A confidence distillation method fusing three-dimensional gaussian splatting and signed distance field

Abstract

The application discloses a confidence distillation method fusing three-dimensional Gaussian splatting and signed distance field. The existing representation based on discrete Gaussian primitives is prone to misjudging appearance changes as geometric undulations when dealing with strong view-dependent effects such as mirror highlights, resulting in a decline in surface reconstruction quality. The application first constructs a double-branch confidence reconstruction module containing a Gaussian branch and a distance field branch. The Gaussian branch gives each Gaussian primitive a learnable saliency attribute to render a confidence map. The distance field branch predicts pixel confidence according to appearance error and geometric uncertainty through a lightweight confidence network. Secondly, the application adopts a two-stage confidence distillation mechanism. In the early stage, a pre-trained normal map and a Gaussian branch confidence map are used to modulate the supervision mask to ensure stable training. After a specified number of iterations, the application switches to the normal map rendered by the distance field branch, and combines the confidence maps of the two branches to jointly modulate the mask, distilling the anti-interference geometric reconstruction advantage of the distance field branch to the Gaussian branch. The method makes three-dimensional Gaussian splatting integrate the advantages of itself, an offline convolutional neural network and a signed distance field, and relies on the efficient rendering speed of three-dimensional Gaussian splatting to realize high-fidelity and high-efficiency scene appearance and geometric joint reconstruction.

Description

Technical Field

[0001] This invention relates to the fields of computer vision and computer graphics, specifically to a method and apparatus for three-dimensional scene reconstruction that integrates three-dimensional Gaussian splashing and symbolic distance field, applicable to high-quality reconstruction of objects and open scenes. Background Technology

[0002] With the rapid development of the digital economy, the demand for 3D digital content in cutting-edge emerging industries such as digital twins, virtual reality, and augmented reality is growing. High-precision 3D scene reconstruction technology serves as a fundamental support for these applications, and breakthroughs in its independent and controllable technology are of great significance for promoting the development of my country's next-generation information technology industry.

[0003] In existing 3D scene reconstruction schemes, explicit representation methods based on 3D Gaussian splashing have attracted widespread attention due to their efficient rendering performance and clear scene representation. This method represents the scene as anisotropic Gaussian primitives and achieves real-time rendering through a differentiable rasterization pipeline, achieving significant progress in reconstruction quality and efficiency in recent years. However, such discrete representation methods have inherent drawbacks when handling complex scenes: because each Gaussian primitive is independently optimized and lacks global geometric constraints, strong viewpoint-dependent effects such as specular highlights in appearance reconstruction are easily misjudged as geometric undulations, leading to a large number of artifacts and noise introduced into surface reconstruction. Traditional methods rely solely on pre-trained offline normal vector estimators for geometric supervision, but such supervision signals themselves are biased and cannot perceive scene characteristics, making it difficult to effectively distinguish appearance changes from true geometric structures.

[0004] Furthermore, while existing methods combining implicit symbolic distance field representation with 3D Gaussian splashing theoretically possess the dual advantages of implicit geometric resistance to interference and high efficiency of explicit rendering, they lack an effective selective control mechanism during geometric prior transfer. Specifically, when the symbolic distance branch transfers geometric supervision to the Gaussian branch, a uniform processing strategy is applied to all pixel regions, failing to consider the differences in sensitivity between appearance and geometric supervision across different regions. This globally uniform distillation method forces appearance reconstruction-dominated regions to accept inappropriate geometric constraints, limiting the expressive power of appearance reconstruction and weakening the anti-interference advantage that the symbolic distance field branch should exert.

