360-degree view synthesis method based on panoramic Gaussian sputtering

Abstract

The invention discloses a 360-degree view synthesis method based on panoramic Gaussian sputtering. According to the method, 360-degree panoramic reconstruction and rendering are realized through the synergistic effect of an anchor point-based Gaussian generation strategy, a spherical multi-level training strategy and a Gaussian ellipsoid ERP projection mechanism. Firstly, a motion recovery structure algorithm is utilized to extract a three-dimensional point cloud from a multi-view panoramic image, anchor points are constructed in a voxelization mode, Gaussian parameters of the anchor points are predicted through a lightweight multilayer perceptron, and then through a spherical multi-level training strategy, the anchor points are activated gradually from coarse to fine and a panoramic radiation field is optimized. Fusing local details and global structure information at each level; and finally, generating a continuous and seamless 360-degree panoramic image through Gaussian ellipsoid equidistant columnar projection. The method significantly reduces redundant gauss, improves storage efficiency and rendering quality, ensures that the geometric structure of the scene is more accurate, enhances the spatial consistency of the panoramic image, and is suitable for real-time application scenes such as virtual reality and augmented reality.

Description

Technical Field

[0001] This invention belongs to the field of computer vision and graphics technology, and relates to view compositing technology, specifically to a 360-degree view compositing method based on panoramic Gaussian sputtering. Background Technology

[0002] With the widespread application of 360° cameras and VR / AR devices, 360° video has rapidly developed in multimedia services. Users can obtain immersive virtual experiences through head-mounted displays, but achieving seamless six degrees of freedom (6DoF) interaction requires a large number of high-resolution panoramic images, posing a significant challenge to data storage and transmission. Therefore, synthesizing new 360° views from a limited number of 360° images has become a key technical problem.

[0003] In recent years, Neural Radiation Fields (NeRF) have achieved significant results in novel view synthesis as an implicit 3D representation method. Some studies have extended it to the 3D reconstruction of panoramic images, improving reconstruction quality through methods such as local radiation field modeling, spherical coordinate optimization, and feature voxelization. However, NeRF models rely on large networks, resulting in low training and rendering efficiency, making it difficult to meet the needs of real-time applications.

[0004] 3D Gaussian Splatting (3DGS) technology, based on explicit 3D representation and parallel rendering pipelines, combines high quality and high efficiency, and is gradually becoming an important direction for novel view compositing. However, existing 3DGS methods are mostly oriented towards perspective images, and when directly applied to panoramic images, they suffer from problems such as Gaussian pixel redundancy and insufficient structural representation, affecting the reconstruction quality and resource utilization efficiency of large-scale scenes. Summary of the Invention

[0005] The purpose of this invention is to provide a 360-degree view synthesis method based on panoramic Gaussian sputtering, to solve the problems of Gaussian redundancy, insufficient structural representation, and low efficiency of panoramic scene reconstruction in existing technologies. This invention utilizes an anchor-point-based Gaussian generation strategy to extract anchor points from the point cloud and predict Gaussian properties. A multi-layer spherical training strategy is used to progressively optimize the global structure and local details, thereby improving training stability and reconstruction accuracy. Finally, the 360-degree image is rendered by implementing equirectangular projection (ERP) of the Gaussian ellipsoid.

[0006] The method of this invention includes: an anchor-based Gaussian generation strategy, a spherical multilayer training strategy, and a Gaussian ellipsoid ERP projection mechanism.

[0007] Anchor-based Gaussian generation strategy: This strategy utilizes the Structure-from-Motion (SfM) algorithm to extract 3D point cloud anchor points from the input multi-view panoramic image, and then uses these anchor points as the basis for voxelization of the scene. A lightweight multilayer perceptron (Tiny MLP) is used to predict the Gaussian parameters corresponding to each anchor point, thereby achieving efficient and structurally constrained Gaussian generation and reducing Gaussian redundancy. Gaussian parameters include position, scale, color, and orientation.

