Smooth surface modeling method based on 3D Gaussian

Abstract

The invention discloses a smooth surface modeling method based on 3D Gaussian, and the method comprises the following steps: carrying out data preprocessing: carrying out the motion structure matching of a collected image group, carrying out the camera parameter reasoning, obtaining a corresponding camera parameter, and constructing an initialization point cloud; performing mask estimation; performing normal estimation; performing training modeling, including loading image group camera parameters obtained by the preprocessing part, initializing a point cloud, a mask image group and normal direction data, initializing to construct a 3D Gaussian set, and obtaining a more accurate sparse Gaussian set through initialization training iteration; modeling training is carried out by starting from a sparse Gaussian set of an initialization module in combination with an image sequence and data in a normal direction, fine adjustment is carried out on the model, finally model modeling is completed, and a 3D Gaussian model is obtained; and carrying out grid conversion, and carrying out grid Poisson reconstruction according to the point cloud with the normal direction obtained by rendering sampling to obtain a final grid model.

Description

Technical Field

[0001] The invention belongs to the technical field of 3D Gaussian modeling, and in particular relates to a smooth surface modeling method based on 3D Gaussian. Background Art

[0002] The task of 3D Gaussian mesh modeling is to take a series of images or video data (including the captured images and the internal and external parameter matrices of the camera corresponding to each image), initialize a 3D Gaussian point set through the captured data, train the 3D Gaussian point set to obtain a representation of a 3D static scene, and finally convert the static scene into a mesh model suitable for industrial editing. 3D Gaussian modeling first constructs a sparse scene point cloud based on the captured images through motion structure matching (Structure From Motion, SFM). The point cloud is used to initialize the 3D Gaussian point cloud set. During the training process, a high-quality 3D static scene representation is obtained by continuously optimizing the 3D Gaussian point cloud set. The 3D static scene representation is converted into a mesh model through mesh reconstruction technology.

[0003] In the real modeling process, the modeled objects often have smooth surfaces (such as the paint surface of a thermos cup). It is very difficult to model the features of such objects. The smooth surface areas in the modeling results will have serious depressions and distortions. The usual modeling processing method is to attach feature label points to the object to increase the standard features of the object and the modeling accuracy, and then remove the label points through post-processing. This processing method requires a lot of manual participation and cannot completely restore the surface of the object. Summary of the invention

[0004] In view of the above problems, the present invention provides a smooth surface modeling method based on 3D Gaussian, which optimizes 3D Gaussian modeling for objects with smooth surfaces, reduces the problems of depression and distortion in the modeling results, and reduces the cost of manual processing after modeling.

[0005] In order to solve the above technical problems, the present invention adopts the following technical solutions:

[0006] A 3D Gaussian-based smooth surface modeling method comprises the following steps:

[0007] Perform data preprocessing, including: performing motion structure matching on the captured image group, inferring camera parameters to obtain corresponding camera parameters, and constructing an initial point cloud; performing mask estimation, directly detecting each frame of the captured image group using a deep learning algorithm, saving the detected objects to be modeled as rgb = [255, 255, 255], and directly saving the parts that do not need to be modeled as rgb = [0, 0, 0] to obtain a binary black-and-white mask image; performing normal estimation, calculating the monocular normal direction data for each frame of the captured image group using a deep learning algorithm, and normalizing the values to [-1, 1] for saving;

[0008] Perform training and modeling, including: loading the image group camera parameters, initial point cloud, mask image group, and normal direction data obtained from the preprocessing part, initializing and constructing a 3D Gaussian set, and obtaining a more accurate sparse Gaussian set through initial training iterations; starting from the sparse Gaussian set of the initialization module, jointly performing modeling training on the image sequence and normal direction data, fine-tuning the model through iterations, and finally completing the model modeling to obtain a 3D Gaussian model;

[0009] Perform mesh conversion, including: loading the modeled 3D Gaussian model and image group camera parameters, performing rendering sampling at each camera angle to obtain a point cloud with normal directions; performing mesh Poisson reconstruction based on the rendered and sampled point cloud with normal directions to obtain the final mesh model.

