A sparse-view 3D Gaussian sintering reconstruction method based on adaptive view sampling

Abstract

The application discloses a sparse view 3D Gaussian sputtering reconstruction method based on adaptive view sampling, which improves the training strategy of 3DGS without introducing external geometric prior, ensures the randomness of view selection, increases the training frequency of large reconstruction error views, reduces the overfitting risk of small reconstruction error views, and reasonably allocates training resources. In addition, while the dropout strategy is used to improve the synthesis result, self-consistency regularization is introduced to measure the robustness of the model, guide the model to further improve the stability of the reconstruction result, and reduce the influence of noise caused by adaptive density control.

Description

TECHNICAL FIELD

[0001] The application belongs to the field of computer vision, and particularly relates to a sparse-viewpoint 3D Gaussian splatting reconstruction method based on adaptive viewpoint sampling. BACKGROUND

[0002] New-viewpoint synthesis technology has attracted wide attention in recent years in the fields of virtual reality, robot vision, digital entertainment and three-dimensional asset generation, and as an explicit three-dimensional scene representation method, 3D Gaussian splatting (3DGS) has become a new generation of mainstream reconstruction framework after NeRF, with efficient rasterization rendering and training speed. However, 3DGS can obtain high-quality reconstruction under the condition of dense input viewpoints, but its performance will sharply decrease when only 3-5 sparse training viewpoints are available, which is manifested in problems such as floating artifacts, geometric distortion and insufficient generalization ability in unseen viewpoints. The root cause lies in that sparse viewpoints cannot provide sufficient geometric constraints, and the original 3DGS optimization process highly depends on photometric error signals, thereby leading to unstable training and structural uncertainty.

[0003] Improvements on the original 3DGS reconstruction method mainly introduce external geometric priors to assist in scene reconstruction. However, the method of introducing external geometric priors needs to use a pre-trained geometric data estimation model (such as a monocular depth estimation model), and the scale ambiguity problem existing in the geometric data estimation model will interfere with the normal training process of the 3DGS model.

[0004] In addition, a recent related literature proposes a training improvement method based on Dropout, which can enhance the robustness of the model without relying on external data, but ignores the structural consistency before and after dropout. When the key Gaussian cell is randomly discarded, chaotic and excessively large gradients will be generated, which will destroy the converged local structure and cause the model to more easily produce floating artifacts and geometric distortion under sparse viewpoints. SUMMARY

[0005] To solve the above problems, the application discloses a sparse-viewpoint 3D Gaussian splatting reconstruction method based on adaptive viewpoint sampling, which improves the training strategy of 3DGS without introducing external geometric priors, so that the model can adaptively identify the reconstruction difficulty of different viewpoints, thereby avoiding overfitting and underfitting under the condition of sparse viewpoints; and a stable internal consistency constraint mechanism is constructed to improve the geometric reliability and new-viewpoint generalization ability of 3DGS under the condition of few viewpoints.

[0006] To achieve the above purpose, the technical scheme of the application is as follows:

[0007] A sparse view 3D Gaussian splatting reconstruction method based on adaptive view sampling, comprising the following steps:

[0008] Step 1: Constructing a 3D GS reconstruction dataset

[0009] For a specific scene, a multi-view dataset containing a set of training view images , a set of test view images , initial point cloud data calculated by an SfM algorithm , and a set of camera pose data is constructed, wherein each camera pose data corresponds to the image data of one of the image sets or ;

[0010] Step 2: Training data initialization

[0011] The view data in step 1 is initialized, the view images are resized and linearly normalized, the initial point cloud data is converted into an initial 3D GS model , and each point in the model is assigned with position , covariance matrix (shape and rotation), spherical harmonic coefficient (color), opacity , etc., wherein the position directly uses the position of the initial point cloud, the average distance of each point and its nearest neighbors is calculated, the logarithm is taken as the initial scaling, and the initial spherical harmonic coefficient order is 0, i.e., the color of the point in the point cloud is directly taken as the initial color;

