Web rendering implementation method suitable for 3DGS large-scale scene of consumer equipment

Abstract

The invention discloses a Web rendering implementation method suitable for a 3DGS large-scale scene of consumption-level equipment, and the method comprises the steps: determining a detail level based on space segmentation and multiple factors through LOD optimization, and building a rapid switching mechanism; compressing geometric, texture and LOD model data, and respectively adopting specific algorithms and formats; a parallel decoding algorithm is designed at an equipment end to realize efficient decoding; a dynamic LOD calculation model based on deep learning is introduced and optimized; combining double-buffer coordinated rendering with edge calculation; adaptively adjusting ray tracing according to the scene and the equipment performance; and establishing an LRU cache management mechanism. According to the method, the hardware dependence is reduced, the data volume is reduced, the dynamic LOD calculation efficiency and rendering performance are improved, the rendering problem of consumption-level equipment is effectively solved, and wide application of the 3DGS technology to the consumption-level equipment is promoted.

Description

Technical Field

[0001] The present invention belongs to the technical field related to three-dimensional graphics rendering, and in particular relates to a Web rendering implementation method for 3DGS large-scale scenes suitable for consumer-grade devices. Background Art

[0002] In the field of 3D graphics rendering, large-scale scene rendering holds broad application prospects, such as virtual city displays and large-scale game scene presentation. However, this field currently faces numerous challenges: First, large-scale scene rendering often requires professional-grade GPU hardware, making smooth rendering difficult on common consumer devices (such as regular computers and mobile phones) due to hardware performance limitations. Second, the mainstream approach to large-scale scene rendering is based on level of detail (LOD). Most current methods rely on preset LOD schemes, such as CityGaussian or Octree-GS. Due to the variable perspectives of 3D rendering, a sufficient number of LODs must be preset, which leads to a rapid increase in data volume and significantly increases memory and video memory requirements, making these methods difficult to apply to common devices. In contrast, dynamic LOD balances storage and computational requirements through real-time calculations without increasing data volume. However, most existing dynamic LOD methods, such as VastGaussian, are computationally complex, and large 3DGS scenes require a large number of Gaussian elements. Simple and efficient dynamic LOD calculation methods are rare. These issues severely limit the widespread application of 3DGS technology on consumer devices. Summary of the Invention

[0003] The purpose of the present invention is to provide a method for implementing Web rendering of 3DGS large-scale scenes suitable for consumer-grade devices to solve the problems raised in the above background technology.

[0004] To achieve the above object, the present invention provides the following technical solutions:

[0005] A method for implementing Web rendering of 3DGS large-scale scenes suitable for consumer-grade devices includes the following steps:

[0006] LOD optimization: Based on the scene's geometric structure and texture information, the scene is divided into multiple sub-areas using a spatial segmentation algorithm. For each sub-area, the appropriate level of detail is determined based on its distance from the viewpoint and its importance, and a fast switching mechanism between LOD models is established.

[0007] Data compression: compress the geometric data, texture data and LOD model data in the 3DGS scene. The geometric data adopts a compression algorithm based on model simplification, the texture data uses an efficient texture compression format, and the LOD model data adopts an incremental encoding method;

[0008] Efficient decoding: On consumer devices, we design fast decoding algorithms for different data types and leverage the device's multi-core processor resources for parallel decoding.

[0009] Dynamic LOD calculation: A deep learning-based dynamic LOD calculation model is introduced. This model uses the scene's real-time state information as input to predict the optimal level of detail. A lightweight neural network structure is used, combined with model pruning and quantization optimization techniques, to design an incremental update mechanism.

[0010] Double-buffered coordinated rendering: Using double-buffering technology, the current frame is rendered in one buffer while the next frame's data is prepared in another buffer. The timing of buffer switching is precisely controlled. Distributed rendering technology based on edge computing is also introduced to dynamically adjust task allocation based on device performance and load.

[0011] Adaptive ray tracing: Dynamically adjusts the accuracy and sampling rate of ray tracing based on scene complexity and device performance, and combines ambient occlusion and shadow mapping technologies to enhance scene realism.

[0012] Cache management optimization: Establish a cache management mechanism based on least recently used (LRU) to cache commonly used rendering data and manage it in different levels according to data access frequency and importance.

[0013] Preferably, in the LOD optimization step, the scene is divided into sub-regions using an octree segmentation algorithm, and the initial detail level is allocated by calculating the Euclidean distance between the sub-region and the viewpoint combined with the importance of key objects in the sub-region. When the LOD model is switched, linear interpolation is used for transition.