[0010] In a possible implementation, performing motion structure matching on the captured image group, inferring camera parameters to obtain corresponding camera parameters, and constructing an initial point cloud specifically includes: performing structure from motion (SFM) on the captured N image groups Y = y i (0 < i ≤ N) to obtain the corresponding camera parameters ImageCams, where the camera parameters include the focal length, distortion parameters, and internal and external camera parameter matrices of the position matrix corresponding to each image, and constructing an initial scene point cloud X = x " (0 < k < M), including the spatial coordinates xyz and color values rgb of M points, and performing model training on ImageCams and the initial scene point cloud,

[0011] ImageCams, X = SFM(Y).

[0012] In a possible implementation, for the mask estimation, directly detecting each frame of the captured image group using a deep learning algorithm, saving the detected objects to be modeled as rgb = [255, 255, 255], and directly saving the parts that do not need to be modeled as rgb = [0, 0, 0] to obtain a binary black-and-white mask image specifically includes: performing structure from motion (SFM) on the captured N image groups Y = y i(0 < i ≤ N) are sequentially input into the mask estimation model. N image groups Y are input, and the mask estimation model outputs N mask map groups Y fg , Y fg represents an image group with an object to be modeled. The object to be modeled is represented by a white area with rgb = [255, 255, 255] as the valid area; the masked area is filled with a filling color of rgb = [0, 0, 0] as the invalid area.

[0013] In a possible implementation, for the normal estimation, the captured picture group is used to calculate the monocular normal direction data frame by frame using a deep learning algorithm, and the values are normalized to [-1, 1] for storage. Specifically, it includes: the N image groups Y = y i (0 < i ≤ N) are sequentially input into the normal estimation model. N image groups Y are input, and the normal estimation model outputs N corresponding normal direction data groups Nor. Nor represents the normal direction corresponding to the image group Y, and the data of the normal direction are normalized to [-1, 1].

[0014] In a possible implementation, loading the image group camera parameters, initializing the point cloud, mask map group, and normal direction data obtained in the preprocessing part, initializing and constructing a 3D Gaussian set, and obtaining a more accurate sparse Gaussian set through initial training iteration specifically includes: loading the camera parameters ImageCams, initializing the point cloud X, normal direction data group Nor, and mask map group Y fg , first according to the loaded mask map group Y fg , delete the points in the masked area in the initialized point cloud X to obtain the point cloud X'. Then, according to the point cloud X', create a group of 3D Gaussian point clouds G. Each point in the point cloud X' corresponds to a 3D Gaussian point. The Gaussian point includes a 3D ellipsoid of position mean, covariance matrix Σ, opacity α, and spherical harmonic function SH. When multiple adjusted ellipsoids are superimposed, the real object will be reproduced;

[0015] The xyz coordinates of the point represent the position mean of the 3D Gaussian point cloud. Initialize the covariance matrix Σ based on the point, and the specific definition is as follows,

[0016]

[0017] According to the order d of the spherical harmonic function SH sh Initialize the parameters with 0, and perform partial initialization for the spherical harmonic function SH based on the rgb value of the point,

[0018] SH = zero((d sh + 1) 2 )[rgb,..., 0]

[0019] The opacity α of each point is initialized to 0.1, and the opacity α is processed by the sigmoid function during calculation, and the adjusted opacity α is output at this time s , α s The value of is restricted to the range [0, 1).

[0020] α s =sigmoid(α)

[0021] ImageCams and 3D Gaussian point clouds use differentiable rasterization to render the effective area in the mask image and obtain the normal direction data predicted by the model. and the predicted image No images are used during the initialization phase Constraint, using only normal direction data The loss is calculated using the normal direction data Nor obtained by the normal estimation model, and the gradient is returned to update the parameters of the 3D Gaussian point cloud. The loss formula is:

[0022]

[0023] in the formula Represents the gradient of the normal direction data, and λ represents the weight of the loss;

[0024] Repeat the above update process t times, 1000<t<3000, and get a sparse Gaussian set X gs .

