A gaussian-splatting based level of detail rendering method

Abstract

The application discloses a kind of detail level rendering method based on Gaussian splatting, can change the detail level of model by the need of user in actual in interactive process, realize high fidelity, low resource loss rendering based on Gaussian splatting, reduce the loss required by rendering.The technical scheme of the application includes preparation stage and real-time rendering stage.Preparation stage, obtain multi-view image, and construct corresponding three-dimensional model or four-dimensional model according to multi-view image, generate Gaussian splatting model with different detail levels.Real-time rendering stage, using Gaussian splatting model with different detail levels, select Gaussian splatting model with different detail levels for real-time rendering according to real-time data of user.

Description

Technical Field

[0001] This invention relates to the fields of computer vision and human-computer interaction, specifically to a detail rendering method based on Gaussian splashing. Background Technology

[0002] Gaussian splashing is a 3D volume rendering technique that renders 3D volume data directly without converting it into lines and planes. With advancements in computer graphics, 3D and 4D Gaussian splashing techniques have been developed and used to provide real-time radiation field rendering and dynamic scene rendering. 3D Gaussian splashing (3DGS) improves rendering quality while enabling real-time model display. This method integrates sparse points generated during camera calibration, represents the scene using a 3D Gaussian function, and preserves the characteristics of a continuous volumetric radiation field. Furthermore, it introduces alternating optimization and density control of the 3D Gaussian function, supporting anisotropic Gaussian splashing. 4D Gaussian splashing adds time to the 3D model, enabling dynamic scene rendering.

[0003] Patent CN117671108A discloses a method for dynamic human body modeling based on 3D Gaussian. This invention discloses a method for dynamic human body modeling based on 3D Gaussian. The method first establishes a 3D Gaussian representation of the human body using multi-view human body videos as input. The method attaches an implicit code and a set of mixing weights to each Gaussian. Then, based on a given target human body pose, a multilayer perceptron is used for correction. Next, the corrected Gaussian sputtering is performed through Gaussian transformation using the mixing weights, and then transformed to the target pose in a linear blending skinning manner. Finally, a realistic human body image under a new viewpoint or pose can be rendered in real time using Gaussian sputtering.

[0004] This invention proposes a method for implementing dynamic Gaussian splashing of the human body, enabling dynamic human body rendering within the Gaussian splashing rendering pipeline. However, because this method is not model-based during Gaussian splashing initialization, the number and position of Gaussian splashes are uncontrollable, thus preventing the implementation of multi-level detail Gaussian splashing rendering. Furthermore, due to the lack of variation in detail levels, adjustable detail levels cannot be achieved during rendering and interaction.

[0005] Paper: GaussianAvatars: Photorealistic Head Avatars with Rigged3DGaussians.

[0006] This paper proposes a dynamic 3D representation method for the human head based on Gaussian splashing. This method involves fixing Gaussian splashes onto a deformable 3D facial model. First, a 3D Gaussian splash is initialized at the center of each triangle in the model. Each Gaussian splash is translated, rotated, and scaled according to the face of its surrounding triangle. For each frame of acquired 4D data, the Gaussian splashes are transformed from the face space of the local triangles to world space, and then rendered using the 3D Gaussian splashes. During network training, the Gaussian splashes in local space are optimized by minimizing color loss during rendering. Simultaneously, new Gaussian splashes are adaptively added and removed based on the existing binding relationships between Gaussian splashes and triangles, and the position and scaling of the 3D Gaussian splashes are regularized. Then, a radiation field is formed on the model surface by training the 3D Gaussian splashes. This method creates a movable head avatar by binding the 3D Gaussian splashes to a parametric mesh model while maintaining high rendering quality. This method provides a solution for 4D dynamic Gaussian splash rendering, focusing on how to make the 3D Gaussian splashes move, but without considering rendering overhead. The location of the Gaussian splashes generated in the result of this method is obtained during the training process, and the number is uncontrollable, which makes it impossible to adjust the level of detail of the trained 3D Gaussian.

[0007] It is evident that existing Gaussian splash rendering methods are unable to adjust the level of detail, resulting in significant rendering losses in practical virtual reality or augmented reality applications. Currently, no technical solution has been found to address this problem. Summary of the Invention

[0008] In view of this, the present invention provides a detail level rendering method based on Gaussian splashing, which can dynamically change the detail level of the model according to the user's actual needs in the interaction process, so as to achieve high-fidelity, low-resource-consumption Gaussian splashing-based rendering and reduce the rendering loss required.