[0014] Preferably, in the data compression step, the geometric data is compressed using a model simplification algorithm based on edge folding to compress the geometric data to 30%-50% of the original data volume. The texture data is compressed in a manner selected according to the device support format, and the compression ratio is controlled at 1:4-1:8. The LOD model data uses incremental encoding to reduce the data volume by 70%-80%.

[0015] Preferably, in the efficient decoding step, independent threads are allocated to the decoding tasks of geometric data, texture data and LOD model data, and decoding operations are performed in parallel, thereby shortening the data decoding time to 1 / 3-1 / 2 of the original time.

[0016] Preferably, in the dynamic LOD calculation step, a large amount of viewpoint position, viewing direction, movement speed and corresponding optimal LOD configuration data in different scenes are collected to construct a training data set, and a lightweight convolutional neural network containing several convolutional layers, pooling layers and fully connected layers is used as a calculation model to train and optimize model parameters, obtain scene status information input into the model in real time, and the dynamic LOD calculation update time is controlled within 10-20 milliseconds.

[0017] Preferably, in the double-buffer coordinated rendering step, buffer A and buffer B are created, buffer A performs current frame rendering, buffer B prepares next frame data, and the display is switched immediately after rendering is completed. The edge server dynamically adjusts the amount of rendering tasks assigned to the device based on the device performance indicators and current load conditions.

[0018] Preferably, in the adaptive ray tracing step, the scene complexity is evaluated and the device performance parameters are obtained before rendering, the initial ray tracing sampling rate and accuracy are set, the device frame rate and resource usage are monitored in real time during the rendering process, the ray tracing parameters are dynamically adjusted, and the ambient occlusion and shadow mapping techniques are combined to enhance the scene realism.

[0019] Preferably, in the cache management optimization step, a special cache space is opened in the memory of the consumer-grade device, which is divided into areas of texture data cache, geometry data cache and LOD model data cache. The LRU cache replacement strategy is adopted to divide the data into three levels: high-frequency important data, medium-frequency general data and low-frequency secondary data. The cache data is cleaned and updated regularly to improve the cache hit rate.

[0020] Compared with the existing technology, the present invention provides a method for implementing Web rendering of 3DGS large-scale scenes suitable for consumer-grade devices, which has the following beneficial effects:

[0021] Reduced hardware dependency: Through data compression, efficient decoding, and optimized rendering algorithms, the hardware performance requirements are significantly reduced, enabling consumer-grade devices to achieve web rendering of large-scale 3DGS scenes without relying on professional-grade GPU hardware;

[0022] Reduce data volume: The application of LOD optimization and data compression technology effectively reduces the amount of scene data, lowers the requirements for memory and video memory, and solves the problem of data expansion caused by the preset LOD method; Improve dynamic LOD calculation efficiency: The dynamic LOD calculation model based on deep learning, combined with a lightweight structure and incremental update mechanism, significantly reduces the calculation complexity while ensuring calculation accuracy, realizing simple and efficient dynamic LOD calculation;

[0023] Improve rendering performance: The comprehensive application of technologies such as double-buffered coordinated rendering, distributed rendering based on edge computing, adaptive ray tracing, and cache management optimization ensures the smoothness of the rendering process, improves rendering efficiency and quality, provides users with a good visual experience, and promotes the widespread application of 3DGS technology on consumer-grade devices. BRIEF DESCRIPTION OF THE DRAWINGS

[0024] Figure 1 Schematic diagram of the method of the present invention. DETAILED DESCRIPTION

[0025] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0026] The present invention provides Figure 1 shown

[0027] A method for implementing Web rendering of 3DGS large-scale scenes suitable for consumer-grade devices includes the following steps:

[0028] LOD optimization: Based on the scene's geometric structure and texture information, the scene is divided into multiple sub-areas using a spatial segmentation algorithm. For each sub-area, the appropriate level of detail is determined based on its distance from the viewpoint and its importance, and a fast switching mechanism between LOD models is established.

[0029] Data compression: compress the geometric data, texture data and LOD model data in the 3DGS scene. The geometric data adopts a compression algorithm based on model simplification, the texture data uses an efficient texture compression format, and the LOD model data adopts an incremental encoding method;

[0030] Efficient decoding: On consumer devices, we design fast decoding algorithms for different data types and leverage the device's multi-core processor resources for parallel decoding.

[0031] Dynamic LOD calculation: A deep learning-based dynamic LOD calculation model is introduced. This model uses the scene's real-time state information as input to predict the optimal level of detail. A lightweight neural network structure is used, combined with model pruning and quantization optimization techniques, to design an incremental update mechanism.