[0012] Step 3: Adaptive view sampling based on reconstruction error distribution

[0013] Based on the number and number of the obtained views, a reconstruction error value is initialized for each view (this index is used to quantify the difference between the model rendering image and the real image under the current view, and in the present application, the loss value of the view is used instead), and is recorded in the vector in the order of numbering, the temperature coefficient of the softmax function is initialized , and the smoothing factor of the reconstruction error update is initialized ;

[0014] The reconstruction error and the temperature coefficient stored in the current step vector are input into the softmax activation function to obtain the probability distribution of each view reconstruction error ; based on the probability distribution of each view, the views are randomly sampled to obtain the views (the training image containing the current viewpoint and the corresponding camera parameters);

[0015] Step 4: Randomly discard the Gaussian primitives

[0016] Calculate the random discard ratio for the Gaussian primitives at this time based on the current iteration step , and randomly discard the Gaussian primitives in the model to obtain the 3D Gaussian model after random primitive discard ;

[0017] Step 5: Two-way rasterization rendering

[0018] Combine the 3D Gaussian model after random discard obtained in step 4 with the viewpoint data obtained in step 3 , project the 3D Gaussian primitives in the view frustum into 2D splats, sort the splats according to the depth information, and perform alpha blending for each pixel to accumulate color and opacity (follow the original 3DGS rasterization renderer process) to generate the rendering result of the model after random discard operation for the current viewpoint and the corresponding depth map ; then, combine the 3D Gaussian model before random discard in step 4 with the viewpoint data obtained in step 3, and render the 3D Gaussian model again through the differentiable rasterization renderer to generate the complete rendering image for the current viewpoint and the corresponding depth map ;

[0019] Step 6: Loss calculation and self-consistency regularization

[0020] Compare the discarded rendering image obtained in step 5 with the training viewpoint image to calculate the photometric loss for the current viewpoint (including L1 pixel distance loss and D-SSIM structural similarity loss); compare the complete rendering image obtained in step 5 with the training viewpoint image to calculate the photometric loss of the complete rendering image ; subtract from to obtain the self-consistency loss , which is used as a regularization term to obtain the total training loss by weighted summation .

[0021] Step 7: Gradient calculation and parameter update

[0022] Based on the total training loss obtained in step 6 Perform backpropagation, calculate the gradient of the current iteration, and update the Gaussian model. Various attribute parameters; smoothing factor obtained in step 3 sum vector The past loss values ​​of the current viewpoint are recorded in the data, and the reconstruction error of the current viewpoint is smoothly updated using an exponential moving average (EMA) strategy.

[0023] Step 8: Adaptive density control and model parameter update

[0024] Within a preset iteration interval (100 intervals were used in the experiment), the cumulative gradient values ​​of each Gaussian element in the view space were calculated. If the current iteration number reaches the iteration number required for adaptive density control, the 3DGS model is adjusted based on the Gaussian element position gradients calculated in the current iteration. The Gaussian primitives are used for cloning, segmentation, and opacity reset operations. For gradients exceeding a set threshold... For Gaussian elements, if their scale is small, they are cloned; if their scale is large, they are segmented; if the opacity of the Gaussian element is less than a threshold... If the opacity is zero, then return to step 3 and continue iterating until the maximum number of iterations is reached.

[0025] Furthermore, the viewpoint reconstruction error vectors initialized in step 3... As shown in Equation 1, it is based on the viewpoint set. The unique identifiers of each viewpoint are sorted, meaning the reconstruction error stored therein is the cumulative loss value within the current iteration step for each viewpoint:

[0026] (1);

[0027] in, Indicates the first The cumulative loss value within the current iteration step for each viewpoint. During training, the algorithm tracks the difficulty of each viewpoint in real time.