[0025] In a possible implementation, starting from the sparse Gaussian set of the initialization module, combining the image sequence and the normal direction data for modeling training, fine-tuning the model through generations, and finally completing the modeling to obtain a 3D Gaussian model specifically includes:

[0026] The sparse Gaussian set X gs As initial input, load the camera parameters ImageCams, normal direction data Nor, mask map Y fg and image group image Y; in mask image Y fg In the effective area, ImageCams and sparse Gaussian set X gs Rendering through differentiable rasterization to obtain predicted normal direction data and the predicted image At the same time, the normal direction data Nor and image Y are used as modeling constraints to calculate the loss, and the gradient is returned to update the parameters of the 3D Gaussian point cloud. The new loss formula is as follows, where L Y is the image consistency constraint, L Nor is the surface consistency constraint, and L is the overall loss of the training process:

[0027] LY =(1-θ)L1+θL ssim

[0028]

[0029] L=L Y +L Nor

[0030] Repeat the training and updating process, the 3D Gaussian modeling model will gradually become clear. When it iterates to the preset number of iterations T, 20000<T<30000, the 3D Gaussian modeling is completed.

[0031] The use of the present invention has the following beneficial effects: due to the reflection and pure color of the smooth surface, it is impossible to accurately model the specific plane depth position of the object during 3D Gaussian modeling, resulting in depressions and distortions in the smooth surface area; the technical solution of the embodiment of the present invention uses the prior monocular normal information and surface consistency constraints predicted by the normal estimation model, so that the modeling process can use the prior knowledge of the normal estimation model to confirm and model the plane depth position of the smooth surface with reference, thereby correctly modeling these areas during the modeling process, reducing the problems of depressions and distortions, and reducing the cost of manual processing after modeling. BRIEF DESCRIPTION OF THE DRAWINGS

[0032] Figure 1 A flowchart of a 3D Gaussian-based smooth surface modeling method according to an embodiment of the present invention;

[0033] Figure 2 A schematic diagram comparing the modeling effect of a 3D Gaussian-based smooth surface modeling method in a specific application example of the present invention with the modeling effect of the prior art;

[0034] Figure 3 It is a schematic diagram comparing the modeling effect of the 3D Gaussian-based smooth surface modeling method in another specific application example of the present invention with the modeling effect of the prior art. DETAILED DESCRIPTION

[0035] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0036] See also Figure 1 , which is a flow chart of a method for modeling a smooth surface based on 3D Gaussian according to an embodiment of the present invention, comprising the following steps:

[0037] Perform data preprocessing, including: performing motion structure matching on the collected image group, inferring camera parameters to obtain corresponding camera parameters, and constructing an initial point cloud; performing mask estimation, directly detecting each frame of the collected image group using a deep learning algorithm, saving the detected objects to be modeled as rgb = [255, 255, 255], and directly saving the parts that do not need to be modeled as rgb = [0, 0, 0] to obtain a binary black and white mask image; performing normal estimation, calculating the monocular normal direction data for each frame of the collected image group using a deep learning algorithm, and normalizing the values to [-1, 1] for saving.

[0038] Perform training and modeling, including: loading the image group camera parameters, initial point cloud, mask image group, and normal direction data obtained in the preprocessing part, initializing and constructing a 3D Gaussian set, and obtaining a more accurate sparse Gaussian set through initial training iteration; starting from the sparse Gaussian set of the initialization module, jointly performing modeling training on the image sequence and normal direction data, fine-tuning the model through iterations, and finally completing the model modeling to obtain a 3D Gaussian model.

[0039] Perform mesh conversion, including: loading the modeled 3D Gaussian model and the image group camera parameters ImageCams, performing rendering sampling at each camera angle to obtain a point cloud M with normal directions; performing mesh Poisson reconstruction based on the rendered and sampled point cloud with normal directions to obtain the final mesh model.

[0040] In a smooth surface modeling method based on 3D Gaussian according to another embodiment of the present invention, performing motion structure matching on the collected image group, inferring camera parameters to obtain corresponding camera parameters, and constructing an initial point cloud specifically includes: performing Structure from Motion (SFM) on the captured N-image group Y = y i (0 < i ≤ N) to obtain the corresponding camera parameters ImageCams, where the camera parameters include the focal length, distortion parameters, and internal and external camera parameter matrices of the position matrix corresponding to each image, and constructing an initial scene point cloud X = x"(0 < k < M), including the spatial coordinates xyz and color value rgb of M points, and performing model training on ImageCams and the initial scene point cloud,

[0041] ImageCams, X = SFM(Y).