[0032] Double-buffered coordinated rendering: Using double-buffering technology, the current frame is rendered in one buffer while the next frame's data is prepared in another buffer. The timing of buffer switching is precisely controlled. Distributed rendering technology based on edge computing is also introduced to dynamically adjust task allocation based on device performance and load.

[0033] Adaptive ray tracing: Dynamically adjusts the accuracy and sampling rate of ray tracing based on scene complexity and device performance, and combines ambient occlusion and shadow mapping technologies to enhance scene realism.

[0034] Cache management optimization: Establish a cache management mechanism based on least recently used (LRU) to cache commonly used rendering data and manage it in different levels according to data access frequency and importance.

[0035] In the LOD optimization step, the octree segmentation algorithm is used to divide the scene into sub-regions. The initial detail level is allocated by calculating the Euclidean distance between the sub-region and the viewpoint and combining the importance of key objects in the sub-region. Linear interpolation is used to transition when the LOD model is switched.

[0036] In the data compression step, geometric data compression uses a model simplification algorithm based on edge folding to compress geometric data to 30%-50% of the original data volume. Texture data is compressed according to the device support format, and the compression ratio is controlled at 1:4-1:8. LOD model data uses incremental encoding to reduce the data volume by 70%-80%.

[0037] In the efficient decoding step, independent threads are assigned to the decoding tasks of geometric data, texture data, and LOD model data, and decoding operations are performed in parallel, shortening the data decoding time to 1 / 3-1 / 2 of the original time.

[0038] In the dynamic LOD calculation step, a large amount of viewpoint position, viewing direction, movement speed and corresponding optimal LOD configuration data in different scenarios are collected to build a training dataset. A lightweight convolutional neural network consisting of several convolutional layers, pooling layers and fully connected layers is used as the calculation model. The model parameters are trained and optimized, and scene status information is obtained in real time and input into the model. The dynamic LOD calculation update time is controlled within 10-20 milliseconds.

[0039] In the double-buffered coordinated rendering step, buffer A and buffer B are created. Buffer A renders the current frame, and buffer B prepares the next frame data. The display is switched immediately after rendering is completed. The edge server dynamically adjusts the amount of rendering tasks assigned to the device based on the device performance indicators and current load conditions.

[0040] In the adaptive ray tracing step, the scene complexity is evaluated and the device performance parameters are obtained before rendering, the initial ray tracing sampling rate and accuracy are set, the device frame rate and resource usage are monitored in real time during the rendering process, the ray tracing parameters are dynamically adjusted, and the ambient occlusion and shadow mapping techniques are combined to enhance the scene realism.

[0041] In the cache management optimization step, a dedicated cache space is opened in the memory of consumer-grade devices and divided into areas for texture data cache, geometry data cache, and LOD model data cache. The LRU cache replacement strategy is adopted to divide the data into three levels: high-frequency important data, medium-frequency general data, and low-frequency secondary data. The cache data is cleaned and updated regularly to improve the cache hit rate.

[0042] Specific operation process

[0043] LOD optimization implementation: First, the octree segmentation algorithm is used to divide the 3DGS large-scale scene into multiple sub-areas. By calculating the Euclidean distance between each sub-area and the viewpoint, and combining the importance of the key objects (such as buildings, people, etc.) contained in the sub-area, an initial level of detail is assigned to each sub-area. For example, the sub-area 50 meters away from the viewpoint and without key objects is set to the lowest level of detail; the sub-area within 10 meters from the viewpoint or containing important key objects is set to the highest level of detail. Then, a switching threshold between LOD models is established. When the distance change between the viewpoint and the sub-area exceeds a certain threshold, the smooth switching of the LOD model is triggered. The switching process uses linear interpolation to transition and avoid visual mutations.

[0044] Data Compression Implementation: For geometric data, a model simplification algorithm based on edge folding is used. The folding cost of each edge is calculated, and edges are folded in ascending order of cost, removing redundant vertex and edge information. The geometric data is compressed to 30%-50% of the original data volume. For texture data, an appropriate compression method is selected based on the texture compression format supported by the device. For example, for mobile devices, the ETC format is preferred for compression, and the texture data compression ratio is controlled at 1:4-1:8. For LOD model data, the difference information between different LOD levels, such as vertex position changes and texture coordinate changes, is recorded and stored using incremental encoding. Compared with storing the complete LOD model data, the data volume can be reduced by 70%-80%.

