Three-dimensional model reconstruction method and device based on gaussian sputtering model
By combining motion recovery structure technology and Gaussian sputtering model, and utilizing monocular depth model and multi-view stereo depth estimation algorithm, the problems of excessive smoothing and artifacts in sparse view 3D reconstruction are solved by introducing depth consistency and dynamic gradient control loss function, thus achieving high-quality 3D model reconstruction.
Patent Information
- Application Number
- CN202511973042.8
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2025-12-25
- Publication Date
- 2026-05-15
- Estimated Expiration
- 2045-12-25
AI Technical Summary
Existing 3D reconstruction methods for scene images based on sparse perspectives suffer from problems such as generating overly smooth surfaces, losing details, floating artifacts, and holes, as well as conflicts between geometric gradients and appearance gradients.
By employing motion recovery structure technology and Gaussian sputtering model, combined with monocular depth model, multi-view stereo depth estimation algorithm, and depth consistency loss function, color loss function, gradient smoothing loss function, and dynamic gradient control loss function, the initialization parameter information is updated to generate a reliable 3D model.
It improves the reliability and robustness of 3D models under sparse viewpoint conditions, resolves the contradiction between geometric constraints and appearance constraints in the gradient optimization direction, and generates high-quality 3D models.
Smart Images

Figure CN121392161B_ABST
Abstract
Description
Technical Field
[0001] This application relates to the field of three-dimensional reconstruction technology, and in particular to a method and apparatus for three-dimensional model reconstruction based on a Gaussian sputtering model. Background Technology
[0002] 3D reconstruction technology can reconstruct the 3D structure and appearance of the real world from images or laser point cloud data, providing a foundation for applications in multiple fields such as virtual reality, augmented reality, cultural heritage protection, and urban planning. Currently, vision-based 3D reconstruction methods have attracted widespread attention and become a research hotspot due to their low cost and ease of operation. Among them, reconstruction methods based on SfM (Structure from Motion) and multi-view stereo (MVS) offer high geometric accuracy and have been applied in fields such as archaeology and surveying. However, these methods have high requirements for image quality and lighting consistency, and are computationally time-consuming. To accelerate image-based 3D modeling, 3D reconstruction methods based on neural radiation fields have been developed, implicitly representing the light and color distribution of a scene. However, their training efficiency is low, and the reconstruction results are difficult to edit. To further accelerate training speed, the 3D Gaussian sputtering method, which uses Gaussian ellipsoids to represent scenes, has received widespread attention in recent years. Its training speed can reach the minute level, and it supports editing the reconstruction results. However, under sparse viewpoint conditions, the 3D Gaussian sputtering method tends to generate overly smooth surfaces, losing sculpted textures and vegetation details. Furthermore, under sparse viewpoints, the 3D Gaussian sputtering method suffers from floating artifacts or voids due to insufficient multi-view constraints. To address this, depth regularization methods based on prior depth maps can improve the modeling performance of 3D Gaussian sputtering under sparse viewpoints, but they suffer from conflicts between geometric gradients and appearance gradients. Therefore, how to establish a reliable 3D scene model based on scene images from sparse viewpoints is a problem that needs to be solved. Summary of the Invention
[0003] Therefore, it is necessary to provide a method and apparatus for 3D model reconstruction based on Gaussian sputtering model that can reconstruct a reliable 3D scene model from scene images with sparse perspectives, in order to address the above-mentioned technical problems.
[0004] In a first aspect, this application provides a method for reconstructing a three-dimensional model based on a Gaussian sputtering model, the method comprising:
[0005] Acquire scene images of the target scene from a sparse perspective, and determine the initialization parameter information of the Gaussian ellipsoid based on the scene images using motion reconstruction structure technology and Gaussian sputtering model;
[0006] Based on the scene image, the prior depth, color loss function, and gradient smoothing loss function are determined using a monocular depth model.
[0007] A reference depth is determined based on the scene image using a multi-view stereo depth estimation algorithm.
[0008] Determine the depth consistency loss function based on the prior depth and the reference depth;
[0009] Using the Gaussian sputtering model, based on the depth consistency loss function, the color loss function, and the gradient smoothing loss function, the initialization parameter information is updated, the updated parameter information is determined, and a 3D model of the target scene is generated according to the updated parameter information.
[0010] In one embodiment, determining a depth consistency loss function based on the prior depth and the reference depth includes:
[0011] Determine the depth confidence weight based on the depth difference data between the prior depth and the reference depth;
[0012] Determine the rendering depth difference between the prior depth and the reference depth, and determine the depth consistency loss function based on the rendering depth difference and the depth confidence weight.
[0013] In one embodiment, the initialization parameter information is updated using a Gaussian sputtering model based on the depth consistency loss function, the color loss function, and the gradient smoothing loss function. Determining the updated parameter information includes:
[0014] Using the Gaussian sputtering model, a dynamic gradient control loss function is determined based on the depth consistency loss function and the color loss function;
[0015] Based on the depth consistency loss function, the color loss function, the gradient smoothing loss function, and the dynamic gradient control loss function, the initialization parameter information is updated, and the updated parameter information is determined.
[0016] In one embodiment, the initialization parameter information is updated using a Gaussian sputtering model based on the depth consistency loss function, the color loss function, the gradient smoothing loss function, and the dynamic gradient control loss function. Determining the updated parameter information includes:
[0017] Set initial values for dynamic weight parameters; the initial values for dynamic weight parameters include initial values for the weights of the depth consistency loss function, the color loss function, the gradient smoothing loss function, and the dynamic gradient control loss function.
[0018] The initial weights of the deep consistency loss function and the dynamic gradient control loss function are dynamically adjusted using a hyperparameter optimization algorithm to determine the adjusted weights of the deep consistency loss function and the dynamic gradient control loss function.
[0019] The weights of the depth consistency loss function, the dynamic gradient control loss function, the initial weights of the color loss function, and the initial weights of the gradient smoothing loss function are used as the weights of the loss function. The total loss function is determined by the Gaussian sputtering model based on the weights of the loss function, the depth consistency loss function, the color loss function, the gradient smoothing loss function, and the dynamic gradient control loss function.
[0020] Based on the total loss function, the initialization parameter information is updated, and the updated parameter information is determined.
[0021] In one embodiment, a hyperparameter optimization algorithm is used to dynamically adjust the initial weights of the depth consistency loss function and the dynamic gradient control loss function to determine the adjusted weights of the depth consistency loss function and the dynamic gradient control loss function, including:
[0022] By using a hyperparameter optimization algorithm, the initial weight values of the deep consistency loss function are dynamically adjusted based on a preset number of iterations, thereby determining the weight adjustment values of the deep consistency loss function.
