7. The 3DGS three-dimensional reconstruction method based on ADC optimization according to claim 6, characterized in that, The actual range of the scenario is determined according to the following formula: In the formula, N SfM The number of point clouds generated by the motion structure recovery method. p represents the average position of the camera in the formula. i Here are the coordinates of the point cloud.
8. The 3DGS three-dimensional reconstruction method based on ADC optimization according to claim 1, characterized in that, In step S3, the 3DGS pruning strategy includes: pruning distributions that contribute little to scene fidelity based on the opacity of the Gaussian distribution and the cumulative value of its alpha blending coefficients across all views.
9. The 3DGS three-dimensional reconstruction method based on ADC optimization according to claim 8, characterized in that, The cumulative value of the alpha mixing coefficient is determined according to the following formula: In the formula, The k-th Gaussian distribution contributes weights to the color of pixel p. This represents the number of pixels in the image.
10. A 3DGS three-dimensional reconstruction system based on ADC optimization, characterized in that, include: Image acquisition module: used to generate an initial sparse point cloud based on multi-view images through motion reconstruction structure; Initialization module: used to initialize a 3D Gaussian distribution based on the initial sparse point cloud. Each 3D Gaussian distribution includes position, covariance matrix, opacity and color feature parameters. Dynamic optimization module: used to improve the adaptive density control algorithm through dynamic gradient threshold strategy, scene range correction strategy and 3DGS pruning strategy, so as to dynamically optimize the quantity and spatial distribution of the 3D Gaussian distribution; The results output module is used to synthesize the optimized 3D Gaussian distribution from a new perspective using a differentiable renderer and output the 3D reconstruction results.