[0008] Spherical Multi-Level Training Strategy: To balance the optimization of local details and global structure of panoramic images across different depth layers, this invention proposes a spherical multi-level training strategy. This strategy progressively optimizes the panoramic radiation field from coarse to fine, integrating local and global feature information at each level, thereby effectively improving training stability and reconstruction accuracy.

[0009] Gaussian ellipsoid ERP projection mechanism: Based on the Gaussian ellipsoid generated by the anchor point, the Gaussian ellipsoid is projected onto the ERP image using Equirectangular Projection (ERP) to render a 360-degree panoramic image and achieve high-fidelity panoramic rendering.

[0010] The specific steps of the method of this invention are as follows:

[0011] Step (1) Obtain multi-view panoramic images and generate initial point clouds using the SfM algorithm;

[0012] Step (2) Voxelize the point cloud to generate anchor points and predict Gaussian parameters using a lightweight multilayer perceptron.

[0013] Step (3) Optimize the panoramic radiation field using a spherical multilayer training strategy and optimize the model using a loss function;

[0014] Step (4) Project the Gaussian ellipsoid onto the ERP image using equidistant cylindrical projection and render and output a 360-degree panoramic image.

[0015] Compared with existing technologies, this invention has the following advantages: 1) Reduced data redundancy: By using the Gaussian generation mechanism with anchor point constraints, redundant Gaussians are significantly reduced, improving storage efficiency and rendering quality; 2) Improved structural fidelity: Anchor point initialization and voxel modeling ensure more accurate scene geometry and enhance the spatial consistency of panoramic images; 3) Improved reconstruction quality and stability: The spherical multi-layer training strategy enhances global structural expression while maintaining local details, significantly improving reconstruction results; 4) High-efficiency rendering performance: The explicit representation based on 3D Gaussian sputtering has efficient parallel characteristics, and the training and inference speeds are significantly improved compared to NeRF-like methods, making it suitable for real-time application scenarios such as virtual reality and augmented reality. Attached Figure Description

[0016] Figure 1 This is an overall flowchart of the method of the present invention;

[0017] Figure 2 This is a diagram showing the internal structure of each module in the method of this invention;

[0018] Figure 3 This is a diagram of the internal structure of MLPs in the method of this invention. Detailed Implementation

[0019] A 360° view synthesis method based on panoramic Gaussian sputtering achieves high-quality 360° panoramic reconstruction and rendering through the synergistic effect of an anchor-point-based Gaussian generation strategy, a spherical multi-level training strategy, and a Gaussian ellipsoidal ERP projection mechanism. The method first extracts 3D point clouds from multi-view panoramic images using the Structure-from-Motion (SfM) algorithm and constructs anchor points in a voxelized manner. A lightweight multilayer perceptron (Tiny MLP) is used to predict the Gaussian parameters (position, scale, color, and orientation, etc.) corresponding to the anchor points, achieving efficient Gaussian generation under structural constraints and effectively reducing redundant Gaussians. Subsequently, a spherical multi-level training strategy is introduced to progressively activate anchor points from coarse to fine and optimize the panoramic radiation field, fusing local details and global structural information at each level, thereby improving training stability and reconstruction accuracy. Finally, a continuous and seamless 360° image is generated through Gaussian ellipsoidal equirectangular projection (ERP), significantly improving geometric consistency and visual continuity. The overall design is efficient and robust, making it particularly suitable for scene reconstruction and high-fidelity view synthesis of 360° images.