[0023] The geometric gradient is determined based on the depth consistency loss function, and the appearance gradient is determined based on the color loss function.
[0024] A dynamic gradient controller is used to detect conflicts between the geometric gradient and the appearance gradient in order to determine the gradient angle between the geometric gradient and the appearance gradient.
[0025] The weight parameters of the dynamic gradient control loss function are determined based on the gradient angle, and the initial weight values of the dynamic gradient control loss function are adjusted according to the weight parameters to determine the weight adjustment values of the dynamic gradient control loss function.
[0026] In one embodiment, updating the initialization parameter information based on the total loss function includes:
[0027] Based on the total loss function, the initialization parameter information is updated, and during the updating of the initialization parameter information, the rate of change of the total loss function is determined in real time.
[0028] If the rate of change of loss is less than the rate of change threshold, or the number of iterations for updating the initialization parameter information is greater than the number of iterations threshold, then the updating of the initialization parameter information is stopped, and the updated initialization parameter information is used as the updated parameter information.
[0029] In one embodiment, using motion recovery structure technology and a Gaussian sputtering model, the initialization parameter information of the Gaussian ellipsoid is determined based on the scene image, including:
[0030] Using motion reconstruction structure technology, the three-dimensional scene structure is determined based on the scene image, and point cloud coordinates are generated based on the three-dimensional scene structure;
[0031] The initialization parameters of the Gaussian ellipsoid are determined based on the point cloud coordinates using the Gaussian sputtering model.
[0032] Secondly, this application also provides a three-dimensional model reconstruction device based on a Gaussian sputtering model, the device comprising:
[0033] The initialization information determination module is used to acquire scene images of the target scene from a sparse perspective, and determine the initialization parameter information of the Gaussian ellipsoid based on the scene images using motion reconstruction structure technology and Gaussian sputtering model.
[0034] The prior depth determination module is used to determine the prior depth, color loss function, and gradient smoothing loss function based on the scene image using a monocular depth model.
[0035] The reference depth determination module is used to determine the reference depth based on the scene image using a multi-view stereo depth estimation algorithm.
[0036] A depth loss function determination module is used to determine a depth consistency loss function based on the prior depth and the reference depth.
[0037] The 3D model building module is used to update the initialization parameter information, determine the updated parameter information, and generate a 3D model of the target scene based on the updated parameter information, using a Gaussian sputtering model and based on the depth consistency loss function, the color loss function, and the gradient smoothing loss function.
[0038] Thirdly, this application also provides a computer device, the computer device including a memory and a processor, the memory storing a computer program, and the processor executing the computer program to perform the following steps:
[0039] Acquire scene images of the target scene from a sparse perspective, and determine the initialization parameter information of the Gaussian ellipsoid based on the scene images using motion reconstruction structure technology and Gaussian sputtering model;
[0040] Based on the scene image, the prior depth, color loss function, and gradient smoothing loss function are determined using a monocular depth model.
[0041] A reference depth is determined based on the scene image using a multi-view stereo depth estimation algorithm.
[0042] Determine the depth consistency loss function based on the prior depth and the reference depth;
[0043] Using the Gaussian sputtering model, based on the depth consistency loss function, the color loss function, and the gradient smoothing loss function, the initialization parameter information is updated, the updated parameter information is determined, and a 3D model of the target scene is generated according to the updated parameter information.
[0044] Fourthly, this application also provides a computer-readable storage medium having a computer program stored thereon, the computer program performing the following steps when executed by a processor:
[0045] Acquire scene images of the target scene from a sparse perspective, and determine the initialization parameter information of the Gaussian ellipsoid based on the scene images using structure of motion reconstruction and Gaussian sputtering model.
[0046] Based on the scene image, the prior depth, color loss function, and gradient smoothing loss function are determined using a monocular depth model.
[0047] A reference depth is determined based on the scene image using a multi-view stereo depth estimation algorithm.
[0048] Determine the depth consistency loss function based on the prior depth and the reference depth;
[0049] Using the Gaussian sputtering model, based on the depth consistency loss function, the color loss function, and the gradient smoothing loss function, the initialization parameter information is updated, the updated parameter information is determined, and a 3D model of the target scene is generated according to the updated parameter information.
[0050] The above three-dimensional model reconstruction method and device based on the Gaussian sputtering model obtain scene images of sparse viewpoints of the target scene, and determine the initialization parameter information of the Gaussian ellipsoid according to the scene images through the structure from motion technology and the Gaussian sputtering model; determine the prior depth, color loss function, and gradient smoothing loss function according to the scene images through the monocular depth model; determine the reference depth according to the scene images through the multi-view stereo depth estimation algorithm; determine the depth consistency loss function according to the prior depth and the reference depth; update the initialization parameter information based on the depth consistency loss function, color loss function, and gradient smoothing loss function through the Gaussian sputtering model, determine the updated parameter information, and generate the three-dimensional model of the target scene according to the updated parameter information. It solves the problem that it is difficult for the Gaussian sputtering model to construct a reliable three-dimensional model of the scene based on the scene images of sparse viewpoints. In the above solution, the depth consistency loss function is determined according to the prior depth determined by the monocular depth model according to the scene images of sparse viewpoints and the reference depth determined by the multi-view stereo depth estimation algorithm, and at the same time, a dynamic gradient control loss function is introduced to solve the contradiction between geometric constraints and appearance constraints in the gradient optimization direction when constructing the three-dimensional model of the target scene. The initialization parameter information is updated according to the depth consistency loss function, dynamic gradient control loss function, color loss function, and gradient smoothing loss function, so as to generate the three-dimensional model of the target scene through the Gaussian sputtering model according to the updated parameter information. It can improve the reliability and robustness of the three-dimensional model through the depth prior constraint information and the dynamic gradient control method. BRIEF DESCRIPTION OF THE DRAWINGS
[0051] Figure 1 It is an application environment diagram of the three-dimensional model reconstruction method based on the Gaussian sputtering model in an embodiment;
[0052] Figure 2 It is a flowchart of the three-dimensional model reconstruction method based on the Gaussian sputtering model in an embodiment;
[0053] Figure 3 It is a flowchart of constructing the depth consistency loss function in an embodiment;
[0054] Figure 4 It is a flowchart of the three-dimensional model reconstruction method based on the Gaussian sputtering model in another embodiment;
[0055] Figure 5 It is a schematic diagram of the method for adjusting the gradient control intensity in an embodiment;
[0056] Figure 6 It is a structural block diagram of the three-dimensional model reconstruction device based on the Gaussian sputtering model in an embodiment;
[0057] Figure 7This is an internal structural diagram of a computer device in one embodiment. Detailed Implementation
[0058] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.