[0005] In summary, existing technologies have not yet established an effective geometric distillation mechanism for the fusion of Gaussian splashing and symbolic distance fields. This makes it difficult to fully utilize the anti-interference geometric reconstruction advantages of the symbolic distance field branch while maintaining appearance reconstruction quality, resulting in insufficient geometric reconstruction accuracy under strong viewpoint correlation effects such as specular highlights. Therefore, a confidence distillation scheme capable of sensing reconstruction quality differences and adaptively modulating geometric supervision transfer is urgently needed. Summary of the Invention

[0006] The technical solution adopted in this invention is: a confidence distillation method integrating three-dimensional Gaussian splashing and symbolic distance field, comprising the following steps:

[0007] S1. Construct a dual-branch scene representation, including a Gaussian branch based on 3D Gaussian splashing and an implicit branch based on the symbolic distance field;

[0008] S2. Render the initial geometric information of the scene normal map and depth map through the Gaussian branch, and render the implicit geometric information of the scene through the implicit branch, while generating the confidence map of the Gaussian branch and the confidence map of the implicit branch.

[0009] S3. Employing a two-stage confidence modulation mechanism, the geometric prior of the implicit branch is distilled into the Gaussian branch:

[0010] S31. First stage: Use the prior normal vector map provided by the pre-trained normal vector estimator as geometric supervision, and use the confidence map generated by the Gaussian branch to modulate the supervision mask to train the Gaussian branch.

[0011] S32. In the second stage, after the preset conditions are met, the normal vector map rendered by the implicit branch is switched to be used as geometric supervision, and the supervision mask is modulated by the confidence map of the Gaussian branch and the implicit branch together to continue training the Gaussian branch.

[0012] S4. Obtain the normal vector map output by the Gaussian branch after the implicit branch geometric prior distillation.

[0013] Furthermore, the generation of the confidence map in step S2 includes:

[0014] S21. Gaussian branch confidence map generation: Each Gaussian primitive in the Gaussian branch is assigned a learnable saliency attribute, the saliency attribute is incorporated into the differentiable rasterization pipeline, and then the confidence map of the Gaussian branch is obtained through Alpha blending rendering; wherein, the higher the pixel value in the confidence map, the lower the confidence of the reconstruction effect of the Gaussian branch at that pixel position.

[0015] The formula for rendering the Gaussian branch confidence plot is:

[0016]

[0017] in For the first The saliency property of each Gaussian element, For the first step in the rendering process The confidence plot of the Gaussian branch can be obtained using the above formula, which uses a Gaussian alpha mixture weight.

[0018] S22. Implicit Branch Confidence Map Generation: A confidence network is set up, which shares the same input single parameter as the radiosity network in the implicit branch and is independent; for each pixel ray, the confidence network predicts the confidence value of each sampling point on it; the confidence value is used for volume rendering along the ray to obtain the confidence map of the implicit branch.

[0019] The volume rendering formula for the implicit branch confidence map is:

[0020]

[0021] in For the first Confidence value of each sampling point and This involves setting the alpha value and sampling interval during the volumetric rendering process. Through the above steps, confidence information is seamlessly embedded into the rendering process of the Gaussian branch and the implicit branch of the signed distance field, enabling real-time optimization while participating in the rendering process. Compared to existing techniques that predict results offline, the confidence results obtained through the above steps... and It produces smoother and more fine-grained optimization results.

[0022] Furthermore, in step S21, the saliency attribute is optimized using the following loss function: a binary target map is constructed, which identifies pixel regions where the radiance loss gradient magnitude is less than a predetermined first threshold and the geometric loss gradient magnitude is greater than a predetermined second threshold; the binary target is used as a supervision signal to optimize the confidence map.

[0023] The optimization loss function for the saliency attribute is:

[0024]

[0025]

[0026] in For the aforementioned binary supervision target, and Let represent the gradients of the radiosity loss with respect to the mean of the Gaussian elements and the gradients of the geometric loss with respect to the rotational properties of the Gaussian elements, respectively. and The first and second thresholds are predetermined.

[0027] Furthermore, in step S22, the confidence network is optimized using the following loss function: an appearance error term is constructed based on the difference between the implicit branch rendered color and the true color; a geometric uncertainty term is constructed based on the Shannon entropy calculated according to the Alpha weight distribution on the ray; and the appearance error term and the geometric uncertainty term are used as joint supervision signals to optimize the confidence map of the implicit branch.