[0042] In a method for smooth surface modeling based on 3D Gaussian in another embodiment of the present invention, mask estimation is performed. The captured image group is directly detected frame by frame using a deep learning algorithm. The detected object to be modeled is saved as rgb = [255, 255, 255], and the part that does not need to be modeled is directly saved as rgb = [0, 0, 0], obtaining a binary black and white mask image. Specifically, it includes: The captured N image groups Y = y i (0 < i ≤ N) are sequentially input into the mask estimation model. By inputting N image groups Y, the mask estimation model outputs N mask image groups Y fg , Y fg represents the image group with the object to be modeled, which will be used as auxiliary data for model modeling. The object to be modeled is represented by a white area with rgb = [255, 255, 255] as the valid area; the masked area is filled with a filling color of rgb = [0, 0, 0] as the invalid area.

[0043] In a method for smooth surface modeling based on 3D Gaussian in another embodiment of the present invention, normal estimation is performed. The captured image group is used to calculate the monocular normal direction data frame by frame using a deep learning algorithm, and the numerical values are normalized to [-1, 1] for storage. Specifically, it includes: The captured N image groups Y = y i (0 < i ≤ N) are sequentially input into the normal estimation model. By inputting N image groups Y, the normal estimation model outputs N corresponding normal direction data groups Nor. Nor represents the normal direction corresponding to the image group Y. The Nor image group will be used as one of the main data for model modeling, and the data of the normal direction is normalized to [-1, 1].

[0044] In a method for smooth surface modeling based on 3D Gaussian in another embodiment of the present invention, the image group camera parameters, initial point cloud, mask image group, and normal direction data obtained in the preprocessing part are loaded, and a 3D Gaussian set is initialized and constructed, and a more accurate sparse Gaussian set is obtained through initialization training iteration. Specifically, it includes: Loading the camera parameters ImageCams, initializing the point cloud X, the normal direction data group Nor, and the mask image group Y fg , first, according to the loaded mask image group Y fg , the points in the masked area in the initial point cloud X are deleted to obtain the point cloud X'. Then, according to the point cloud X', a group of 3D Gaussian point clouds G is created. Each point in the point cloud X' corresponds to a 3D Gaussian point. The Gaussian point includes a 3D ellipsoid of position mean, covariance matrix ∑, opacity α, and spherical harmonic function SH. When multiple adjusted ellipsoids are superimposed, the real object will be reproduced;

[0045] The xyz coordinates of the points represent the position mean of the 3D Gaussian point cloud. The covariance matrix ∑ is initialized based on the points, and the specific definition is as follows,

[0046]

[0047] According to the order d of the spherical harmonic function SH sh Initialize the parameters with 0, and partially initialize the spherical harmonics SH based on the rgb value of the point.

[0048] SH=zero((d sh +1) 2 )[rgb, ..., 0]

[0049] The opacity α of each point is initialized to 0.1, and the opacity α is processed by the sigmoid function during calculation, and the adjusted opacity α is output at this time s , α s The value of is restricted to the range [0, 1).

[0050] α s =sigmoid(α)

[0051] ImageCams and 3D Gaussian point clouds use differentiable rasterization to render the effective area in the mask image and obtain the normal direction data predicted by the model. and the predicted image No images are used during the initialization phase Constraint, using only normal direction data The loss is calculated using the normal direction data Nor obtained by the normal estimation model, and the gradient is returned to update the parameters of the 3D Gaussian point cloud. The loss formula is:

[0052]

[0053] in the formula Represents the gradient of the normal direction data, and λ represents the weight of the loss;

[0054] Repeat the above update process t times, 1000<t<3000, and get a sparse Gaussian set X gs , sparse Gaussian set X gs Will be used for joint training.

[0055] In a smooth surface modeling method based on 3D Gaussian in another embodiment of the present invention, starting from a sparse Gaussian set of an initialization module, modeling training is performed in conjunction with an image sequence and normal direction data, the model is fine-tuned through generations, and finally the model modeling is completed to obtain a 3D Gaussian model, which specifically includes:

[0056] The sparse Gaussian set X gs As initial input, load the camera parameters ImageCams, normal direction data Nor, mask map Y fg and image group image Y; in mask image Yfg In the effective area, ImageCams and sparse Gaussian set X gs Rendering through differentiable rasterization to obtain predicted normal direction data and the predicted image At the same time, the normal direction data Nor and image Y are used as modeling constraints to calculate the loss, and the gradient is returned to update the parameters of the 3D Gaussian point cloud. The new loss formula is as follows, where L Y is the image consistency constraint, L Nor is the surface consistency constraint, and L is the overall loss of the training process:

[0057] L Y =(1-θ)L1+θL ssim

[0058]

[0059] L=L Y +L Nor

[0060] Repeat the training and updating process, the 3D Gaussian modeling model will gradually become clear. When it iterates to the preset number of iterations T, 20000<T<30000, the 3D Gaussian modeling is completed.