[0059] The 3D model reconstruction method based on the Gaussian sputtering model provided in this application can be applied to, for example... Figure 1 In the application environment shown, terminal 102 communicates with server 104 via a network. A data storage system can store the data that server 104 needs to process. The data storage system can be integrated onto server 104 or placed in the cloud or on another network server. Server 104 acquires scene images of the target scene from a sparse perspective, and determines the initialization parameters of the Gaussian ellipsoid based on the scene images using motion reconstruction structure technology and a Gaussian sputtering model; it also determines the prior depth, color loss function, and gradient smoothing loss function based on the scene images using a monocular depth model; it determines a reference depth based on the scene images using a multi-view stereo depth estimation algorithm; it determines a depth consistency loss function based on the prior depth and the reference depth; it updates the initialization parameters based on the depth consistency loss function, the color loss function, and the gradient smoothing loss function using the Gaussian sputtering model, determines the updated parameters, and generates a 3D model of the target scene based on the updated parameters. The 3D model of the target scene is then sent to terminal 102 via the communication network. The terminal 102 can be, but is not limited to, various personal computers, laptops, smartphones, tablets, IoT devices, and portable wearable devices. IoT devices can include smart speakers, smart TVs, smart air conditioners, and smart in-vehicle systems. Portable wearable devices can include smartwatches, smart bracelets, and head-mounted devices. The server 104 can be implemented using a standalone server or a server cluster consisting of multiple servers.
[0060] In one embodiment, such as Figure 2 As shown, a 3D model reconstruction method based on a Gaussian sputtering model is provided. This embodiment illustrates the application of this method to a terminal. It is understood that this method can also be applied to a server, and further to a system including both a terminal and a server, and implemented through interaction between the terminal and the server. In this embodiment, the method includes the following steps:
[0061] S210. Obtain scene images of the target scene from a sparse perspective, and determine the initialization parameter information of the Gaussian ellipsoid based on the scene images using motion reconstruction structure technology and Gaussian sputtering model.
[0062] The target scene refers to the scene requiring 3D modeling. Sparse viewpoint scene images refer to a small number of images of the target scene taken from different angles, exhibiting sparse viewpoint distribution, meaning the shooting angles are not densely packed. Structure-on-motion (SOP) technology is a technique that recovers camera parameters and 3D scene structure by analyzing image sequences. Its core processes include feature point extraction, image matching, camera pose estimation, and 3D point cloud reconstruction. It should be noted that Gaussian sputtering technology relies on the point cloud generated by SOP as initial data. SOP recovers the 3D scene structure and generates point cloud coordinates by analyzing multi-view image sequences; while Gaussian sputtering converts the point cloud into a Gaussian distribution, achieving efficient rendering through parameter training and optimization.
[0063] Specifically, the process involves acquiring scene images of the target scene from sparse perspectives, captured by an image acquisition device, forming a scene image set. I M Let M represent the scene image with the sparse viewpoint, where M ≥ 1 and M is an integer. The initialization parameters of the Gaussian ellipsoid are determined based on the scene image using structure-of-motion reconstruction (SOR) and a Gaussian sputtering model.
[0064] For example, by using motion-reconstruction structure technology and a Gaussian sputtering model, the initialization parameter information of the Gaussian ellipsoid is determined based on the scene image, including:
[0065] Using motion reconstruction technology, the 3D scene structure is determined based on the scene image, and point cloud coordinates are generated based on the 3D scene structure; using a Gaussian sputtering model, the initialization parameter information of the Gaussian ellipsoid is determined based on the point cloud coordinates.
[0066] The initialization parameters can include the scale parameters, rotation parameters, spherical harmonics, and opacity of the Gaussian ellipsoid. The scale parameters of the Gaussian ellipsoid primarily include the scaling factor, used to adjust the proportional relationship between latitude and longitude and planar coordinates. The scaling factor is a core scale parameter in the Gaussian-Krüger projection, determining the corresponding length of a unit length on the Earth's surface in the planar coordinate system. By adjusting the scaling factor, the shape and scale of the projected map can be optimized. The Gaussian-Krüger projection is a conformal transverse cylindrical projection used to convert the Earth's ellipsoid into a planar map, widely used in large-scale topographic mapping. The rotation parameters of the Gaussian ellipsoid usually refer to the rotation matrix, used to describe the rotation of the ellipsoid in three-dimensional space. In the context of the Gaussian distribution, the rotation parameters are usually represented by quaternions, which, after being converted into a rotation matrix, together with the scale matrix, determine the covariance matrix. The spherical harmonics are parameters of the Gaussian ellipsoid used in computer graphics to represent view-dependent color or illumination distribution. In the Gaussian sputtering model, each Gaussian ellipsoid controls its color variation with viewing angle through spherical harmonic coefficients. Low-order coefficients can smoothly fit complex lighting or material reflection effects. The opacity of the Gaussian ellipsoid is a parameter that controls the degree to which the Gaussian ellipsoid contributes to the final rendering result, and its value typically ranges from 0 to 1. When the opacity is 0, the Gaussian ellipsoid is completely transparent; when the opacity is 1, the Gaussian ellipsoid is completely opaque.
[0067] Specifically, using motion reconstruction technology, the 3D scene structure of the target scene is reconstructed from the scene image, and the point cloud coordinates of the 3D scene structure are initialized. This allows the geometric position of the initialized Gaussian ellipsoid to be determined using a Gaussian sputtering model based on the initialized point cloud coordinates. The initialization parameters of the Gaussian ellipsoid are then determined using the Gaussian sputtering model and the point cloud coordinates. These initialization parameters include the scale parameter, rotation parameter, spherical harmonic coefficients, and opacity of the Gaussian ellipsoid. For example, the scale parameter in the initialization parameter information... Set as Rotation parameters Set as spherical harmonic coefficient Set as Opacity Set as .
[0068] The above scheme initializes the point cloud coordinates of the target scene based on motion recovery structure technology, and determines the initialization parameter information of the Gaussian ellipsoid based on the point cloud coordinates using a Gaussian sputtering model, which can improve the efficiency of determining the initialization parameter information.
[0069] S220. Using a monocular depth model, determine the prior depth, color loss function, and gradient smoothing loss function based on the scene image.