[0020] Anchor-based Gaussian generation strategies are used to achieve structured representation and efficient parametric modeling of 3D scenes. Their structure is as follows: Figure 2 and 3 As shown, firstly, sparse point cloud data is generated from the input multi-view panoramic image using the Structure-from-Motion (SfM) algorithm, and the point cloud is voxelized to obtain the voxel centers as anchor points. Each anchor point is represented by its spatial coordinates and a set of attributes, including the anchor point feature vector, scale parameter, and offset. Subsequently, based on the anchor point position, a lightweight multilayer perceptron (Tiny MLP) is used to predict multiple sets of Gaussian parameters associated with it, including opacity, color, rotation quaternion, and scale factor, thereby generating learnable Gaussian primitives. This strategy, with anchor points as the core, constructs a mapping relationship from sparse point cloud to continuous Gaussian distribution, which not only preserves the spatial structure information of the scene but also significantly reduces Gaussian redundancy, providing structural constraints and efficient expressive capabilities for subsequent spherical mapping and rendering.

[0021] The spherical multilayer training strategy is used to achieve efficient and stable training, and its structure is as follows: Figure 2 As shown, this strategy dynamically divides the hierarchy based on the distance between the anchor point and the camera center, prioritizing anchor points that are closer and have a greater impact on the rendering results, thus significantly improving training and rendering efficiency. In the hierarchy estimation stage, this method determines the number of levels by the maximum and minimum distance range, achieving multi-layer anchor point modeling from coarse to fine. Subsequently, in the dynamic hierarchy allocation stage, the hierarchy of the anchor point is dynamically determined based on the real-time distance between the camera and the anchor point in the current training image, allowing the same anchor point to adaptively adjust to different viewpoints, thereby enhancing the model's adaptability to depth changes and viewpoint differences in wide-angle panoramic scenes. In the progressive training stage, a layer-by-layer activation mechanism is adopted, prioritizing the optimization of anchor points that are closer to the camera and have lower hierarchy levels, and gradually introducing higher-level anchor points according to the iteration step size during training. This mechanism effectively reduces redundant representations and balances training time and reconstruction quality. During the optimization process, high-quality panoramic image reconstruction and rendering are achieved by minimizing the combined loss between the rendered image and the real image (consisting of L1 loss and structural similarity loss D-SSIM weighted). This spherical multi-layer training strategy significantly reduces storage and computational overhead while maintaining rendering accuracy, providing a robust solution for efficient modeling of panoramic scenes.

[0022] The ERP projection mechanism of the Gaussian ellipsoid has the following structure: Figure 2 As shown, all 3D Gaussian ellipsoids are unfolded into 2D image coordinates using equirectangular projection (ERP), achieving a continuous mapping from the 3D scene to the panoramic image. Specifically, this is based on the position of the Gaussian center in the camera coordinate system. It can calculate its projected coordinates in the ERP image plane. This establishes the correspondence between spatial points and pixel points. This is achieved by introducing the ERP projection Jacobian matrix. Substituting this into the Gaussian covariance projection formula, we can obtain the covariance matrix projected onto the two-dimensional plane. This process involves the projection of an effective Gaussian ellipsoid from three-dimensional space onto the ERP image.

[0023] like Figure 1 The specific steps of this method are as follows:

[0024] Step (1) Obtain multi-view panoramic images and generate an initial point cloud using the SfM algorithm, as follows:

[0025] (1-1) Acquire multi-view images: Acquire panoramic images from multiple different perspectives around the scene to obtain comprehensive multi-view input data.

[0026] (1-2) Camera pose estimation: The camera pose of the acquired panoramic images is estimated using a 3D reconstruction algorithm based on Structure from Motion (SfM).

[0027] (1-3) Point cloud initialization: Calculate the sparse initial point cloud of the generated scene to provide basic geometric information for subsequent anchor point construction and Gaussian parameter generation.

[0028] Step (2) involves voxelizing the point cloud to generate anchor points, and then predicting Gaussian parameters using Tiny MLP, as detailed below:

[0029] (2-1) Perform voxelization on the sparse point cloud data generated in step (1), and the point cloud... M represents the number of points contained in the point cloud. Represents the three-dimensional spatial coordinate components of each point cloud. R represents the real number field; in terms of voxel size Divide the three-dimensional space uniformly by step size and calculate the voxel center positions: ;in, Let N represent the set of voxel center coordinates, and N be the number of effective voxels obtained after voxelization. This is a rounding operation that can effectively remove redundant coordinates in space, resulting in a regularly distributed set of anchor points.