[0009] To achieve the above objectives, the technical solution of the present invention includes a preparation stage and a real-time rendering stage.

[0010] In the preparation phase, multi-view images are acquired, and corresponding 3D or 4D models are constructed based on the multi-view images to generate Gaussian splash models with different levels of detail.

[0011] During the real-time rendering phase, Gaussian splash models with different levels of detail are used, and the model is selected based on the user's real-time data for real-time rendering.

[0012] Furthermore, in the preparation stage, multi-view images are images of the same object taken from multiple different perspectives, used to represent objects that require the application of Gaussian splash radiation fields. Both 3D and 4D models consist of a series of vertices and connecting lines, and the model is divided into a series of model surfaces.

[0013] Furthermore, the multi-view images can be single frames or sequences of frames; if they are single-frame multi-view images, they are used to construct a detailed 3D model of Gaussian splashes; if they are sequences of multi-view images, they are used to construct a detailed 4D model of Gaussian splashes.

[0014] Furthermore, in the preparation stage, Gaussian splash models with different levels of detail are generated, specifically through the following steps:

[0015] Step 101: Initialize Gaussian splashes at the centroid of each vertex and each face of the 3D or 4D model. Each Gaussian splash has five basic parameters, including position, rotation, scaling, opacity, and a set of spherical harmonic functions representing color information. For the Gaussian splash at the centroid of each face, its spatial position, spatial rotation, and spatial scaling are obtained from the Gaussian splash information of the vertex position in the same plane corresponding to the centroid, thereby obtaining the position of the initialized Gaussian splash.

[0016] Step 102: The initial Gaussian splash locations are used for training the Gaussian splash network. The parameters of the Gaussian points are continuously optimized to generate different levels of detail, thus obtaining Gaussian splash models with different levels of detail.

[0017] Further, in step 102, the positions of the initialized Gaussian splashes are used for training the Gaussian splash network. The training process specifically involves optimizing the position, rotation, scaling, opacity, and color information of the Gaussian points. During training, for Gaussian splashes at model vertices, their spatial positions are not changed; only their rotation, scaling, opacity, and color information are altered. For Gaussian splashes at the centroid, their position, rotation, scaling, opacity, and color information all change. The number of Gaussian splashes remains unchanged during network training. After optimizing all Gaussian splashes through the training process, a Gaussian splash model with a level of detail is obtained.

[0018] Further, in step 102, different levels of detail are generated to obtain Gaussian splash models with different levels of detail. Specifically, the model surfaces are subdivided. Specifically, for a 3D or 4D model, the Gaussian splashes at the centroid of each model surface are moved to new positions after network training. By connecting the new position points with the original vertices, a series of new model surfaces are formed.

[0019] While keeping the position, rotation, scaling, opacity, and color information of the existing Gaussian splashes unchanged, initialize a series of new Gaussian splashes at the centroid position of the new model face.

[0020] By repeating the training process of the Gaussian splash network and continuously optimizing the new Gaussian splashes, new details can be added to the existing level of detail, thus obtaining a new level of detail Gaussian splash model.

[0021] The new Gaussian splash model at a level of detail is evaluated by reprojecting it onto the camera position. If the image similarity between the new Gaussian splash model and the original multi-view image reaches a set index range, the new Gaussian splash model at a level of detail is saved; otherwise, it is deleted. The camera position refers to the position of the camera that acquired the multi-view image.

[0022] Furthermore, during the real-time rendering stage, the user's real-time data includes: the user's position, the direction of the user's head, and the real-time position of the user's hands.

[0023] Furthermore, in the real-time rendering stage, different levels of detail are selected based on the user's real-time data. Specifically, decision tree models, neural symbolic decision models, or multilayer perceptrons are used to map the user's real-time data onto Gaussian splash models of different levels of detail.

[0024] Beneficial effects:

[0025] 1. This invention proposes a level-of-detail rendering method based on Gaussian splashing. This method dynamically changes the level of detail of the model according to the user's needs during actual interaction, achieving high-fidelity, low-resource-consumption Gaussian splashing-based rendering. This invention introduces level-of-detail into Gaussian splashing rendering, which can generate 3D or 4D models of varying precision, making it applicable to a wider range of application scenarios. Furthermore, by changing the level of detail of the model during real-time interaction, the rendering overhead is reduced.