[0070] Prior depth, in the field of deep learning, is a method that utilizes implicit prior knowledge within the network structure to handle tasks such as image generation and restoration, generating or restoring signals without external data training. Image prior depth is a concept in deep learning used to describe the statistical features and structural patterns of images automatically extracted by neural networks during training. The color loss function is a loss function used in computer vision and deep learning to measure color differences between images, primarily used to control the color consistency between the generated and target images. The gradient smoothing loss function is an improvement on traditional loss functions that suffer from abrupt gradient changes in specific regions. It smooths gradient changes to make them more stable, thereby optimizing the model training process. A monocular depth model can be the Depth Anything V2 model. Prior depth, color loss function, and gradient smoothing loss function can all be obtained by processing scene images using a monocular depth model.
[0071] Specifically, the scene image is input into the monocular depth model, and the scene image is processed according to the monocular depth model to determine the prior depth of the scene image, the color loss function, and the gradient smoothing loss function.
[0072] S230. Determine the reference depth based on the scene image using a multi-view stereo depth estimation algorithm.
[0073] Among them, the Multi-View Stereo Depth Estimation (MVS) algorithm estimates the depth information of each point in a scene by analyzing multiple images with overlapping fields of view and utilizing the principle of geometric consistency. Its core lies in extracting consistent 3D information from images from different viewpoints through multi-view matching and optimization algorithms. The reference depth in the MVS depth estimation algorithm typically refers to the depth information of a reference image, that is, the depth values of each point in the scene calculated through multi-view image matching.
[0074] For example, in the MVS algorithm, one or more keyframe images are typically selected as reference images, while images from other perspectives are used as auxiliary images. Algorithms such as SIFT (Scale-Invariant Feature Transform) are used to extract feature points from the auxiliary images and match them with feature points in the reference images. Through multi-view geometric constraints and optimization algorithms, the depth value of the corresponding point in the scene, i.e., the reference depth, is calculated from the coordinates of the matched points.
[0075] S240. Determine the depth consistency loss function based on the prior depth and the reference depth.
[0076] For example, the method for determining the depth consistency loss function includes: determining a depth confidence weight based on the depth difference data between the prior depth and the reference depth; determining the rendering depth difference between the prior depth and the reference depth; and determining the depth consistency loss function based on the rendering depth difference and the depth confidence weight.
[0077] Among them, rendering depth difference usually refers to the difference in how different rendering methods or technologies process the depth information of objects in a scene in computer graphics rendering.
[0078] Specifically, the flowchart for constructing the deep consistency loss function is as follows: Figure 3 As shown, scene images with sparse viewpoints of the target scene are acquired, and the initialization parameters of the Gaussian ellipsoid are determined based on the scene images using structure-of-motion reconstruction (SOR) and a Gaussian sputtering model. A prior depth is determined based on the scene images using a monocular depth model. A reference depth is determined based on the scene images using a multi-view stereo depth estimation algorithm. The difference between the prior depth and the reference depth is used as depth difference data. Depth confidence weights are determined based on the depth difference data and the adaptive uncertainty factor of the monocular depth model. The formula for calculating the depth confidence weight is shown in formula (1):
[0079] (1).
[0080] in, As an adaptive uncertainty factor, Use 0.1 in areas with rich texture and 0.5 in areas with weak texture; Let be the gradient norm of the scene image at pixel i.
[0081] Determine the rendering depth difference between the prior depth and the reference depth. The rendering depth difference includes... and The formula for calculating the deep consistency loss function is shown in formula (2):
[0082] (2).
[0083] in, It is the rendering depth of the i-th pixel in the scene image; It refers to the depth value of the i-th pixel in the scene image determined by the monocular depth model, that is, the prior depth of the i-th pixel; It refers to the depth value of the i-th pixel in the scene image determined by the multi-view stereo depth estimation algorithm, that is, the reference depth of the i-th pixel. It is the balance coefficient, in When greater than 0.3 The value is 0.1. When less than or equal to 0.3 The value is 1.0.
[0084] The above scheme combines the global structural robustness of the monocular depth estimation model with the local geometric high accuracy of the multi-view stereo depth estimation algorithm. It can dynamically adjust the depth confidence weight according to the texture complexity of the scene, and apply strong depth difference sensitive constraints in texture-rich areas to suppress reconstruction jitter; in weak texture areas, the depth difference tolerates weak constraints to prevent overfitting.
[0085] S250. Using the Gaussian sputtering model, based on the depth consistency loss function, color loss function, and gradient smoothing loss function, the initialization parameter information is updated, the updated parameter information is determined, and a 3D model of the target scene is generated based on the updated parameter information.
[0086] Understandably, the updated parameter information includes the updated scale parameters, updated rotation parameters, updated spherical harmonic coefficients, and updated opacity. A 3D model refers to the three-dimensional structural model of the target scene.
[0087] Specifically, using a Gaussian sputtering model, based on depth consistency loss, color loss, and gradient smoothing loss functions, and employing the backpropagation algorithm, the initialization parameters are updated. These updated parameters are then used as the updated parameters to generate a 3D model of the target scene. The backpropagation algorithm is a core algorithm in neural network training, using a chain rule to calculate the error gradient layer by layer, thereby optimizing the network parameters.
[0088] The aforementioned 3D model reconstruction method based on the Gaussian sputtering model acquires scene images of the target scene from sparse perspectives. Using structure-of-motion (SOR) reconstruction and the Gaussian sputtering model, initialization parameters of the Gaussian ellipsoid are determined from the scene images. A monocular depth model is used to determine prior depth, color loss function, and gradient smoothing loss function based on the scene images. A multi-view stereo depth estimation algorithm is used to determine a reference depth based on the scene images. A depth consistency loss function is determined based on the prior and reference depths. The Gaussian sputtering model is then used to update the initialization parameters based on the depth consistency loss function, color loss function, and gradient smoothing loss function, determining the updated parameters and generating a 3D model of the target scene based on these updated parameters. This method solves the problem that the Gaussian sputtering model struggles to construct reliable 3D scene models from scene images with sparse perspectives. The above scheme uses a monocular depth model to determine the prior depth based on the scene image from a sparse viewpoint and the reference depth determined by the multi-view stereo depth estimation algorithm. It then determines a depth consistency loss function and introduces a dynamic gradient control loss function to resolve the contradiction between geometric and appearance constraints in the gradient optimization direction when constructing the 3D model of the target scene. Based on the depth consistency loss function, dynamic gradient control loss function, color loss function, and gradient smoothing loss function, the initial parameters are updated. A Gaussian sputtering model is then used to generate the 3D model of the target scene based on the updated parameters. This approach improves the reliability and robustness of the 3D model through depth prior constraints and dynamic gradient control.