[0028] The definition of the appearance error term is:

[0029]

[0030] in These are the pixel values ​​sampled from the truth image used in the implicit branch. Normalized image values ​​are used consistently throughout the process, therefore... .

[0031] The geometric uncertainty term is defined as follows:

[0032]

[0033] in This represents the normalized Alpha weight value for that ray.

[0034] The optimized loss function of the confidence network is:

[0035]

[0036] in Weights are predefined for the loss function. Through the above steps, the confidence attributes in the Gaussian branch and the implicit branch of the signed distance field are finely optimized and accurately targeted to the target pixel region. This refers to regions in the Gaussian rendering results that have low appearance loss but high geometric loss, i.e., areas that require additional geometric supervision. The region with low appearance loss and low geometric Shannon entropy is identified as the region that can provide high-precision geometric supervision. Compared with existing technologies, the above setting establishes an independent and reliable optimization path for the confidence attribute, ensuring that Gaussian units and implicit branches can adaptively optimize the confidence attribute during training and incorporate it into the scene reconstruction process to improve its reconstruction accuracy.

[0037] Furthermore, the predetermined condition in step S3 is that the number of training iterations reaches a predetermined threshold; in the first stage S31, the confidence map of the Gaussian branch and the truth mask are multiplied element-wise to obtain the modulated supervision mask; in the second stage S32, the confidence map of the Gaussian branch, the confidence map of the implicit branch of the symbolic distance field, and the truth mask are multiplied element-wise to obtain the modulated joint supervision mask.

[0038] The formula for the two-stage supervision switching strategy is expressed as follows:

[0039]

[0040] in and This refers to the normal vector supervision and mask used in the confidence supervision process described above.

[0041] The distilled Gaussian branch combines the advantages of efficient rendering, stable initial supervision by offline convolutional neural networks, and accurate geometric reconstruction by the symbolic distance field. While maintaining the quality of appearance reconstruction, it significantly improves the accuracy of surface geometry, achieving high-fidelity and high-efficiency joint reconstruction of scene appearance and geometry.

[0042] Compared with the prior art, the advantages of the present invention are as follows:

[0043] 1. The confidence mechanism proposed in this scheme achieves adaptive and selective supervision of geometric supervision signals. Existing methods mostly use uniform supervision weights to supervise geometric results, which do not consider the sensitivity of different pixel regions during optimization and are prone to introducing supervision noise. This invention introduces independent confidence rendering and supervision mechanisms for the Gaussian branch and the symbolic distance field branch, respectively. By using the rendered confidence map and attributes, the Gaussian branch focuses on pixel regions with unequal reconstruction quality, such as those with reasonable appearance reconstruction but failed geometric reconstruction, thereby improving the reconstruction quality of the scene.

[0044] 2. The two-stage distillation strategy balances training stability and reconstruction accuracy. This invention employs different distillation strategies in the early and late stages of training to address different optimization needs. In the early stages, the geometric rendering results of the Gaussian branch and the implicit branch have not yet converged, leading to erroneous supervision of the Gaussian branch due to geometric bias, thus causing training instability. Therefore, in the first stage, the prior normal map used in this invention is generated by an offline convolutional neural network pre-trained on a large-scale dataset, exhibiting good generalization and stability. In the later stages of training, the focus of optimization shifts to fine-tuning the rendering results. Therefore, the normal map used in the second stage originates from the rendering results of the implicit branch. In this stage, the reconstructed normal maps of both branches are superior to the prior normal maps used in the first stage. Therefore, it is only necessary to distill the implicit branch, which is superior to the Gaussian branch, into the Gaussian branch using our proposed confidence mechanism.

[0045] 3. The confidence mechanism proposed in this solution has the advantage of integrating geometric supervision from multiple sources. This invention, through its proposed confidence distillation method, ultimately integrates the advantages of three sources of geometric supervision: offline convolutional neural network prediction normal vector results, implicit branch continuous symbolic distance field rendering normal vector results, and Gaussian branch rendering normal vector results. The confidence mechanism achieves complementary advantages among these three sources. Experimental results demonstrate that, in scenes with complex lighting conditions, this invention significantly improves appearance and geometry reconstruction and rendering.