[0045] Efficient decoding implementation: On consumer devices, the device's multi-core processors are leveraged to allocate independent threads for different types of data decoding tasks. For example, the geometry data decoding thread, texture data decoding thread, and LOD model data decoding thread execute in parallel. For geometry data decoding, the edge expansion operation is reversed according to the edge collapse algorithm used during compression to restore the original geometry model. For texture data decoding, the compressed texture data is restored to the original texture according to the decoding rules of the selected compression format. For LOD model data decoding, the base LOD model is updated based on the incremental encoding information to construct a complete LOD model. Through parallel decoding and optimized decoding algorithms, data decoding time is reduced to 1 / 3-1 / 2 of the original time.

[0046] Implementation of dynamic LOD calculation: First, collect a large amount of viewpoint position, viewing direction, movement speed and corresponding optimal LOD configuration data in different scenes to build a training data set. Use a lightweight convolutional neural network (CNN) as the dynamic LOD calculation model, and the network structure contains several convolution layers, pooling layers and fully connected layers. The model is trained with the training data set, and the model parameters are optimized so that it can accurately predict the optimal level of detail for each sub-area under different scene states. In the actual rendering process, the state information of the scene is obtained in real time and input into the trained model to quickly obtain the dynamic LOD results for each sub-area. When the scene state changes, the incremental update mechanism is used to perform local dynamic LOD calculation updates only on the affected sub-areas, and the update time is controlled within 10-20 milliseconds.

[0047] Double-buffered coordinated rendering implementation: Two render buffers, buffer A and buffer B, are created on consumer devices. Buffer A performs rendering operations for the current frame, including drawing scene geometry, applying textures, and calculating lighting. Simultaneously, buffer B prepares data for the next frame, updating scene data based on the latest scene state information, calculating dynamic LODs, and loading new textures. When the rendering operation in buffer A is completed, it is immediately displayed on the screen, and the rendering operation switches to buffer B for the next frame. Meanwhile, buffer A begins preparing data for the next frame. During double-buffered coordinated rendering, a connection is established with an edge server, sending some complex rendering tasks (such as ray tracing calculations and drawing large-scale geometry) to the edge server for processing. The edge server dynamically adjusts the amount of rendering tasks allocated to the device based on the device's performance indicators (such as the number of CPU cores and GPU memory size) and current load, ensuring load balancing between the device and the edge server, thereby improving overall rendering efficiency.

[0048] Adaptive ray tracing implementation: Before rendering begins, the complexity of the scene is evaluated, and the complexity level of the scene is determined by calculating indicators such as the number of geometric objects, the number of polygons, and the complexity of the texture in the scene. At the same time, the performance parameters of consumer-grade devices (such as CPU frequency, GPU model, etc.) are obtained, and the initial ray tracing sampling rate and accuracy are set according to the scene complexity and device performance. For example, for simple scenes and low device performance, the ray tracing sampling rate is set to 8, and a simplified ray tracing algorithm is used; for complex scenes and good device performance, the ray tracing sampling rate is set to 32, and a more accurate ray tracing algorithm is used. During the rendering process, the frame rate and resource usage of the device are monitored in real time. When the frame rate is lower than the set threshold or the resource usage is too high, the ray tracing sampling rate and accuracy are reduced; when the frame rate is higher than the set threshold and there are surplus resources, the ray tracing sampling rate and accuracy are increased. At the same time, ambient occlusion and shadow mapping technologies are combined to enhance the realism of the scene.

[0049] Cache management optimization implementation: Dedicated cache space is allocated in the memory of consumer devices to store data commonly used during the rendering process. The cache space is divided into multiple cache areas, respectively used to store texture data cache, geometry data cache, and LOD model data cache. A least recently used (LRU) cache replacement strategy is adopted. When cache space is insufficient, the least recently used data is found and removed from the cache to make room for new data. At the same time, data is divided into three levels based on data access frequency and importance: high-frequency important data, medium-frequency general data, and low-frequency secondary data. High-frequency important data is stored in the high-speed cache area to improve its access speed; low-frequency secondary data has its cache priority appropriately lowered to ensure efficient use of cache space. Cached data is regularly cleaned and updated, deleting unused data, improving cache hit rates, and reducing data loading time.