[0028] In step 3, when selecting training viewpoints, a probabilistic sampling strategy based on the reconstruction error of the training viewpoints is adopted. That is, at the beginning of each training iteration, as shown in Equation 2, the loss vector of each viewpoint is used. Calculate the sampling probability distribution, the first... The probability of a viewpoint being selected The calculation method is as follows:

[0029] (2);

[0030] in, Indicates the first The cumulative loss value within the current iteration step for each viewpoint, and , This is the temperature coefficient of the softmax function, used to control the magnitude of the difference in probability as a function of reconstruction error. In experiments, it is typically set to a certain value. This ensures that viewpoints with higher reconstruction errors (i.e., viewpoints that are more difficult to reconstruct and have worse reconstruction results) have a higher probability of being selected for training.

[0031] Furthermore, in step 6, this invention, based on the Dropout algorithm, uses a self-consistency loss regularization method to guide the model training. A diagram of the self-consistency loss is shown below. Figure 3 As shown, its significance lies in improving the robustness and stability of the 3DGS model itself, ensuring that the model reconstruction result conforms to the real geometric structure.

[0032] Through the dual-path rasterization rendering process in step 5, the rendering results of the model after random discarding are obtained respectively. Rendering results of the model without random dropout The two images are compared with the training viewpoint image. Calculate the photometric loss. This includes the total loss during the training process. The calculation is shown in Equation 3:

[0033] (3);

[0034] in, is the proportionality coefficient of the regularization term. express and Photometric loss, which constitutes the main part of the total loss. The calculation method is shown in Equation 4:

[0035] (4);

[0036] in, represent The loss ratio coefficient. express and The loss of luminosity between them The calculation method is shown in Equation 5:

[0037] (5);

[0038] This is a model consistency regularization term, representing the degree of difference between the model and the original model after a slight perturbation caused by random dropout operations. To ensure the geometric consistency of the model, a depth map is added during rendering. and The L1 loss is used as a geometric regularization term. The calculation method is shown in Equation 6:

[0039] (6);

[0040] in, The weighting coefficients represent the depth L1 error. The higher the value, the higher the instability of the model itself.

[0041] Furthermore, the viewpoint loss update strategy used in step 7 is not a direct overlay, but rather an exponential moving average (EMA) strategy. As shown in Equation 7, the EMA strategy avoids drastic changes in viewpoint sampling probability caused by excessively large or small single loss values, and is specifically expressed in the following form:

[0042] (7);

[0043] in The smoothing coefficient set during the initialization phase.

[0044] The beneficial effects of this invention are as follows:

[0045] (1) This invention proposes a viewpoint sampling strategy based on viewpoint reconstruction error. Existing technologies typically employ "blind" uniform random sampling under sparse viewpoints, leading to overfitting on viewpoints with small errors and underfitting on viewpoints with large errors. This invention innovatively introduces reconstruction error (i.e., reconstruction loss value) as a feedback signal to dynamically adjust the selection probability of each viewpoint in real time. This allows the model to automatically concentrate computational resources on the viewpoint with the worst reconstruction effect, thereby maximizing training efficiency even under conditions of extremely scarce data.

[0046] (2) A self-supervised geometric constraint framework based on the Dropout strategy is constructed. The core challenge of sparse viewpoint reconstruction is the lack of sufficient geometric constraints. Existing methods usually rely on pseudo-true values ​​provided by external sources (such as monocular depth estimation), which introduces additional errors. At the same time, although some algorithms have begun to adopt the Dropout strategy for Gaussian units, they lack comparison with complete models. This invention takes a different approach, using the Dropout strategy as an active geometric perturbation method. Through its own consistency regularization, it forces the model to output consistent rendering results in both the "complete state" and the "incomplete state," thus constructing a self-supervised consistency constraint. This method does not require any external prior data and can effectively suppress geometric noise and floating artifacts by utilizing only the internal structural changes of the model itself, thereby improving the robustness of the model. Attached Figure Description

[0047] Figure 1 This is the overall flowchart of the algorithm of the present invention.

[0048] Figure 2 The overall architecture diagram of the strategy adopted from the perspective of this invention is shown.