[0046] 4. The confidence distillation mechanism proposed in this scheme does not significantly increase the computational overhead of the inference stage. The confidence network, saliency attribute, and signed distance field branch used in this invention only participate in loss calculation and gradient backpropagation during the training process; only the Gaussian branch is used in the inference rendering stage. After the training stage is completed, the Gaussian branch has internalized the geometric prior from the implicit branch, and its rendering pipeline in the inference stage is completely consistent with standard 3D Gaussian splashing. Furthermore, the saliency attribute does not need further optimization; therefore, only the following steps are required:

[0047] 5. High-fidelity rendering and high-precision reconstruction results can be obtained with a single forward rasterization, maintaining the core advantages of real-time 3D Gaussian splash rendering.

[0048] Table 1 below presents a comparison of experimental results between our proposed method and existing technologies on two public benchmark datasets, BlenderMVS and DTU. Existing technologies include SuGaR, GaussianSurfels, GSDF, and ProbeSDF. The experimental comparison metrics include appearance rendering quality metrics such as Peak Signal-to-Noise Ratio (PSNR) and Structural Similarity (SSIM), and geometric reconstruction metrics such as Chamfer Distance (CD).

[0049] The BlenderMVS dataset focuses on multi-view outdoor and real-world scene images, while the DTU dataset is experimental data captured in a laboratory under complex lighting conditions. Although they have different focuses, they both place high demands on the fitting ability and robustness of the proposed method. The confidence distillation method that integrates three-dimensional Gaussian splash and symbolic distance field proposed in this scheme fully demonstrates its effectiveness on the above two datasets.

[0050] Table 1 compares the experimental results of the confidence distillation method integrating 3D Gaussian splash and symbolic distance field with existing methods on the BlenderMVS and DTU datasets.

[0051] Attached Figure Description

[0052] Figure 1This is a schematic diagram of the overall process of a confidence distillation method that integrates three-dimensional Gaussian splashing and symbolic distance field according to the present invention;

[0053] Figure 2 This is a detailed flowchart of the Gaussian branch rendering and supervised confidence graph described in this invention;

[0054] Figure 3 This is a detailed flowchart of the symbolic distance field implicit branch rendering and supervised confidence graph described in this invention. Detailed Implementation

[0055] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. The present invention is a confidence distillation method that integrates three-dimensional Gaussian splashing and symbolic distance field. The detailed process is as follows:

[0056] This embodiment provides a confidence distillation method that integrates three-dimensional Gaussian splashing and symbolic distance field, the overall process of which is as follows: Figure 1 As shown, this method first constructs a bi-branch scene representation (S1) containing a Gaussian branch and a symbolic distance field implicit branch. The initial geometric information and implicit geometric information of the scene are rendered through the two branches respectively, and their respective confidence maps are generated (S2). Based on this, a two-stage confidence modulation mechanism is used to distill the geometric prior of the implicit branch to the Gaussian branch (S3), and finally the normal vector map output by the distilled Gaussian branch is obtained (S4).

[0057] Specifically, in step S1, the bi-branch scene representation consists of the following two parts: the Gaussian branch explicitly represents the scene using anisotropic Gaussian elements, with each Gaussian element having its center position. Rotation Scaling ,transparency Normal vector and direction-related color sphere harmonic parameters Attribute composition; where rotation Quaternion representation is used to ensure its validity; normal vector Determined by the third direction of the rotation matrix, therefore scaling The third axis of the attribute is forced to be set to 0 to ensure the normal vector is reasonable.

[0058] The implicit branch performs volume rendering based on the symbolic distance field, modeling the symbolic distance function values ​​of the scene using a multilayer perceptron. Both branches share the input multi-view image but perform forward rendering independently.

[0059] In step S2, while rendering the scene geometry information through Gaussian branch and implicit branch respectively, confidence maps are generated for each. The generation and supervision process for the Gaussian branch confidence map is as follows: Figure 2As shown, the generation and supervision process of the implicit branch confidence graph is as follows: Figure 3 As shown.