[0089] In one embodiment, such as Figure 4 As shown, using the Gaussian sputtering model, based on the depth consistency loss function, color loss function, and gradient smoothing loss function, the initialization parameter information is updated, and the updated parameter information is determined, including:
[0090] S310. Using the Gaussian sputtering model, determine the dynamic gradient control loss function based on the depth consistency loss function and the color loss function.
[0091] For example, the formula for calculating the dynamic gradient control loss function is shown in formula (3):
[0092] (3).
[0093] in, Let be the dynamic gradient control loss function corresponding to the Sth Gaussian parameter; This represents the set of all parameters of the Gaussian ellipsoid. Is the depth consistency loss function with respect to the th Geometric gradient of Gaussian parameters, It is the color loss function with respect to the first... The appearance gradient of a Gaussian parameter.
[0094] S320. Based on the depth consistency loss function, color loss function, gradient smoothing loss function, and dynamic gradient control loss function, update the initialization parameter information and determine the updated parameter information.
[0095] Specifically, the Gaussian sputtering model determines the total loss function based on the depth consistency loss function, color loss function, gradient smoothing loss function, and dynamic gradient control loss function. It then updates the initialization parameter information in reverse based on the total loss function to determine the updated parameter information.
[0096] The above scheme constructs a dynamic gradient control loss function based on the depth consistency loss function and the color loss function. Based on the depth consistency loss function, the color loss function, the gradient smoothing loss function, and the dynamic gradient control loss function, it updates the initial parameter information and determines the updated parameter information. This can resolve the contradiction between geometric constraints and appearance constraints in the gradient optimization direction and enhance the robustness of the 3D model built based on sparse viewpoint images.
[0097] In one embodiment, using a Gaussian sputtering model, based on depth consistency loss function, color loss function, gradient smoothing loss function, and dynamic gradient control loss function, the initialization parameter information is updated, and the updated parameter information is determined, including:
[0098] Initial values for dynamic weight parameters are set. These initial values include the initial weights of the depth consistency loss function, color loss function, gradient smoothing loss function, and dynamic gradient control loss function. The initial weights of the depth consistency loss function and dynamic gradient control loss function are dynamically adjusted using a hyperparameter optimization algorithm to determine their adjusted values. These adjusted values are then used as the weights for the loss functions. A Gaussian sputtering model is used to determine the total loss function based on the loss function weights, depth consistency loss function, color loss function, gradient smoothing loss function, and dynamic gradient control loss function. Finally, the initial parameter information is updated based on the total loss function to determine the updated parameter information.
[0099] For example, the initial value of the dynamic weight parameter is: , , ,in These are the initial weights for the gradient smoothing loss function. These are the initial weights for the deep consistency loss function. The initial weights for the dynamic gradient control loss function are 1, and the initial weights for the color loss function are 1.
[0100] For example, by using a hyperparameter optimization algorithm, the initial weights of the depth consistency loss function and the initial weights of the dynamic gradient control loss function are dynamically adjusted to determine the adjusted weights of the depth consistency loss function and the dynamic gradient control loss function, including:
[0101] Using a hyperparameter optimization algorithm, the initial weights of the depth consistency loss function are dynamically adjusted based on a preset number of iterations to determine the adjusted weights of the depth consistency loss function. The geometric gradient is determined based on the depth consistency loss function, and the appearance gradient is determined based on the color loss function. A dynamic gradient controller is used to detect conflicts between the geometric and appearance gradients to determine their gradient angle. The weight parameters of the dynamic gradient control loss function are determined based on the gradient angle, and the initial weights of the dynamic gradient control loss function are adjusted according to these weight parameters to determine the adjusted weights of the dynamic gradient control loss function.
[0102] Among them, the dynamic gradient controller is the core component of the magnetic resonance imaging system used to dynamically adjust the gradient magnetic field strength. Its core function is to adjust the gradient field parameters in real time according to imaging requirements to ensure imaging quality and efficiency.
[0103] For example, the weight adjustment method of the dynamic gradient control loss function is as follows: Figure 5 As shown, through a hyperparameter optimization algorithm, the gradient data corresponding to the initialization parameter information determined by the depth consistency loss function is used as the geometric gradient, and the gradient data corresponding to the initialization parameter information determined by the color loss function is used as the appearance gradient. A dynamic gradient controller performs gradient conflict detection on the geometric and appearance gradients to determine the gradient angle between them. Specifically, the geometric and appearance gradients are input into the dynamic gradient controller, which then determines the gradient angle between them. If the cosine function of the gradient angle is less than -0.8, the gradient control strength is increased, i.e., the weight of the dynamic gradient control loss function is increased; if the cosine function of the gradient angle is greater than -0.2, the gradient control strength is decreased, i.e., the weight of the dynamic gradient control loss function is decreased; if the cosine function of the gradient angle is greater than or equal to -0.8 and less than or equal to -0.2, the gradient control strength remains unchanged.
[0104] For example, the weight adjustment method of the deep consistency loss function includes: determining the weight adjustment value of the deep consistency loss function based on the number of algorithm iterations and a preset iteration threshold through a hyperparameter optimization algorithm. The calculation formula of the weight adjustment value of the deep consistency loss function is shown in formula (4):
[0105] (4).
[0106] in, t is the weight adjustment value of the deep consistency loss function; e is a natural number; t is the number of algorithm iterations; T is the preset iteration threshold.
[0107] The formula for calculating the total loss function is shown in formula (5):
[0108] (5).
[0109] in, This is the total loss function; Here is the color loss function; The depth consistency loss function; The gradient smoothing loss function; This is the dynamic gradient control loss function.
[0110] The above scheme provides a way to dynamically adjust the weights of the depth consistency loss function and the dynamic gradient control loss function. The total loss function is determined by the weights of the adjusted loss function, which can solve the problem of geometric ambiguity of sparse viewpoints, avoid multi-objective conflicts, and improve the reliability of the total loss function.
[0111] In one embodiment, updating the initialization parameter information based on the total loss function includes:
[0112] Based on the total loss function, the initialization parameter information is updated, and the rate of change of the total loss function is determined in real time during the update process. If the rate of change of loss is less than the rate of change threshold, or the number of iterations for updating the initialization parameter information is greater than the number of iterations threshold, the update of the initialization parameter information is stopped, and the updated initialization parameter information is used as the updated parameter information.
[0113] The loss change rate refers to the rate of change between the currently determined real-time total loss function and the loss function determined in the previous period during the update of the initialization parameter information. The change rate threshold and the iteration number threshold can be set according to actual needs.