[0049] Figure 3 This is a schematic diagram illustrating the self-consistency loss of the present invention. Detailed Implementation

[0050] The present invention will be further illustrated below with reference to the accompanying drawings and specific embodiments. It should be understood that the following specific embodiments are for illustrative purposes only and are not intended to limit the scope of the present invention.

[0051] like Figure 1 As shown, the sparse viewpoint 3D Gaussian sputtering reconstruction method based on adaptive viewpoint sampling of the present invention includes the following steps:

[0052] Step 1: Constructing a 3DGS Reconstruction Dataset

[0053] For a specific scenario, construct a set of training viewpoint images. Test viewpoint image set Initial point cloud data obtained by SfM algorithm and camera pose data set A multi-view dataset containing pose data for each camera. Both are related to image sets or The image data corresponds one-to-one;

[0054] Step 2: Initialize training data

[0055] Initialize the viewpoint data from step 1, resize and linearly normalize the viewpoint image, and convert the initial point cloud data... Convert to initial 3DGS model Assign a position to each of the points. Covariance matrix (Shape and Rotation), Spherical Harmonics (Color), Opacity The properties include position, which directly uses the initial point cloud position, and then calculates the nearest neighbor for each point. The average distance between the neighbors is taken as the logarithm of the initial scaling, and the initial spherical harmonic coefficient order is 0, that is, the color of the point cloud is directly used as the initial color.

[0056] Step 3: Adaptive viewpoint sampling based on reconstruction error distribution

[0057] Based on the number and number of viewpoints obtained, a reconstruction error value is initialized for each viewpoint (this metric is used to quantify the degree of difference between the model-rendered image and the real image at the current viewpoint; in this invention, the viewpoint's loss value is used instead), and recorded in a vector according to the number order. In the initialization of the temperature coefficient of the softmax function Initialize the smoothing factor for reconstruction error update ;

[0058] Initialized reconstruction error vectors for each viewpoint As shown in Equation 1, it is based on the viewpoint set. The unique identifiers of each viewpoint are sorted, meaning the reconstruction error stored therein is the cumulative loss value within the current iteration step for each viewpoint:

[0059] (1);

[0060] in Indicates the first The cumulative loss value within the current iteration step for each viewpoint. During training, the algorithm tracks the difficulty of each viewpoint in real time;

[0061] The current step vector The reconstruction error and temperature coefficient stored in the database are fed into the softmax activation function to obtain the probability distribution of the reconstruction error for each viewpoint. Viewpoints are obtained by randomly sampling viewpoints based on their probability distribution. (This includes the training image for the current viewpoint and the corresponding camera parameters);

[0062] When selecting training viewpoints, such as Figure 2 As shown in Equation 2, this invention employs a probabilistic sampling strategy based on the difficulty of reconstructing training viewpoints. Specifically, at the beginning of each training iteration, as shown in Equation 2, the loss vectors of each viewpoint are used... Calculate the sampling probability distribution, the first... The probability of a viewpoint being selected The calculation method is as follows:

[0063] (2);

[0064] in, This is the temperature coefficient of the softmax function, used to control the magnitude of the difference in probability as a function of reconstruction error. In experiments, it is typically set to a certain value. This ensures that viewpoints with higher reconstruction errors (i.e., viewpoints that are more difficult to reconstruct and have worse reconstruction results) have a higher probability of being selected for training.

[0065] In step 3, besides using the probability distribution normalized by Softmax for random sampling, other weighting strategies can also be employed. For example, the Top-K strategy directly selects the top K viewpoints with the largest loss values ​​and combines them with M random viewpoints from the remaining viewpoints for training. Alternatively, a threshold-triggered strategy can be used, setting a dynamic threshold that increases the sampling frequency of a viewpoint only when its loss value exceeds the threshold; otherwise, uniform sampling is maintained.