[0060] Combination Figure 2 The generation of the Gaussian branch confidence plot includes the following process:

[0061] First, each Gaussian basis in the Gaussian branch is assigned a learnable saliency property. The attribute value is constrained by the Sigmoid function. The interval is used to characterize the reliability of each Gaussian branch during training. The saliency attribute is incorporated into a differentiable rasterization pipeline, and the confidence map of the Gaussian branch is obtained through alpha blending rendering. The rendering formula is as follows:

[0062]

[0063] in, For the first The saliency property of each Gaussian element, For the first step in the rendering process The Gaussian branch is used to mix alpha weights. In the confidence map, the higher the pixel value, the lower the confidence of the reconstruction effect of the Gaussian branch at that pixel location, meaning the Gaussian branch is less reliable at that location.

[0064] To optimize the saliency attribute, a binary target graph is constructed. The target image indicates that the magnitude of the radiance loss gradient is less than a predetermined first threshold. Furthermore, the geometric loss gradient is greater than the predetermined second threshold. The pixel region. The specific optimization loss function is:

[0065]

[0066]

[0067] in, and These represent the gradients of the radiosity loss with respect to the Gaussian cell position and the geometric loss with respect to the Gaussian cell rotation attribute, respectively. The meaning of this target map is that when the appearance reconstruction error of a certain pixel region is small, but the geometric reconstruction error is still large, this region is marked as requiring special attention, and the saliency attribute of the Gaussian cell is trained to output a higher response in such regions. Through the above supervision, the confidence map of the Gaussian branch can automatically identify regions where appearance supervision and geometric supervision conflict.

[0068] Combination Figure 3 The generation of the implicit branch confidence graph includes the following process:

[0069] Set up a lightweight confidence network that shares inputs with the radiosity network in the implicit branch but has independent parameters; the number of parameters is approximately half that of the radiosity network. For each pixel ray... The confidence network predicts the confidence value of each sampling point. Perform volume rendering along the ray to obtain the confidence map of the implicit branch. :

[0070]

[0071] in, For the first Confidence value of each sampling point and These represent the alpha value and sampling interval during the volume rendering process, respectively.

[0072] The confidence network optimization employs a joint supervision strategy. First, it renders colors based on implicit branches. With true color The difference is used to construct an appearance error term. :

[0073]

[0074] in, The pixel values ​​sampled from the truth image used for the implicit branch are uniformly normalized values. The smaller the appearance error, The closer the value is to 1, the higher the confidence level of the appearance reconstruction of that pixel.

[0075] Secondly, Shannon entropy is calculated based on the Alpha weight distribution on the ray, and a geometric uncertainty term is constructed. :

[0076]

[0077] in This represents the normalized Alpha weight value along that ray. The higher the Shannon entropy, the more dispersed the geometric distribution and the higher the uncertainty along the ray. The closer it is to 0.

[0078] Using appearance error and geometric uncertainty as joint supervision signals, the optimal loss function of the confidence network is:

[0079]

[0080] in Predetermine weight values ​​for the loss function. Through the above joint supervision, the confidence plot of the implicit branch can simultaneously reflect the uncertainty of both the appearance reconstruction quality and the geometric distribution.

[0081] In step S3, a two-stage confidence modulation mechanism is used for geometric prior distillation, which is implemented through the following piecewise function:

[0082]

[0083] in, and These are the normal vector supervision and mask used in the confidence supervision process, respectively. The prior normal vector graph provided for the pre-trained normal vector estimator The normal map rendered for implicit branches. For the truth mask, This represents element-wise multiplication. This is the preset threshold for the number of iterations.

[0084] Specifically, in the first stage, when the number of training iterations has not reached the threshold... At that time, the prior normal vector graph provided by the pre-trained normal vector estimator is used. As a geometric supervision, and utilizing the confidence graph of Gaussian branches With truth mask Perform element-wise multiplication to obtain the modulated supervision mask. In this stage, the pre-trained normal vector map provides stable initial geometric supervision, and the confidence map of the Gaussian branch modulates the supervision mask, suppressing the geometric constraint strength in the conflict region between appearance supervision and geometric supervision, thus ensuring stable convergence in the early stage of training.