[0030] (2-2) Define the feature set of each voxel center as an anchor point. ,in This represents an anchor feature vector with dimension 32. This represents a scale parameter with a dimension of 3. For use in generating the corresponding A 3D offset vector of the center of a neuron Gaussian.

[0031] Calculate the k Gaussian center positions corresponding to each anchor point: ;in, This indicates the center position of the anchor point. This indicates the corresponding anchor point. A set of Gaussian offset vectors This represents the center position of the k Gaussians corresponding to each anchor point.

[0032] (2-3) Calculate the distance between the anchor point and the camera center. and line-of-sight vector This provides perspective constraint information for subsequent network predictions, among which Indicates the camera's position. This represents the Euclidean distance between vectors.

[0033] (2-4) The anchor point feature vector ,distance and line-of-sight vector As input, these parameters are fed into three independent lightweight multilayer perceptron (Tiny MLP) networks to predict the k Gaussian parameters associated with the anchor point, including an opacity prediction network. Color prediction network and geometric parameter prediction network :

[0034] 1. An opacity prediction network, consisting of two linear layers, a ReLU activation layer, and a Tanh activation layer, is used to map input features to intervals. Output the set of opacity parameters for the k Gaussian ellipsoids corresponding to the anchor point. , ;

[0035] 2. The color prediction network consists of two linear layers, one ReLU activation layer, and one Sigmoid activation layer. It is used to predict the color parameter set of k Gaussian ellipsoids corresponding to the output anchor point. , ;

[0036] 3. A geometric parameter prediction network, consisting of two linear layers and one ReLU activation layer, is used to output the rotation quaternions of the k Gaussian ellipsoids corresponding to the anchor points. and scale factor parameter set , .

[0037] Through the synergistic effect of the three networks mentioned above, Gaussian parameters corresponding to each anchor point can be adaptively generated, realizing an effective mapping from sparse point clouds to learnable Gaussian primitives, and providing a structured parameter representation for subsequent panoramic rendering.

[0038] Step (3) employs a spherical multilayer training strategy to optimize the panoramic radiation field and uses a loss function to optimize the model, as detailed below:

[0039] (3-1) Hierarchical Initialization: Based on the spatial distance between the anchor points and the centers of each camera, an octree structure based on distance is established to achieve multi-level partitioning of the anchor points. First, the Euclidean distance distribution between the anchor points and the camera centers is calculated, and the maxima in the distance distribution are removed. and minimum value This eliminates outliers. Then, based on the processed distance range, the anchor point level is determined using a logarithmic ratio. This serves as the basis for subsequent anchor point grading and multi-layer modeling.

[0040] (3-2) Dynamic Anchor Point Allocation: During training, due to the sparse distribution of camera viewpoints in 3D space, a dynamic allocation strategy is adopted for anchor point levels to improve adaptability. For each forward propagation, the anchor point level is calculated in real time based on the Euclidean distance between the camera position of the current training image and each anchor point. Anchor Point In the camera The dynamic level below ,in The range normalization function scales the range to , This is the focal length scaling factor. anchor point To the camera The distance. This mechanism allows the anchor points to dynamically adjust their hierarchy as the camera pose changes, improving the model's adaptability to different viewpoints and its spatial representation ability.

[0041] (3-3) Progressive Training: To further improve the efficiency of multi-layer anchor point modeling, a layer-by-layer activation training method is adopted. In the initial stage, only anchor points that are close to each other and at a low level are activated for optimization; subsequently, after every K iterations, higher-level anchor points are gradually introduced to participate in the training. This strategy allows the model to prioritize important regions, thereby reducing computational and storage overhead while ensuring reconstruction quality.