[0066] Step 4: Randomly discard Gaussian elements

[0067] Calculate the random drop rate for Gaussian elements based on the current number of iterations. Then, the Gaussian primitives in the model are randomly discarded to obtain the Gaussian model after random primitive discarding. ;

[0068] Step 5: Dual-path rasterization rendering

[0069] The 3DGS model obtained in step 4, after being randomly discarded. Combined with the viewpoint data obtained in step 3 The process involves projecting 3D Gaussian primitives within the view frustum into 2D sputterings; sorting the sputterings based on depth information; and performing alpha blending on each pixel to accumulate color and opacity (following the original 3DGS rasterization renderer workflow) to generate the rendered model of the current viewpoint after random dropout operations. and the corresponding depth map Then, the 3DGS models that were not randomly discarded before step 4 are... Combining the viewpoint data obtained in step 3, the 3DGS model is again rendered using a differentiable rasterizer. Perform rendering operations to generate a complete rendered image of the current viewpoint. and the corresponding depth map ;

[0070] Step 6: Loss Calculation and Self-Consistency Regularization

[0071] Discard the rendered image obtained in step 5. With training viewpoint images Compare and calculate the photometric loss at the current viewpoint. (Including L1 pixel distance loss and D-SSIM structural similarity loss); the complete rendered image obtained in step 5 With training viewpoint images Compare and calculate the photometric loss of the fully rendered image. ;make and Subtraction yields the self-consistency loss. We use this as a regularization term and obtain the total training loss through weighted summation. .

[0072] This invention, based on the Dropout algorithm, uses a self-consistency loss regularization method to guide model training. A diagram of the self-consistency loss is shown below. Figure 3 As shown, its significance lies in improving the robustness and stability of the 3DGS model itself, ensuring that the model reconstruction result conforms to the real geometric structure.

[0073] Through the dual-path rasterization rendering process in step 5, the rendering results of the model after random discarding are obtained respectively. Rendering results of the model without random dropout The two images are compared with the training viewpoint image. Calculate the photometric loss. This includes the total loss during the training process. The calculation is shown in Equation 3:

[0074] (3);

[0075] in, is the proportionality coefficient of the regularization term. express and Photometric loss, which constitutes the main part of the total loss. The calculation method is shown in Equation 4:

[0076] (4);

[0077] in, represent The loss ratio coefficient. express and The loss of luminosity between them The calculation method is shown in Equation 5:

[0078] (5);

[0079] This is a model consistency regularization term, representing the degree of difference between the model and the original model after a slight perturbation caused by random dropout operations. To ensure the geometric consistency of the model, a depth map is added during rendering. and The L1 loss is used as a geometric regularization term. The calculation method is shown in Equation 6:

[0080] (6);

[0081] in, The weighting coefficients represent the depth L1 error. The higher the value, the higher the instability of the model itself.

[0082] In step 6, the calculation of self-consistency loss is not limited to differences at the RGB image level. For example, a depth map or an opacity accumulation map can be rendered simultaneously to calculate the consistency loss of the complete model and the discarded model in terms of geometric output, thereby enhancing the stability of the geometric structure.

[0083] Step 7: Gradient Calculation and Parameter Update

[0084] Based on the total training loss obtained in step 6 Perform backpropagation, calculate the gradient of the current iteration, and update the Gaussian model. Various attribute parameters; smoothing factor obtained in step 3 sum vector The past loss values ​​of the current viewpoint are recorded in the data, and the reconstruction error of the current viewpoint is smoothly updated using an exponential moving average (EMA) strategy.

[0085] The viewpoint loss update strategy employed is not direct overlay, but rather an Exponential Moving Average (EMA) strategy. As shown in Equation 7, the EMA strategy can avoid drastic changes in viewpoint sampling probability caused by excessively large or small single loss values. Specifically, it can be expressed as follows:

[0086] (7);

[0087] in, The smoothing coefficient set during the initialization phase.