[0114] The above scheme stops updating the initialization parameter information when the rate of change of loss is less than the rate of change threshold, or when the number of iterations for updating the initialization parameter information is greater than the number of iterations threshold. This can improve the efficiency of determining the updated parameter information while ensuring its reliability.
[0115] For example, based on the above embodiments, the method for reconstructing a 3D model based on a Gaussian sputtering model includes:
[0116] The process involves acquiring sparse viewpoint images of the target scene, reconstructing the 3D scene structure using structure-of-motion (SOP) techniques, and initializing the point cloud coordinates of the 3D scene structure. This allows for the determination of the geometric position of an initialized Gaussian ellipsoid based on these initial point cloud coordinates using a Gaussian sputtering model. The Gaussian sputtering model then determines the initialization parameters of the Gaussian ellipsoid, including its scale, rotation, spherical harmonics, and opacity. The scene image is then input into a monocular depth model, which processes the image to determine the prior depth, color loss function, and gradient smoothing loss function. Finally, a multi-view stereo depth estimation algorithm is used to determine a reference depth based on the scene image.
[0117] Based on the depth difference data between the prior depth and the reference depth, the depth confidence weight is determined; the rendering depth difference between the prior depth and the reference depth is determined, and the depth consistency loss function is determined based on the rendering depth difference and the depth confidence weight.
[0118] Using a Gaussian sputtering model, a dynamic gradient control loss function is determined based on the depth consistency loss function and the color loss function. Initial values for the dynamic weight parameters are set; these initial values include the initial weights of the depth consistency loss function, the color loss function, the gradient smoothing loss function, and the dynamic gradient control loss function. A hyperparameter optimization algorithm is used to dynamically adjust the initial weights of the depth consistency loss function and the dynamic gradient control loss function, determining their adjusted values. Finally, the adjusted weights of the depth consistency loss function, the dynamic gradient control loss function, the color loss function, and the gradient smoothing loss function are... The values are used as weights for the loss function, and the total loss function is determined using a Gaussian sputtering model based on the loss function weights, depth consistency loss function, color loss function, gradient smoothing loss function, and dynamic gradient control loss function. Based on the total loss function, the initialization parameter information is updated, and the rate of change of the total loss function is determined in real time during the update process. If the rate of change of the loss is less than the rate of change threshold, or the number of iterations for updating the initialization parameter information is greater than the number of iterations threshold, the update of the initialization parameter information is stopped, and the updated initialization parameter information is used as the updated parameter information to generate a 3D model of the target scene.
[0119] The aforementioned 3D model reconstruction method based on the Gaussian sputtering model acquires scene images of the target scene from sparse perspectives. Using structure-of-motion (SOR) reconstruction and the Gaussian sputtering model, initialization parameters of the Gaussian ellipsoid are determined from the scene images. A monocular depth model is used to determine prior depth, color loss function, and gradient smoothing loss function based on the scene images. A multi-view stereo depth estimation algorithm is used to determine a reference depth based on the scene images. A depth consistency loss function is determined based on the prior and reference depths. The Gaussian sputtering model is then used to update the initialization parameters based on the depth consistency loss function, color loss function, and gradient smoothing loss function, determining the updated parameters and generating a 3D model of the target scene based on these updated parameters. This method solves the problem that the Gaussian sputtering model struggles to construct reliable 3D scene models from scene images with sparse perspectives. The above scheme uses a monocular depth model to determine the prior depth based on the scene image from a sparse viewpoint and the reference depth determined by the multi-view stereo depth estimation algorithm. It then determines a depth consistency loss function and introduces a dynamic gradient control loss function to resolve the contradiction between geometric and appearance constraints in the gradient optimization direction when constructing the 3D model of the target scene. Based on the depth consistency loss function, dynamic gradient control loss function, color loss function, and gradient smoothing loss function, the initial parameters are updated. A Gaussian sputtering model is then used to generate the 3D model of the target scene based on the updated parameters. This approach improves the reliability and robustness of the 3D model through depth prior constraints and dynamic gradient control.
[0120] Based on the same inventive concept, this application also provides a Gaussian sputtering model-based 3D model reconstruction apparatus for implementing the above-described Gaussian sputtering model-based 3D model reconstruction method. The solution provided by this apparatus is similar to the implementation described in the above-described method. Therefore, the specific limitations of one or more Gaussian sputtering model-based 3D model reconstruction apparatus embodiments provided below can be found in the above-described limitations of the Gaussian sputtering model-based 3D model reconstruction method, and will not be repeated here.
[0121] In one embodiment, such as Figure 6 As shown, a 3D model reconstruction device based on a Gaussian sputtering model is provided, including: an initialization information determination module 601, a priori depth determination module 602, a reference depth determination module 603, a depth loss function determination module 604, and a 3D model construction module 605, wherein:
[0122] The initialization information determination module 601 is used to acquire scene images of the target scene from a sparse perspective, and determine the initialization parameter information of the Gaussian ellipsoid based on the scene images using motion recovery structure technology and Gaussian sputtering model.
[0123] The prior depth determination module 602 is used to determine the prior depth, color loss function and gradient smoothing loss function based on the scene image using a monocular depth model.
[0124] The reference depth determination module 603 is used to determine the reference depth based on the scene image using a multi-view stereo depth estimation algorithm.
[0125] The depth loss function determination module 604 is used to determine the depth consistency loss function based on the prior depth and the reference depth.
[0126] The 3D model construction module 605 is used to update the initialization parameter information, determine the updated parameter information, and generate a 3D model of the target scene based on the updated parameter information, using a Gaussian sputtering model and based on the depth consistency loss function, the color loss function, and the gradient smoothing loss function.
[0127] For example, the depth loss function determination module 604 is specifically used for:
[0128] Determine the depth confidence weight based on the depth difference data between the prior depth and the reference depth;
[0129] Determine the rendering depth difference between the prior depth and the reference depth, and determine the depth consistency loss function based on the rendering depth difference and the depth confidence weight.
[0130] For example, the 3D model building module 605 is specifically used for:
[0131] Using the Gaussian sputtering model, a dynamic gradient control loss function is determined based on the depth consistency loss function and the color loss function;
[0132] Based on the depth consistency loss function, the color loss function, the gradient smoothing loss function, and the dynamic gradient control loss function, the initialization parameter information is updated, and the updated parameter information is determined.
[0133] Furthermore, the 3D model building module 605 is also specifically used for:
[0134] Set initial values for dynamic weight parameters; the initial values for dynamic weight parameters include initial values for the weights of the depth consistency loss function, the color loss function, the gradient smoothing loss function, and the dynamic gradient control loss function.