[0026] 2. The method of this invention, after obtaining Gaussian splashing 3D or 4D models with different levels of detail, selects and replaces models with different levels of detail based on the user's real-time data during real-time interaction. During user interaction, real-time data, including but not limited to user position, head orientation, and hand position, is first acquired. Then, the level of detail to be rendered in the current situation is selected based on the user's real-time data. The selection methods include, but are not limited to, decision trees, neural symbolic decision models, and multilayer perceptrons. For example, one possible method is to use a decision tree model to map the user's real-time information onto models with different levels of detail. When the user is far from the model, a model with a lower level of detail is used for rendering; when the user is close to the model, a higher level of detail is used. This method can reduce the cost of rendering computation while ensuring high rendering quality for the user's subjective experience, achieving multi-level-of-detail rendering based on Gaussian splashing. Attached Figure Description

[0027] Figure 1 This is a schematic diagram of the detail rendering method based on Gaussian splashing;

[0028] Figure 2 This is a schematic diagram of model surface subdivision and Gaussian splashing network training. Detailed Implementation

[0029] The present invention will now be described in detail with reference to the accompanying drawings and embodiments.

[0030] This invention aims to introduce level of detail into Gaussian splash rendering to achieve high-quality, low-computational-cost Gaussian splash rendering. 3D and 4D Gaussian splashing techniques can render highly realistic 3D or 4D data and offer currently superior rendering performance. However, the number and position of points in a Gaussian splash model are uncontrollable, leading to two problems: first, in some situations, Gaussian splashing can be too dense, sometimes making it unnecessary to render all details; second, due to the aforementioned problems, it is impossible to generate Gaussian splash models with different levels of detail, nor is dynamic level of detail replacement possible.

[0031] Based on this, the present invention provides a detail level rendering method based on Gaussian splashing. The process of the present invention is as follows: Figure 1 As shown, the method consists of two parts: a level-of-detail preparation stage and a real-time rendering stage.

[0032] In the preparation phase, multi-view images are acquired, and corresponding 3D or 4D models are constructed based on these images, generating Gaussian splash models with different levels of detail. The multi-view images are images of the same object taken from multiple different perspectives, used to represent the object requiring the application of the Gaussian splash radiation field. These multi-view images can be single frames or a sequence of frames. If it's a single-frame multi-view image, it's used to construct a 3D model with detailed levels of Gaussian splash detail; if it's a sequence of frames, it's used to construct a 4D model with detailed levels of Gaussian splash detail. Both the 3D and 4D models consist of a series of vertices and connecting lines, and the model is divided into a series of model faces.

[0033] In this embodiment of the invention, the level of detail preparation stage is a preliminary offline process used to prepare Gaussian splashes of different levels of detail required for real-time rendering. In this stage, different Gaussian splashes are first generated for the 3D model of the object to be rendered, and these splashes are bound to the vertices and centroids of the faces on the model surface (taking triangular faces as an example). Subsequently, the Gaussian splash network is trained using multi-view images from a sequence of acquired frames. During training, the splitting and deletion of Gaussian splashes are restricted to control the number of splashes, resulting in a Gaussian splash model with a low level of detail. Then, the 3D model surface is subdivided into meshes. While retaining the existing Gaussian splashes, the splashes are reinitialized at the centroid of each face of the subdivided 3D model, and the optimization process of the Gaussian splash network training is repeated to obtain a Gaussian splash model with a higher level of detail.

[0034] The following steps are used to generate Gaussian splash models with different levels of detail:

[0035] Step 101: Initialize Gaussian splashes at the centroid of each vertex and each face of the 3D or 4D model. Each Gaussian splash has five basic parameters, including position, rotation, scaling, opacity, and a set of spherical harmonic functions representing color information. For the Gaussian splash at the centroid of each face, its spatial position, spatial rotation, and spatial scaling are obtained from the Gaussian splash information of the vertex position in the same plane corresponding to that centroid, thereby obtaining the position of the initialized Gaussian splash.

[0036] Step 102: The initial Gaussian splash locations are used for training the Gaussian splash network. The parameters of the Gaussian points are continuously optimized to generate different levels of detail, thus obtaining Gaussian splash models with different levels of detail.