[0085] In the second phase, the number of training iterations reaches or exceeds the threshold. Then, switch to the normal map rendered with implicit branching. As a geometric supervision, and utilizing Gaussian branch confidence plots With implicit branch confidence plot Common and truth mask Element-wise multiplication is performed to obtain the modulated joint supervision mask. In this stage, the normal vector map rendered by the implicit branch has higher accuracy and can perceive scene characteristics compared to the pre-trained normal vector map. The combination of the two branch confidence maps provides more refined selective control: when a certain pixel region has low confidence in both the Gaussian branch and the implicit branch, the geometric supervision of that region is effectively suppressed, avoiding inappropriate geometric constraints from interfering with appearance reconstruction.

[0086] In step S4, after the above two-stage distillation training, the normal vector diagram output by the Gaussian branch is... It combines the advantages of pre-trained offline convolutional neural networks providing stable initial supervision with the advantages of implicit branching of the symbolic distance field for interference-resistant geometric reconstruction, while retaining the efficient rendering characteristics of 3D Gaussian splashing itself.

[0087] This normal vector map can be directly used for subsequent scene geometry representation, or it can be used as a geometry regularization supervision signal for downstream tasks.

[0088] It should be noted that the above embodiments are not intended to limit the scope of protection of the present invention. Equivalent transformations or substitutions made based on the above technical solutions all fall within the scope of protection of the claims of the present invention.

Claims

1. A confidence distillation method integrating three-dimensional Gaussian distance fields, characterized in that, Includes the following steps: S1. Construct a dual-branch scene representation, including a Gaussian branch based on 3D Gaussian splashing and an implicit branch based on the conforming distance field; S2. Render the initial geometry of the scene using Gaussian branch, and render the implicit geometry of the scene using implicit branch; S3. A two-stage confidence modulation mechanism is used to distill the geometric prior of the implicit branch into a Gaussian branch: S31. In the first stage, the prior normal map provided by the pre-trained normal vector estimator is used as geometric supervision, and the Gaussian branch is trained by modulating the supervision mask with the confidence map generated by the Gaussian branch. S32. In the second stage, after the preset conditions are met, the normal map rendered by the implicit branch of the symbolic distance field is switched to be used as geometric supervision, and the supervision mask is modulated by the confidence maps of the Gaussian branch and the implicit branch, and the training of the Gaussian branch continues. S4. Obtain the normal map output by the Gaussian branch after implicit branch geometric prior distillation.

2. The method according to claim 1, characterized in that, The confidence plot generated by the Gaussian branch in step S2 is obtained through the following method: S21. Assign a learnable saliency property to each Gaussian element in the Gaussian branch; S22. Incorporate the saliency attribute into the differentiable rasterization pipeline and obtain the confidence map through Alpha blending rendering.

3. The method according to claim 2, characterized in that, In step S21, the significance attribute is optimized using the following loss function: S211. Construct a binary target map, wherein the circular target map identifies pixel regions where the radiance loss gradient magnitude is less than a predetermined first threshold and the geometric loss gradient magnitude is greater than a predetermined second threshold. S212. Using the binary target map as a supervision signal, optimize the confidence map.

4. The method according to claim 1, characterized in that, The confidence plot of the distance field branch in step S2 is obtained in the following way: S23. Set up a confidence network, and make the confidence network and the radiometric network in the implicit branch share the same input single parameter but are independent; S24. For each pixel ray, the confidence network predicts the confidence value of each sampling point on it; S25. Perform volume rendering along the ray to obtain the confidence map of the implicit branch.

5. The method according to claim 4, characterized in that, In step S23, the confidence network is optimized using the following loss function: S231. Construct an appearance error term based on the difference between the implicit branch rendered color and the true color; S232. Calculate the Shannon entropy based on the Alpha weight distribution on the ray and construct the geometric uncertainty term; S233. Using the appearance error term and the geometric uncertainty term as a joint monitoring signal, optimize the confidence map of the implicit branch.