[0135] The initial weights of the deep consistency loss function and the dynamic gradient control loss function are dynamically adjusted using a hyperparameter optimization algorithm to determine the adjusted weights of the deep consistency loss function and the dynamic gradient control loss function.
[0136] The weights of the depth consistency loss function, the dynamic gradient control loss function, the initial weights of the color loss function, and the initial weights of the gradient smoothing loss function are used as the weights of the loss function. The total loss function is determined by the Gaussian sputtering model based on the weights of the loss function, the depth consistency loss function, the color loss function, the gradient smoothing loss function, and the dynamic gradient control loss function.
[0137] Based on the total loss function, the initialization parameter information is updated, and the updated parameter information is determined.
[0138] Furthermore, the 3D model building module 605 is also specifically used for:
[0139] By using a hyperparameter optimization algorithm, the initial weight values of the deep consistency loss function are dynamically adjusted based on a preset number of iterations, thereby determining the weight adjustment values of the deep consistency loss function.
[0140] The geometric gradient is determined based on the depth consistency loss function, and the appearance gradient is determined based on the color loss function.
[0141] A dynamic gradient controller is used to detect conflicts between the geometric gradient and the appearance gradient in order to determine the gradient angle between the geometric gradient and the appearance gradient.
[0142] The weight parameters of the dynamic gradient control loss function are determined based on the gradient angle, and the initial weight values of the dynamic gradient control loss function are adjusted according to the weight parameters to determine the weight adjustment values of the dynamic gradient control loss function.
[0143] Furthermore, the 3D model building module 605 is also specifically used for:
[0144] Based on the total loss function, the initialization parameter information is updated, and during the updating of the initialization parameter information, the rate of change of the total loss function is determined in real time.
[0145] If the rate of change of loss is less than the rate of change threshold, or the number of iterations for updating the initialization parameter information is greater than the number of iterations threshold, then the updating of the initialization parameter information is stopped, and the updated initialization parameter information is used as the updated parameter information.
[0146] For example, the initialization information determination module 601 is specifically used for:
[0147] Using motion reconstruction structure technology, the three-dimensional scene structure is determined based on the scene image, and point cloud coordinates are generated based on the three-dimensional scene structure;
[0148] The initialization parameters of the Gaussian ellipsoid are determined based on the point cloud coordinates using the Gaussian sputtering model.
[0149] The modules in the aforementioned 3D model reconstruction device based on the Gaussian sputtering model can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in the processor of a computer device in hardware form or independent of it, or stored in the memory of a computer device in software form, so that the processor can call and execute the corresponding operations of each module.
[0150] In one embodiment, a computer device is provided, which may be a terminal, and its internal structure diagram may be as follows: Figure 7 As shown, the computer device includes a processor, memory, input / output interfaces, a communication interface, a display unit, and an input device. The processor, memory, and input / output interfaces are connected via a system bus, and the communication interface, display unit, and input device are also connected to the system bus via the input / output interfaces. The processor provides computational and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The input / output interfaces are used for exchanging information between the processor and external devices. The communication interface is used for wired or wireless communication with external terminals; wireless communication can be achieved through Wi-Fi, mobile cellular networks, NFC (Near Field Communication), or other technologies. When executed by the processor, the computer program implements a 3D model reconstruction method based on a Gaussian sputtering model. The display unit is used to form a visually visible image and can be a display screen, a projection device, or a virtual reality imaging device. The display screen can be an LCD screen or an e-ink screen. The input device of the computer device can be a touch layer covering the display screen, or buttons, trackballs, or touchpads set on the casing of the computer device, or external keyboards, touchpads, or mice, etc.
[0151] Those skilled in the art will understand that Figure 7 The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the computer device to which the present application is applied. Specific computer devices may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.
[0152] In one embodiment, a computer device is provided, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to perform the following steps:
[0153] Step 1: Obtain scene images of the target scene from a sparse perspective, and determine the initialization parameter information of the Gaussian ellipsoid based on the scene images using motion reconstruction structure technology and Gaussian sputtering model;
[0154] Step 2: Using a monocular depth model, determine the prior depth, color loss function, and gradient smoothing loss function based on the scene image;
[0155] Step 3: Determine the reference depth based on the scene image using a multi-view stereo depth estimation algorithm;
[0156] Step 4: Determine the depth consistency loss function based on the prior depth and the reference depth;
[0157] Step 5: Using the Gaussian sputtering model, based on the depth consistency loss function, the color loss function, and the gradient smoothing loss function, update the initialization parameter information, determine the updated parameter information, and generate a 3D model of the target scene according to the updated parameter information.
[0158] In one embodiment, a computer-readable storage medium is provided having a computer program stored thereon, the computer program performing the following steps when executed by a processor:
[0159] Step 1: Obtain scene images of the target scene from a sparse perspective, and determine the initialization parameter information of the Gaussian ellipsoid based on the scene images using motion reconstruction structure technology and Gaussian sputtering model;
[0160] Step 2: Using a monocular depth model, determine the prior depth, color loss function, and gradient smoothing loss function based on the scene image;
[0161] Step 3: Determine the reference depth based on the scene image using a multi-view stereo depth estimation algorithm;
[0162] Step 4: Determine the depth consistency loss function based on the prior depth and the reference depth;
[0163] Step 5: Using the Gaussian sputtering model, based on the depth consistency loss function, the color loss function, and the gradient smoothing loss function, update the initialization parameter information, determine the updated parameter information, and generate a 3D model of the target scene according to the updated parameter information.
[0164] In one embodiment, a computer program product is provided, including a computer program that, when executed by a processor, performs the following steps:
[0165] Step 1: Obtain scene images of the target scene from a sparse perspective, and determine the initialization parameter information of the Gaussian ellipsoid based on the scene images using motion reconstruction structure technology and Gaussian sputtering model;
[0166] Step 2: Using a monocular depth model, determine the prior depth, color loss function, and gradient smoothing loss function based on the scene image;
[0167] Step 3: Determine the reference depth based on the scene image using a multi-view stereo depth estimation algorithm;
[0168] Step 4: Determine the depth consistency loss function based on the prior depth and the reference depth;
[0169] Step 5: Using the Gaussian sputtering model, based on the depth consistency loss function, the color loss function, and the gradient smoothing loss function, update the initialization parameter information, determine the updated parameter information, and generate a 3D model of the target scene according to the updated parameter information.
[0170] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, data stored, data displayed, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties, and the collection, use and processing of related data must comply with the relevant laws, regulations and standards of the relevant countries and regions.