[0037] The training process specifically involves: optimizing the position, rotation, scaling, opacity, and color information of Gaussian points; during training, for Gaussian splashes at model vertices, their spatial position remains unchanged, only their rotation, scaling, opacity, and color information are altered; for Gaussian splashes at the centroid, their position, rotation, scaling, opacity, and color information are all changed; the number of Gaussian splashes remains unchanged during network training; after optimizing all Gaussian splashes through the training process, a Gaussian splash model with a level of detail is obtained.

[0038] During training, the accuracy of the Gaussian splash model is improved by restricting the splitting and deletion of Gaussian splashes and by subdividing the model surfaces, thereby obtaining models with different levels of detail.

[0039] After training using the initialized Gaussian splash locations, a Gaussian splash model with the lowest level of detail is obtained.

[0040] The model surfaces are subdivided; specifically, for a 3D or 4D model, the Gaussian splashes at the centroid of each model surface are moved to new positions after network training; by connecting the new position points with the original vertices, a series of new model surfaces are formed.

[0041] While keeping the position, rotation, scaling, opacity, and color information of the existing Gaussian splashes unchanged, initialize a series of new Gaussian splashes at the centroid position of the new model face;

[0042] By repeating the training process of the Gaussian splash network and continuously optimizing the new Gaussian splash, new details can be added to the existing level of detail, thereby obtaining a new level of detail Gaussian splash model.

[0043] The new Gaussian splash model at a level of detail is evaluated by reprojecting it onto the camera position. If the image similarity between the new Gaussian splash model and the original multi-view image reaches a set index range, the new Gaussian splash model at a level of detail is saved; otherwise, it is deleted. The camera position refers to the position of the camera that acquired the multi-view image.

[0044] During the real-time rendering phase, Gaussian splash models with different levels of detail are used, and the model is selected based on the user's real-time data for real-time rendering.

[0045] In this embodiment of the invention, during the real-time rendering stage, different levels of detail models are dynamically selected and replaced based on the user's real-time data, achieving real-time rendering with varying levels of detail based on Gaussian splashing. The core of this research lies in introducing levels of detail into Gaussian splashing technology. This method can reduce resource consumption in real-time rendering while maintaining high-quality rendering. This method is applicable to various real-time interactive scenarios based on Gaussian splashing rendering.

[0046] Building upon existing levels of detail, during real-time rendering and use, models in the virtual environment are selected and replaced by capturing real-time user data, achieving real-time rendering based on Gaussian splashing. This approach reduces computational overhead while maintaining the rendering quality required by the user.

[0047] Example 1

[0048] The process in this embodiment is as follows: Figure 1 As shown, the process includes two parts: a preliminary level-of-detail preparation process and a real-time rendering process. In the level-of-detail preparation process, the invention generates Gaussian splash models with different levels of detail using 3D or 4D multi-view images and corresponding 3D models. In the real-time rendering process, using the aforementioned different level-of-detail models, different levels of detail are selected based on the user's real-time data, enabling the replacement of different levels of detail, thereby achieving real-time level-of-detail rendering based on Gaussian splashing.

[0049] Data preparation and Gaussian splash initialization

[0050] In the level-of-detail preparation stage, the first step is to acquire the multi-view images and corresponding 3D models required for Gaussian splashing. Multi-view images are images of the same object taken from multiple different perspectives, used to represent the object for which the Gaussian splashing radiation field will be applied. Since the more multi-view images used in Gaussian splashing, the better the viewing effect from different angles, preferably, these images need to capture information about the object's various angles and directions in 3D space. Multi-view images can be single frames or a sequence of frames. Single-frame multi-view images are used for the level-of-detail model of 3D Gaussian splashing, while a sequence of frame multi-view images is used for the level-of-detail model of 4D Gaussian splashing.

[0051] A three-dimensional model consists of a series of vertices and connecting lines. Typically, such a model can be represented by a series of triangular or quadrilateral faces. In obtaining a three-dimensional model of an object represented by a Gaussian splash radiation field using this method, the three-dimensional model can be obtained by hand-making, image-based generation (such as FLAME[1]), or by using an existing standard model (such as SMPL[2]).