[0088] Step 8: Adaptive density control and model parameter update

[0089] Within a preset iteration interval (100 intervals were used in the experiment), the cumulative gradient values ​​of each Gaussian element in the view space were calculated. If the current iteration number reaches the iteration number required for adaptive density control, the 3DGS model is adjusted based on the Gaussian element position gradients calculated in the current iteration. The Gaussian primitives are used for cloning, segmentation, and opacity reset operations. For gradients exceeding a set threshold... For Gaussian elements, if their scale is small, they are cloned; if their scale is large, they are segmented; if the opacity of the Gaussian element is less than a threshold... If the opacity is zero, then return to step 3 and continue iterating until the maximum number of iterations is reached.

[0090] This invention maintains the randomness of viewpoint selection while increasing the training iterations for viewpoints with large reconstruction errors, and reduces the risk of overfitting for viewpoints with small reconstruction errors, thus rationally allocating training resources. Experiments show that this improved reconstruction result improves PSNR by nearly 0.5 dB and SSIM by nearly 0.02 compared to the original algorithm with randomly selected viewpoints in synthesizing new viewpoints.

[0091] This invention utilizes the Dropout strategy to improve the synthesis results while introducing its own consistency regularization to measure the robustness of the model itself, guiding the model to further improve the stability of the reconstruction results and reducing the impact of noise caused by adaptive density control.

[0092] It should be noted that the above content merely illustrates the technical concept of the present invention and should not be construed as limiting the scope of protection of the present invention. For those skilled in the art, various improvements and modifications can be made without departing from the principle of the present invention, and all such improvements and modifications fall within the scope of protection of the claims of the present invention.

Claims

1. A sparse viewpoint 3D Gaussian sputtering reconstruction method based on adaptive viewpoint sampling, characterized in that: Includes the following steps: Step 1: Constructing a 3DGS Reconstruction Dataset For a specific scenario, construct a set of training viewpoint images. Test viewpoint image set Initial point cloud data obtained by SfM algorithm and camera pose data set A multi-view dataset containing pose data for each camera. Both are related to image sets or The image data corresponds one-to-one; Step 2: Initialize training data Initialize the viewpoint data from step 1, resize and linearly normalize the viewpoint image, and convert the initial point cloud data... Convert to initial 3DGS model Assign a position to each of the points. Covariance matrix spherical harmonic coefficient Opacity The attributes, where position is directly taken from the initial point cloud, and then the nearest neighbor of each point is calculated. The average distance of the neighbors is taken as the logarithm of the initial scaling, and the initial spherical harmonic coefficient order is 0. Step 3: Adaptive viewpoint sampling based on reconstruction error distribution Based on the number and number of viewpoints obtained, a reconstruction error value is initialized for each viewpoint and recorded in a vector according to the number order. In the initialization of the temperature coefficient of the softmax function Initialize the smoothing factor for reconstruction error update , the current step vector The reconstruction error and temperature coefficient stored in the database are fed into the softmax activation function to obtain the probability distribution of the reconstruction error for each viewpoint. Viewpoints are obtained by randomly sampling viewpoints based on their probability distribution. ; Step 4: Randomly discard Gaussian elements Calculate the random drop rate for Gaussian elements based on the current number of iterations. Then, the Gaussian primitives in the model are randomly discarded to obtain the Gaussian model after random primitive discarding. ; Step 5: Dual-path rasterization rendering The 3DGS model obtained in step 4, after being randomly discarded. Combined with the viewpoint data obtained in step 3 The 3D Gaussian elements within the view frustum are projected as 2D sputterings. The sputterings are sorted according to depth information, and alpha blending is performed on each pixel to accumulate color and opacity, generating the rendering result of the model from the current viewpoint after random dropout operations. and the corresponding depth map Then, the 3DGS models that were not randomly discarded before step 4 are... Combining the viewpoint data obtained in step 3, the 3DGS model is again rendered using a differentiable rasterizer. Perform rendering operations to generate a complete rendered image of the current viewpoint. and the corresponding depth map ; Step 6: Loss Calculation and Self-Consistency Regularization Discard the rendered image obtained in step 5. With training viewpoint images Compare and calculate the photometric loss at the current viewpoint. The complete rendered image obtained in step 5 With training viewpoint images Compare and calculate the photometric loss of the fully rendered image. ,make and Subtraction yields the self-consistency loss. We use this as a regularization term and obtain the total training loss through weighted summation. ; Step 7: Gradient Calculation and Parameter Update Based on the total training loss obtained in step 6 Perform backpropagation, calculate the gradient of the current iteration, and update the Gaussian model. Each attribute parameter is based on the smoothing factor obtained in step 3. sum vector The past loss values ​​of the current viewpoint are recorded in the data, and the EMA strategy is used to smoothly update the reconstruction error of the current viewpoint. Step 8: Adaptive density control and model parameter update Within a preset iteration interval, the cumulative gradient values ​​of each Gaussian element in the view space are calculated. If the current iteration number reaches the iteration number required for adaptive density control, the 3DGS model is adjusted based on the Gaussian element position gradients calculated in the current iteration. The Gaussian unit is cloned, segmented, and its opacity reset is performed. For gradients exceeding a set threshold... For Gaussian elements, if their scale is small, they are cloned; if their scale is large, they are segmented. If the opacity of the Gaussian element is less than a threshold... If the opacity is set to 0, then return to step 3 to continue iterating until the maximum number of iterations is reached.