6. The method according to claim 1, characterized in that, The generation of the confidence map in step S2 includes: S21. Gaussian Branch Confidence Map Generation: Each Gaussian primitive in the Gaussian branch is assigned a learnable saliency attribute. This saliency attribute is incorporated into a differentiable rasterization pipeline, and then the confidence map of the Gaussian branch is obtained through alpha blending rendering. The higher the pixel value in the confidence map, the lower the confidence of the reconstruction effect of the Gaussian branch at that pixel location. The formula for rendering a Gaussian branch confidence plot is: in For the first The saliency property of each Gaussian element, For the first step in the rendering process Gaussian Alpha mixture weights, and depth-sorted Gaussian elements preceding the current Gaussian element are used. It indicates that, through the previous The transmittance is obtained by multiplying the Alpha mixed weights of the Gaussian elements to achieve the attenuation of the current Gaussian element weights. The confidence plot of the Gaussian branch is obtained through the above formula. S22. Implicit Branch Confidence Map Generation: A confidence network is set up, which shares the same input single parameter as the radiosity network in the implicit branch and is independent; for each pixel ray, the confidence network predicts the confidence value of each sampling point on it; the confidence value is used for volume rendering along the ray to obtain the confidence map of the implicit branch. The formula for volume rendering of implicit branch confidence plots is: in For the first Confidence value of each sampling point and To determine the alpha value and sampling interval during volumetric rendering, the transmittance of the sampling point during rendering is determined by the previous... Alpha value of each sampling point It is obtained by accumulating exponents.

7. The method according to claim 1, characterized in that, In step S21, the saliency attribute is optimized using the following loss function: a binary target map is constructed, which identifies pixel regions where the radiometric loss gradient magnitude is less than a predetermined first threshold and the geometric loss gradient magnitude is greater than a predetermined second threshold; the binary target map is then used as a supervision signal to optimize the confidence map. The optimization loss function for salient attributes is: in For dual supervision objectives, and Let represent the gradients of the radiosity loss with respect to the mean of the Gaussian elements and the gradients of the geometric loss with respect to the rotational properties of the Gaussian elements, respectively. and The first and second thresholds are predetermined; In step S22, the confidence network is optimized using the following loss function: an appearance error term is constructed based on the difference between the implicit branch's rendered color and the ground truth color; a geometric uncertainty term is constructed by calculating the Shannon entropy based on the alpha weight distribution on the ray; and the appearance error term and the geometric uncertainty term are used as joint supervision signals to optimize the confidence map of the implicit branch. The definition of the appearance error term is: in The pixel values ​​sampled from the truth image used in the implicit branch are normalized image values ​​used throughout the process. , The geometric uncertainty term is defined as follows: in This represents the normalized Alpha weight value on that ray. The optimized loss function of the confidence network is: in Predetermine weight values ​​for the loss function. This represents the L2 norm.

8. The method according to claim 1, characterized in that, In step S3, the predetermined condition is that the number of training iterations reaches a predetermined threshold. In the first stage S31, element-wise multiplication is performed using the confidence map of the Gaussian branch and the truth mask to obtain the modulated supervision mask. In the second stage S32, element-wise multiplication is performed using the confidence map of the Gaussian branch, the confidence map of the implicit branch of the symbolic distance field, and the truth mask to obtain the modulated joint supervision mask. The formula for the two-stage supervision switching strategy is expressed as follows: in and This refers to the normal vector supervision and mask used in the confidence supervision process described above.

9. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the program, it implements a confidence distillation method for fusing three-dimensional Gaussian distance fields as described in any one of claims 1 to 8.

10. A computer-readable storage medium storing computer instructions thereon, characterized in that, When executed by the processor, the computer instructions implement a confidence distillation method for fusing three-dimensional Gaussian distance fields as described in any one of claims 1-8.

Images