[0061] In a specific application example, a mesh Poisson reconstruction is performed based on a point cloud with a normal direction sampled by rendering to obtain a final mesh model. Specifically, the Poisson reconstruction part loads the point cloud M obtained by rendering sampling, constructs a mesh patch based on the neighborhood relationship and normal direction of the point cloud M, and then uniformly calculates the mesh patch color based on multiple vertex colors of the mesh patch, and finally obtains the constructed mesh model.

[0062] See also Figure 2 and Figure 3 , respectively, are schematic diagrams comparing the modeling effects of the smooth surface modeling method based on 3D Gaussian and the modeling effects of the prior art in two specific application examples, wherein the left side is the modeling effect diagram of the prior art, and the right side is the modeling effect diagram of the method using the embodiment of the present invention. It can be seen that the method using the embodiment of the present invention solves the problem of large-area depression and distortion in 3D Gaussian modeling when modeling objects with smooth surfaces. Targeted optimization effectively improves the modeling quality and reduces the labor cost of post-processing.

[0063] It should be understood that the exemplary embodiments described herein are illustrative rather than restrictive. Although one or more embodiments of the present invention have been described in conjunction with the accompanying drawings, it should be understood by those skilled in the art that various changes in form and detail may be made without departing from the spirit and scope of the present invention as defined by the appended claims.

Claims

1. A smooth surface modeling method based on 3D Gaussian, characterized in that: Including the following steps: Perform data preprocessing, including: performing motion structure matching on the collected image group, inferring the corresponding camera parameters through camera parameter inference, and constructing an initial point cloud; performing mask estimation, directly detecting each frame of the collected image group using a deep learning algorithm, saving the detected objects to be modeled as rgb = [255, 255, 255], and directly saving the parts that do not need to be modeled as rgb = [0, 0, 0] to obtain a binary black and white mask image; performing normal estimation, calculating the monocular normal direction data frame by frame on the collected image group using a deep learning algorithm, and normalizing the values to [-1, 1] for saving; Perform training and modeling, including: loading the image group camera parameters, initial point cloud, mask image group, and normal direction data obtained in the preprocessing part, initializing and constructing a 3D Gaussian set, and obtaining a more accurate sparse Gaussian set through initial training iteration; starting from the sparse Gaussian set of the initialization module, jointly performing modeling training on the image sequence and normal direction data, fine-tuning the model through iterations, and finally completing the model modeling to obtain a 3D Gaussian model; Perform mesh conversion, including: loading the well-modeled 3D Gaussian model and image group camera parameters, performing rendering sampling at each camera angle to obtain a point cloud with normal directions; performing mesh Poisson reconstruction based on the point cloud with normal directions sampled by rendering to obtain the final mesh model.

2. The 3D Gaussian-based smooth surface modeling method according to claim 1, characterized in that: Performing motion structure matching on the collected image group, inferring camera parameters to obtain corresponding camera parameters, and constructing an initial point cloud specifically includes: The N captured image groups Y = y i (0 < i ≤ N) are processed through Structure from Motion (SFM) to obtain the corresponding camera parameters ImageCams, where the camera parameters include the focal length, distortion parameters, and internal and external parameter matrices of the position matrix corresponding to each image, and an initial scene point cloud X = x k (0 < k < M), including the spatial coordinates xyz and color values rgb of M points, and performing model training on ImageCams and the initial scene point cloud ImageCams, X = SFM(Y).