[0052] Next, the positions of the Gaussian splashes are initialized. At each vertex and the centroid of each surface in this 3D model, the Gaussian splashes need to be initialized, such as... Figure 2 As shown on the left. Each Gaussian splash has five basic parameters, including position, rotation, scaling, opacity, and a set of spherical harmonic functions representing color information. For each Gaussian splash at the centroid of a surface, its spatial position, spatial rotation, and spatial scaling can be obtained from the Gaussian splash information at the vertex position in the same plane corresponding to that centroid. The positions of the initialized Gaussian splashes can then be used to train the Gaussian splash network, continuously optimizing the parameters of the Gaussian points and generating different levels of detail.

[0053] Gaussian splashing network training

[0054] After obtaining the initial Gaussian splatter spatial position, the position, rotation, scaling, opacity and color information of all Gaussian splatters are optimized according to the acquired single frame or multiple frames of multi-view images. This method adopts the method in the literature "3DGaussian Splatting for Real-Time Radiance Field Rendering" [3], projects the existing Gaussian splatters onto the camera imaging plane position corresponding to each frame of the multi-view image, calculates the loss, and optimizes the Gaussian points using gradient descent. In this scheme, there are two main steps: the first step is to optimize the position, rotation, scaling, opacity attributes and color information of the Gaussian points, and the second step is adaptive density control. Unlike this scheme, in the training process of this invention, in the first step, for the Gaussian splatter at the vertex of the model, its spatial position is not changed, only its rotation, scaling, opacity attributes and color information are changed; for the Gaussian splatter at the centroid position, its position, rotation, scaling, opacity and color information can all change. Furthermore, this invention eliminates adaptive density control, i.e., it eliminates the steps of Gaussian splash splitting and deletion, thus maintaining the number of Gaussian splashes during network training. After optimizing all Gaussian splashes through network training, a Gaussian splash model with the lowest level of detail can be obtained.

[0055] Model surface subdivision and level of detail implementation

[0056] By subdividing the surfaces of a 3D model, more surfaces can be generated within the model. This allows for the re-initialization of Gaussian splash positions on these new surfaces, increasing the number of Gaussian splashes and achieving model partitioning at different levels of detail. Figure 2 As shown.

[0057] The subdivision method for model faces can be limited to loop subdivision. Taking a triangular face model as an example, after network training, the Gaussian splashes at the centroid of each face are moved to new positions. By connecting these points with the positions of the original vertices, a series of new triangular faces can be formed. While keeping the position, rotation, scaling, opacity, and color information of the existing Gaussian splashes unchanged, a series of new Gaussian splashes are initialized at the centroid positions of the newly generated triangular faces. Then, the network training is repeated, and by continuously optimizing the new Gaussian splashes, new details can be added to the existing level of detail, thereby obtaining a higher level of detail.

[0058] When evaluating the level of detail in a Gaussian splash model, image similarity metrics can be used, including but not limited to peak signal-to-noise ratio, structural similarity, and learned perceptual patch similarity. In practice, the range of image similarity metrics for level of detail can be preset in advance. When the Gaussian splash model is reprojected onto the camera position during iterations and the image similarity with the original input image reaches the metric range, the model is saved as a level-of-detail model. This allows for the continuous addition of new details to the existing low-level detail.

[0059] For example, this embodiment can use an image similarity metric – structural similarity. The value of structural similarity ranges from -1 to 1, and the larger the value, the better the image quality.

[0060] After the initial network training step of the Gaussian splash model, the resulting level of detail is called LOD 0, with a structural similarity of 0.2. Subsequently, a range of structural similarity metrics is manually defined; for example, LOD 1 is defined as structural similarity of 0.4, LOD 2 as 0.6, and LOD 3 as 0.8. This process is iterated continuously, with the structural similarity of the current model being assessed during each iteration. When the structural similarity reaches 0.4, the model is saved as an LOD 1 model, and so on, thus obtaining Gaussian splash models with varying levels of detail.