[0042] (3-4) Loss Function Optimization: The multilayer perceptron (MLP) used in the network for predicting 3D Gaussian parameters is optimized using a combined loss function. The loss function is composed of... loss With structural similarity loss Composition, loss function ;in, The weighting coefficients are used to balance the influence of the two types of losses in the overall optimization process. By minimizing this loss function, efficient optimization of the panoramic radiation field and a significant improvement in rendering quality can be achieved.

[0043] Step (4) Project the Gaussian ellipsoid onto the ERP image using equidistant cylindrical projection, and render and output a 360-degree panoramic image, as follows:

[0044] (4-1) Calculate the coordinates on the Gaussian ellipsoid ERP image: the position of the Gaussian center in the camera coordinate system. , The position along the horizontal direction of the camera. The position is along the vertical direction of the camera. This represents the depth along the forward direction. Gaussian center point. Corresponding ERP image coordinates The calculation is as follows:

[0045] , ; These represent the height and width of the ERP image, respectively.

[0046] (4-2) Using ERP projection Jacobian matrix Substituting this into the Gaussian covariance projection formula, we obtain the covariance moments projected onto the two-dimensional plane. This is used to describe the shape and orientation of the Gaussian ellipsoid in two-dimensional projection space, where For view transformation matrix, Let T be the covariance matrix in three-dimensional space, and let T denote the transpose.

[0047]

[0048] (4-3) Based on the two-dimensional Gaussian parameters obtained from the projection, alpha blending rendering is performed on all projection results to generate a 360° panoramic image, realizing the effective mapping of the Gaussian ellipsoid from three-dimensional space to the ERP plane image.

[0049] In summary, the anchor-point-based panoramic Gaussian sputtering view synthesis method proposed in this invention achieves efficient, stable, and high-fidelity 360° panoramic scene reconstruction and rendering through the collaborative design of anchor-point-constrained Gaussian generation, spherical multi-level training, and Gaussian ellipsoid ERP projection. Structurally, this method fully integrates the sparse geometric priors generated by SfM with the continuous representation capability of Gaussian distribution. The anchor-point-driven Gaussian generation strategy achieves structured modeling of the 3D scene, significantly reducing Gaussian redundancy. The spherical multi-level training mechanism progressively optimizes the anchor point distribution and radiation field parameters from coarse to fine, improving the model's convergence efficiency and rendering stability. Finally, the Gaussian ellipsoid ERP projection mechanism achieves continuous mapping from 3D Gaussian space to 2D panoramic images, ensuring geometric consistency and visual continuity. The overall method balances rendering quality with storage capacity, making it particularly suitable for applications such as panoramic visual reconstruction, virtual reality, and immersive graph synthesis, demonstrating high engineering application value and promotion potential.

Claims

1. A method for 360 degree view synthesis based on panoramic Gaussian spheri- cal sputtering, characterized by: By the synergistic effect of anchor-based Gaussian generation strategy, spherical multi-level training strategy and Gaussian ellipsoid ERP projection mechanism, 360° panoramic reconstruction and rendering are realized. Firstly, the three-dimensional point cloud is extracted from the multi-view panoramic image by using the motion recovery structure algorithm, and the anchor point is constructed in the voxelization manner. The anchor point corresponding Gaussian parameters are predicted by lightweight multi-layer perception, realizing efficient Gaussian generation under the structure constraint. Then, through the spherical multi-level training strategy, the anchor points are gradually activated from coarse to fine, and the panoramic radiation field is optimized. At each level, local details and global structure information are fused. Finally, the continuous and seamless 360° panoramic image is generated by the Gaussian ellipsoid equidistant cylindrical projection.

2. The 360-degree view synthesis method based on panoramic Gaussian sputtering according to claim 1, characterized in that: The anchor-based Gaussian generation strategy is used to realize the structured representation and efficient parameterized modeling of the three-dimensional scene. Firstly, the sparse point cloud data is generated from the input multi-view panoramic image by using the motion recovery structure algorithm, and the point cloud is voxelized to obtain the voxel center as the anchor point position. Each anchor point is represented by its spatial coordinates and attribute set, including anchor point feature vector, scale parameter and offset. Then, according to the anchor point position, the multiple groups of Gaussian parameters associated with it are predicted by lightweight multi-layer perception, including opacity, color, rotation quaternion and scale factor.