3. The 3D Gaussian-based smooth surface modeling method according to claim 2, characterized in that: Perform the above-mentioned mask estimation, directly detect each frame of the collected image group using a deep learning algorithm, save the detected objects to be modeled as rgb = [255, 255, 255], and directly save the parts that do not need to be modeled as rgb = [0, 0, 0], to obtain a binary black-and-white mask image, which specifically includes: input the captured N image groups Y = y i (0 < i ≤ N) into the mask estimation model in sequence. Input the N image groups Y, and the mask estimation model outputs N mask image groups Y fg , Y fg represents the image group with the objects to be modeled, and the objects to be modeled are represented by the white area of rgb = [255, 255, 255] as the valid area; the masked area is filled with the fill color of rgb = [0, 0, 0] as the invalid area.

4. The 3D Gaussian-based smooth surface modeling method according to claim 3, characterized in that: Perform the normal estimation. Use a deep learning algorithm to calculate the monocular normal direction data frame by frame for the captured image group, and normalize the values to [-1, 1] for saving. Specifically, it includes: Input the captured N-image group Y = y i (0 < i ≤ N) into the normal estimation model in sequence. Input the N-image group Y, and the normal estimation model outputs N corresponding normal direction data groups Nor. Nor represents the normal direction corresponding to the image group Y, and the data of the normal direction is normalized to [-1, 1].

5. The 3D Gaussian-based smooth surface modeling method according to claim 4, characterized in that: The loading of the image group camera parameters, initialization of point cloud, mask group and normal direction data obtained in the preprocessing part, initialization of a 3D Gaussian set, and obtaining a more accurate sparse Gaussian set through initialization training iterations specifically include: loading camera parameters ImageCams, initialization of point cloud X, normal direction data group Nor and mask group Y fg , first according to the loaded mask group Y fg , delete the points in the masked area of ​​the initial point cloud X to obtain the point cloud X', and then create a set of 3D Gaussian point cloud sets G based on the point cloud X'. Each point in the point cloud X' corresponds to a 3D Gaussian point. The Gaussian point includes a 3D ellipsoid of position mean, covariance matrix Σ, opacity α and spherical harmonic function SH. When multiple ellipsoids with adjusted parameters are superimposed, the real object will be reproduced; The xyz coordinates of the points represent the position mean of the 3D Gaussian point cloud. Based on the points, the covariance matrix Σ is initialized, and the specific definition is as follows. According to the order d of the spherical harmonic function SH sh Initialize the parameters with 0, and partially initialize the spherical harmonics SH based on the rgb value of the point. SH=zero((d sh +1) 2 )[rgb,…,0] The opacity α of each point is initialized to 0.1, and the opacity α is processed by the sigmoid function during calculation, and the adjusted opacity α is output at this time s , α s The value of is restricted to the range [0,1). a s =sigmoid(a) ImageCams and 3D Gaussian point clouds use differentiable rasterization to render the effective area in the mask image and obtain the normal direction data predicted by the model. and the predicted image No images are used during the initialization phase Constraint, using only normal direction data The loss is calculated using the normal direction data Nor obtained by the normal estimation model, and the gradient is returned to update the parameters of the 3D Gaussian point cloud. The loss formula is: in the formula Represents the gradient of the normal direction data, and λ represents the weight of the loss; Repeat the above update process \(t\) times, where \(1000 < t < 3000\), to obtain a sparse Gaussian set \(X\) gs .

6. The 3D Gaussian-based smooth surface modeling method according to claim 5, characterized in that: The specific process of starting from the sparse Gaussian set of the initialization module, jointly performing modeling training on the image sequence and normal direction data, fine-tuning the model through iterations, and finally completing the model modeling to obtain a 3D Gaussian model includes: The sparse Gaussian set X gs As initial input, load the camera parameters ImageCams, normal direction data Nor, mask map Y fg and image group image Y; in mask image Y fg In the effective area, ImageCams and sparse Gaussian set X gs Rendering through differentiable rasterization to obtain predicted normal direction data and the predicted image At the same time, the normal direction data Nor and image Y are used as modeling constraints to calculate the loss, and the gradient is returned to update the parameters of the 3D Gaussian point cloud. The new loss formula is as follows, where L Y is the image consistency constraint, L Nor is the surface consistency constraint, and L is the overall loss of the training process: L Y =(1-θ)L1+θL ssim L=L Y +L Nor Repeating the training and update process, the model of 3D Gaussian modeling will gradually become clear. When the iteration reaches the preset iteration number T, 20000 < T < 30000, the 3D Gaussian modeling is completed.

Images