[0171] Those skilled in the art will understand that all or part of the processes in the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium. When executed, the computer program can include the processes of the embodiments described above. Any references to memory, databases, or other media used in the embodiments provided in this application can include at least one of non-volatile and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, optical memory, high-density embedded non-volatile memory, resistive random access memory (ReRAM), magnetic random access memory (MRAM), ferroelectric random access memory (FRAM), phase change memory (PCM), graphene memory, etc. Volatile memory can include random access memory (RAM) or external cache memory, etc. By way of illustration and not limitation, RAM can take many forms, such as Static Random Access Memory (SRAM) or Dynamic Random Access Memory (DRAM). The databases involved in the embodiments provided in this application may include at least one type of relational database and non-relational database. Non-relational databases may include, but are not limited to, blockchain-based distributed databases. The processors involved in the embodiments provided in this application may be general-purpose processors, central processing units, graphics processing units, digital signal processors, programmable logic devices, quantum computing-based data processing logic devices, etc., and are not limited to these.
[0172] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.
[0173] The embodiments described above are merely illustrative of several implementation methods of this application, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of this patent application. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these all fall within the protection scope of this application. Therefore, the protection scope of this application should be determined by the appended claims.
Claims
1. A method for reconstructing a three-dimensional model based on a Gaussian sputtering model, characterized in that, include: Acquire scene images of the target scene from a sparse perspective, and determine the initialization parameter information of the Gaussian ellipsoid based on the scene images using structure of motion reconstruction and Gaussian sputtering model. Based on the scene image, the prior depth, color loss function, and gradient smoothing loss function are determined using a monocular depth model. A reference depth is determined based on the scene image using a multi-view stereo depth estimation algorithm. Determine the depth consistency loss function based on the prior depth and the reference depth; Using the Gaussian sputtering model, a dynamic gradient control loss function is determined based on the depth consistency loss function and the color loss function; Set initial values for dynamic weight parameters; the initial values for dynamic weight parameters include initial values for the weights of the depth consistency loss function, the color loss function, the gradient smoothing loss function, and the dynamic gradient control loss function. The initial weights of the deep consistency loss function and the dynamic gradient control loss function are dynamically adjusted using a hyperparameter optimization algorithm to determine the adjusted weights of the deep consistency loss function and the dynamic gradient control loss function. The weights of the depth consistency loss function, the dynamic gradient control loss function, the initial weights of the color loss function, and the initial weights of the gradient smoothing loss function are used as the weights of the loss function. The total loss function is determined by the Gaussian sputtering model based on the weights of the loss function, the depth consistency loss function, the color loss function, the gradient smoothing loss function, and the dynamic gradient control loss function. Based on the total loss function, the initialization parameter information is updated, the updated parameter information is determined, and a 3D model of the target scene is generated according to the updated parameter information.
2. The method according to claim 1, characterized in that, Based on the prior depth and the reference depth, a depth consistency loss function is determined, including: Determine the depth confidence weight based on the depth difference data between the prior depth and the reference depth; Determine the rendering depth difference between the prior depth and the reference depth, and determine the depth consistency loss function based on the rendering depth difference and the depth confidence weight.
3. The method according to claim 1, characterized in that, The initial weights of the depth consistency loss function and the dynamic gradient control loss function are dynamically adjusted using a hyperparameter optimization algorithm. This process determines the adjusted weight values for both the depth consistency loss function and the dynamic gradient control loss function, including: By using a hyperparameter optimization algorithm, the initial weight values of the deep consistency loss function are dynamically adjusted based on a preset number of iterations, thereby determining the weight adjustment values of the deep consistency loss function. The geometric gradient is determined based on the depth consistency loss function, and the appearance gradient is determined based on the color loss function. A dynamic gradient controller is used to detect conflicts between the geometric gradient and the appearance gradient in order to determine the gradient angle between the geometric gradient and the appearance gradient. The weight parameters of the dynamic gradient control loss function are determined based on the gradient angle, and the initial weight values of the dynamic gradient control loss function are adjusted according to the weight parameters to determine the weight adjustment values of the dynamic gradient control loss function.
4. The method according to claim 1, characterized in that, Based on the total loss function, the initialization parameter information is updated, including: Based on the total loss function, the initialization parameter information is updated, and during the updating of the initialization parameter information, the rate of change of the total loss function is determined in real time. If the rate of change of loss is less than the rate of change threshold, or the number of iterations for updating the initialization parameter information is greater than the number of iterations threshold, then the updating of the initialization parameter information is stopped, and the updated initialization parameter information is used as the updated parameter information.
5. The method according to claim 1, characterized in that, Using motion recovery structure technology and a Gaussian sputtering model, the initialization parameter information of the Gaussian ellipsoid is determined based on the scene image, including: Using motion reconstruction structure technology, the three-dimensional scene structure is determined based on the scene image, and point cloud coordinates are generated based on the three-dimensional scene structure; The initialization parameters of the Gaussian ellipsoid are determined based on the point cloud coordinates using the Gaussian sputtering model.
6. A three-dimensional model reconstruction device based on a Gaussian sputtering model, characterized in that, The three-dimensional model reconstruction device based on the Gaussian sputtering model includes: The initialization information determination module is used to acquire scene images of the target scene from a sparse perspective, and determine the initialization parameter information of the Gaussian ellipsoid based on the scene images using motion reconstruction structure technology and Gaussian sputtering model. The prior depth determination module is used to determine the prior depth, color loss function, and gradient smoothing loss function based on the scene image using a monocular depth model. The reference depth determination module is used to determine the reference depth based on the scene image using a multi-view stereo depth estimation algorithm. A depth loss function determination module is used to determine a depth consistency loss function based on the prior depth and the reference depth. A 3D model construction module is used to determine a dynamic gradient control loss function based on the depth consistency loss function and the color loss function using a Gaussian sputtering model; set initial values for dynamic weight parameters, including initial weights for the depth consistency loss function, color loss function, gradient smoothing loss function, and dynamic gradient control loss function; dynamically adjust the initial weights of the depth consistency loss function and dynamic gradient control loss function using a hyperparameter optimization algorithm to determine adjusted weights for the depth consistency loss function and dynamic gradient control loss function; use these adjusted weights as loss function weights, and determine a total loss function based on the loss function weights, depth consistency loss function, color loss function, gradient smoothing loss function, and dynamic gradient control loss function using the Gaussian sputtering model; update the initialization parameter information based on the total loss function, determine updated parameter information, and generate a 3D model of the target scene based on the updated parameter information.
7. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the method according to any one of claims 1 to 5.
8. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the steps of the method according to any one of claims 1 to 5.