3. The 360-degree view synthesis method based on panoramic Gaussian sputtering according to claim 1, characterized in that: The spherical multi-level training strategy is used to realize efficient and stable training. The strategy dynamically divides the levels according to the distance between the anchor point and the camera center. Anchor points with short distances are given priority in optimization. In the level estimation stage, the number of levels is determined by the maximum and minimum distance range, realizing multi-level anchor point modeling from coarse to fine. In the dynamic level allocation stage, the level of each anchor point is dynamically determined according to the real-time distance between the camera and the anchor point in the current training image, so that the same anchor point can be adjusted adaptively with different perspectives. In the progressive training stage, a layer-by-layer activation mechanism is adopted to preferentially optimize anchor points with short distances and low levels, and higher level anchor points are gradually introduced in the training process according to the iteration step.

4. The 360-degree view synthesis method based on panoramic Gaussian sputtering according to claim 1, characterized in that: The Gaussian ellipsoid ERP projection mechanism projects all three-dimensional Gaussian ellipsoids into two-dimensional image coordinates by equidistant rectangular projection, realizing the continuous mapping relationship from three-dimensional scene to panoramic image. According to the position of the Gaussian center in the camera coordinate system, the projection coordinates of the Gaussian center in the ERP image plane are calculated to establish the correspondence between the spatial points and the pixel points. The ERP projection Jacobian matrix is introduced into the Gaussian covariance projection formula to obtain the covariance matrix projected onto the two-dimensional plane.

5. The method of claim 1, wherein the panoramic Gaussian splay is based on a 360 degree view synthesis. 5 The steps are as follows: Step (1) Obtain multi-view panoramic images and generate initial point cloud by using SfM algorithm; Step (2) Voxelize the point cloud to generate anchor points and predict Gaussian parameters by lightweight multi-layer perception; Step (3) Optimize the panoramic radiation field by using spherical multi-level training strategy and optimize the model by using loss function. Step (4) renders a 360-degree panoramic image by projecting the Gaussian ellipsoid onto the ERP image in an equidistant cylindrical projection manner.

6. A method for combining 360 degree views based on panoramic Gaussian splottering as claimed in claim 5, wherein, Step (1) is specifically as follows: (1-1) Collecting multi-view images; a plurality of panoramic images of different perspectives are collected by surrounding the scene to obtain complete multi-view input data; (1-2) Camera pose estimation; a three-dimensional reconstruction algorithm based on structure from motion is used to estimate the camera pose of the collected panoramic images; (1-3) Point cloud initialization; a sparse initial point cloud of the scene is calculated and generated to provide basic geometric information for subsequent anchor point construction and Gaussian parameter generation.

7. A method for combining 360 degree views based on panoramic Gaussian splottering as claimed in claim 6, wherein, Step (2) is specifically as follows: (2-1) voxelization processing is performed on the generated sparse point cloud data, and the point cloud , M represents the number of point clouds contained in the point cloud, , and X represents the three-dimensional space coordinate components of each point cloud , R represents the real field; the three-dimensional space is uniformly divided with a step size of voxel size , and the voxel center position is calculated: ; wherein, , represents the voxel center coordinate set, N is the number of effective voxels obtained after voxelization, is a rounding operation. (2-2) Define its feature set as the anchor point of each voxel center wherein denotes an anchor feature vector of dimension 32, denotes a scale parameter of dimension 3, is a 3-dimensional offset vector for generating the corresponding neural Gaussian centers. Calculate the k Gaussian center positions corresponding to each anchor point: ;in, This indicates the center position of the anchor point. This indicates the corresponding anchor point. A set of Gaussian offset vectors This represents the center position of the k Gaussians corresponding to each anchor point; (2-3) Calculate the distance between the anchor point and the camera center and the line-of-sight direction vector where denotes the position of the camera, denotes the Euclidean distance of the vectors; (2-4) the anchor point feature vector , distance and line of sight direction vector are input as inputs into three independent lightweight multi-layer perceptron networks for predicting k Gaussian parameters associated with the anchor point, including opacity prediction network , color prediction network and geometry parameter prediction network : Through the cooperative action of the three networks, the Gaussian parameters corresponding to each anchor point are adaptively generated, and effective mapping from sparse point cloud to learnable Gaussian primitives is realized.