[0050] Finally, it should be noted that the above is only a preferred embodiment of the present invention and is not intended to limit the present invention. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art can still modify the technical solutions described in the aforementioned embodiments or make equivalent substitutions for some of the technical features therein. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. A method for implementing Web rendering of 3DGS large-scale scenes suitable for consumer-grade devices, characterized in that: The following steps are involved: LOD optimization: Based on the scene's geometric structure and texture information, the scene is divided into multiple sub-areas using a spatial segmentation algorithm. For each sub-area, the appropriate level of detail is determined based on its distance from the viewpoint and its importance, and a fast switching mechanism between LOD models is established. Data compression: compress the geometric data, texture data and LOD model data in the 3DGS scene. The geometric data adopts a compression algorithm based on model simplification, the texture data uses an efficient texture compression format, and the LOD model data adopts an incremental encoding method; Efficient decoding: On consumer devices, we design fast decoding algorithms for different data types and leverage the device's multi-core processor resources for parallel decoding. Dynamic LOD calculation: A deep learning-based dynamic LOD calculation model is introduced. This model uses the scene's real-time state information as input to predict the optimal level of detail. A lightweight neural network structure is used, combined with model pruning and quantization optimization techniques, to design an incremental update mechanism. Double-buffered coordinated rendering: Using double-buffering technology, the current frame is rendered in one buffer while the next frame's data is prepared in another buffer. The timing of buffer switching is precisely controlled. Distributed rendering technology based on edge computing is also introduced to dynamically adjust task allocation based on device performance and load. Adaptive ray tracing: Dynamically adjusts the accuracy and sampling rate of ray tracing based on scene complexity and device performance, and combines ambient occlusion and shadow mapping technologies to enhance scene realism. Cache management optimization: Establish a cache management mechanism based on least recently used data to cache commonly used rendering data, and manage it in different levels according to data access frequency and importance.

2. The method for implementing Web rendering of 3DGS large-scale scenes suitable for consumer-grade devices according to claim 1, characterized in that: In the LOD optimization step, the scene is divided into sub-regions using an octree segmentation algorithm. The initial level of detail is allocated by calculating the Euclidean distance between the sub-region and the viewpoint and combining the importance of key objects in the sub-region. Linear interpolation is used for transition when the LOD model is switched.

3. The method for implementing Web rendering of 3DGS large-scale scenes suitable for consumer-grade devices according to claim 1, characterized in that: In the data compression step, the geometric data compression adopts a model simplification algorithm based on edge folding to compress the geometric data to 30%-50% of the original data volume. The texture data selects the compression method according to the device support format, and the compression ratio is controlled at 1:4-1:

8. The LOD model data adopts incremental encoding to reduce the data volume by 70%-80%.

4. The method for implementing Web rendering of 3DGS large-scale scenes suitable for consumer-grade devices according to claim 1, characterized in that: In the efficient decoding step, independent threads are allocated to the decoding tasks of geometric data, texture data and LOD model data, and decoding operations are performed in parallel, shortening the data decoding time to 1 / 3-1 / 2 of the original time.

5. The method for implementing Web rendering of 3DGS large-scale scenes suitable for consumer-grade devices according to claim 1, characterized in that: In the dynamic LOD calculation step, a large amount of viewpoint position, viewing direction, movement speed and corresponding optimal LOD configuration data in different scenes are collected to construct a training data set. A lightweight convolutional neural network containing several convolutional layers, pooling layers and fully connected layers is used as the calculation model. The model parameters are trained and optimized, and the scene status information is obtained in real time and input into the model. The dynamic LOD calculation update time is controlled within 10-20 milliseconds.

6. The method for implementing Web rendering of 3DGS large-scale scenes suitable for consumer-grade devices according to claim 1, characterized in that: In the double-buffered coordinated rendering step, buffer A and buffer B are created. Buffer A renders the current frame, and buffer B prepares the next frame data. The display is switched immediately after the rendering is completed. The edge server dynamically adjusts the amount of rendering tasks assigned to the device based on the device performance indicators and current load conditions.

7. The method for implementing Web rendering of 3DGS large-scale scenes suitable for consumer-grade devices according to claim 1, characterized in that: In the adaptive ray tracing step, the scene complexity is evaluated and the device performance parameters are obtained before rendering, the initial ray tracing sampling rate and accuracy are set, the device frame rate and resource usage are monitored in real time during the rendering process, the ray tracing parameters are dynamically adjusted, and the ambient occlusion and shadow mapping technologies are combined to enhance the scene realism.

8. The method for implementing Web rendering of 3DGS large-scale scenes suitable for consumer-grade devices according to claim 1, characterized in that: In the cache management optimization step, a dedicated cache space is opened in the memory of the consumer-grade device and divided into areas for texture data cache, geometry data cache, and LOD model data cache. An LRU cache replacement strategy is adopted to divide the data into three levels: high-frequency important data, medium-frequency general data, and low-frequency secondary data. The cache data is cleaned and updated regularly to improve the cache hit rate.

Images