[0061] Real-time selection, replacement, and rendering at a level of detail

[0062] After acquiring Gaussian splashing 3D or 4D models with different levels of detail, this method selects and replaces models with different levels of detail based on real-time user data during real-time interaction. During user interaction, real-time data, including but not limited to user position, head orientation, and hand position, is first acquired. Then, the level of detail to be rendered in the current situation is selected based on the user's real-time data. The selection methods include, but are not limited to, decision trees, neural symbolic decision systems, and multilayer perceptrons. For example, one possible method is to use a decision tree model to map the user's real-time information onto models with different levels of detail. When the user is far from the model, a lower level of detail model is used for rendering; when the user is close to the model, a higher level of detail is used. This method can reduce the cost of rendering computation while ensuring high rendering quality for the user's subjective experience, achieving multi-level-of-detail rendering based on Gaussian splashing.

[0063] In summary, the above are merely preferred embodiments of the present invention and are not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for rendering level of detail based on Gaussian splatting, characterized in that, The method comprises a preparation stage and a real-time rendering stage; The preparation stage comprises: acquiring multi-view images, and constructing a corresponding three-dimensional model or four-dimensional model according to the multi-view images to generate Gaussian spatter models with different levels of detail; the multi-view images are images of the same object taken from multiple different angles and are used to represent an object that needs to use a Gaussian spatter radiation field; the three-dimensional model or the four-dimensional model is composed of a series of vertices and connecting lines, and the model is divided into a series of model surfaces; The real-time rendering stage uses the Gaussian spatter models with different levels of detail to select Gaussian spatter models with different levels of detail according to real-time data of a user for real-time rendering; The Gaussian spatter models with different levels of detail are generated by the following steps: Step 101: initializing a Gaussian spatter at the center of gravity of each vertex and each model surface of the three-dimensional model or the four-dimensional model, each Gaussian spatter having five basic parameters, including position, rotation, scaling, opacity, and a set of spherical harmonic functions representing color information; for the Gaussian spatter at the center of gravity of each model surface, the spatial position, spatial rotation, and spatial scaling are obtained from the Gaussian spatter information of the vertex position in the same plane corresponding to the center of gravity, thereby obtaining the position of the initialized Gaussian spatter; Step 102: the position of the initialized Gaussian spatter is used for training of a Gaussian spatter network to continuously optimize the parameters of the Gaussian points to generate different levels of detail and obtain Gaussian spatter models with different levels of detail; The training process comprises: optimizing the position, rotation, scaling, opacity, and color information of the Gaussian points; during the training process, for the Gaussian spatter at the model vertex, the spatial position is not changed, and only the rotation, scaling, opacity, and color information are changed; for the Gaussian spatter at the center of gravity, the position, rotation, scaling, opacity, and color information are all changed; during the network training process, the number of Gaussian spatters is not changed; after all the Gaussian spatters are optimized through the training process, a Gaussian spatter model with a level of detail is obtained; Subdividing the model surface; specifically, for the three-dimensional model or the four-dimensional model, the Gaussian spatter at the center of gravity of each model surface is moved to a new position after the network training; a series of new model surfaces are formed by connecting the new position points and the original vertices; Under the condition that the position, rotation, scaling, opacity, and color information of the existing Gaussian spatter are unchanged, a series of new Gaussian spatters are initialized at the center of gravity of the new model surface; The training process of the Gaussian spatter network is repeated, and new details can be added to the existing level of detail by continuously optimizing the new Gaussian spatters, thereby obtaining a new Gaussian spatter model with a level of detail; The new level of detail Gaussian splat model is evaluated, i.e. the new level of detail Gaussian splat model is re-projected to the camera position, and the image similarity with the original multi-view image is within the set index range, then the new level of detail Gaussian splat model is saved, otherwise the new level of detail Gaussian splat model is deleted; the camera position is the position of the camera for acquiring the multi-view image.

2. A Gaussian-spur based level of detail rendering method according to claim 1, wherein, The multi-view image is a single frame or a sequence frame. If the multi-view image is a single frame, a level of detail three-dimensional model for constructing a Gaussian splat is used. If the multi-view image is a sequence frame, a level of detail four-dimensional model for constructing a Gaussian splat is used.

3. The Gaussian-spur based level of detail rendering method of claim 1, wherein, The real-time rendering stage, the real-time data of the user, includes real-time data of the user position, the user head orientation, and the user hand position.

4. A Gaussian-spur based level of detail rendering method according to claim 3, wherein, The real-time rendering stage selects different levels of detail according to the real-time data of the user, specifically: The real-time data of the user is mapped to different levels of detail Gaussian splat models by using a decision tree model, a neural symbolic decision model, or a multi-layer perception machine.

Images