2. The sparse viewpoint 3D Gaussian sputtering reconstruction method based on adaptive viewpoint sampling according to claim 1, characterized in that: The initialization of each viewpoint reconstruction error vector in step 3 As shown in equation (1), it is based on the viewpoint set. The unique identifiers of each viewpoint are sorted, meaning the reconstruction error stored therein is the cumulative loss value within the current iteration step for each viewpoint: （1）； in, Indicates the first The cumulative loss value within the current iteration step for each viewpoint.

3. The sparse viewpoint 3D Gaussian sputtering reconstruction method based on adaptive viewpoint sampling according to claim 1, characterized in that: In step 3, when selecting training viewpoints, a probabilistic sampling strategy based on the reconstruction error of the training viewpoints is adopted. That is, at the beginning of each training iteration, as shown in equation (2), the loss vector of each viewpoint is used. Calculate the sampling probability distribution, the first... The probability of a viewpoint being selected The calculation method is as follows: （2）； in, Indicates the first The cumulative loss value within the current iteration step for each viewpoint, and , This is the temperature coefficient of the softmax function.

4. The sparse viewpoint 3D Gaussian sputtering reconstruction method based on adaptive viewpoint sampling according to claim 1, characterized in that: In step 6, based on the Dropout algorithm, a self-consistency loss regularization method is used to guide the training of the model; Through the dual-path rasterization rendering process in step 5, the rendering results of the model after random discarding are obtained respectively. Rendering results of the model without random dropout The two images are compared with the training viewpoint image. Calculate the photometric loss, where the total loss during the training process is... The calculation is shown in equation (3): （3）； in, The proportionality coefficient of the regularization term. express and Light loss, The calculation method is shown in equation (4): （4）； in, represent The loss ratio coefficient express and The loss of luminosity between them The calculation method is shown in equation (5): （5）； This is a model consistency regularization term, representing the degree of difference between the model and the original model after a slight perturbation caused by random dropout operations. To ensure the geometric consistency of the model, a depth map is added during rendering. and The L1 loss is used as a geometric regularization term. The calculation method is shown in equation (6): （6）； in, The weighting coefficients represent the depth L1 error.

5. The sparse viewpoint 3D Gaussian sputtering reconstruction method based on adaptive viewpoint sampling according to claim 1, characterized in that: The viewpoint loss update strategy used in step 7 adopts the EMA strategy, as shown in equation (7), specifically expressed in the following form: （7）； in, The smoothing coefficient set during the initialization phase.

Images