8. A method for combining 360 degree views based on panoramic Gaussian splottering as claimed in claim 7, wherein, Step (3) is specifically as follows: (3-1) Hierarchical initialization: According to the spatial distance between anchor points and the center of each camera, the distance-based octree structure is established to realize the multi-level division of anchor points; First, the Euclidean distance distribution between the anchor points and the camera center is calculated, and the maximum value in the distance distribution is removed and the minimum value ; Then according to the processed distance range, the logarithmic proportional relationship is used to determine the number of anchor point levels , which is used as the level basis for subsequent anchor point grading and multi-layer modeling; (3-2) Dynamic anchor point assignment: In the training process, a dynamic assignment strategy is adopted for anchor point levels; for each forward propagation, the level to which each anchor point belongs is calculated in real time according to the Euclidean distance between the camera position of the current training image and each anchor point; the anchor points are assigned to different levels according to the calculated level In the dynamic level assignment of the camera under the camera , wherein is a range normalization function that scales the range to , is a focal length scaling factor, is the distance from the anchor point to the camera ; (3-3) Progressive training: a layer-by-layer activation training method is adopted, and only anchor points with short distances and low levels are activated for optimization in the initial stage; then, after a plurality of iterations, higher level anchor points are gradually introduced to participate in the training; (3-4) Loss function optimization: the multi-layer perceptron in the network for predicting the three-dimensional Gaussian parameters is optimized using a combined loss function; the loss function ; is a structural similarity loss, loss, is a structural similarity loss, is a balance weight coefficient.

9. A method for combining 360 degree views based on panoramic Gaussian splottering as claimed in claim 8, wherein, Step (4) is specifically as follows: (4-1) Calculate the coordinates on the Gauss ellipsoid ERP image: the position of the Gauss center in the camera coordinate system , is the position along the horizontal direction of the camera, is the position along the vertical direction of the camera, is the depth along the forward direction; the Gauss center point the corresponding ERP image coordinates is calculated as follows: , ; H and W are the height and width of the ERP image, respectively; (4-2) Using the ERP projection Jacobian matrix Substituting this into the Gaussian covariance projection formula, we obtain the covariance matrix projected to the two-dimensional plane to describe the shape and direction of the Gaussian ellipsoid in the two-dimensional projection space, where is the view transformation matrix, is the covariance matrix of the three-dimensional space, and T denotes the transpose; (4-3) Based on the two-dimensional Gaussian parameters obtained by projection, alpha blending rendering is performed on all projection results to generate a 360-degree panoramic image, realizing effective mapping of the Gaussian ellipsoid from three-dimensional space to the ERP plane image.

10. The 360-degree view synthesis method based on panoramic Gaussian sputtering according to claim 7, characterized in that: The opacity prediction network consists of two linear layers, one ReLU activation layer and one Tanh activation layer, which is used to map the input features to the interval , and outputs the opacity parameter set of the k Gaussian ellipsoids corresponding to the anchor points , ; The color prediction network consists of two linear layers, a ReLU activation layer and a Sigmoid activation layer, and is used for predicting a color parameter set of the k Gaussian ellipsoids corresponding to the output anchor point , ; The geometric parameter prediction network consists of two linear layers and a ReLU activation layer, and is used for outputting a rotation quaternion of the k Gaussian ellipsoids corresponding to the anchor points and scale factors a parameter set , .

Images