CN-122023744-A - 3D Gaussian model non-rigid correction method, system and medium based on spatial deformation field

CN122023744ACN 122023744 ACN122023744 ACN 122023744ACN-122023744-A

Abstract

The embodiment of the invention discloses a non-rigid correction method, a system and a medium of a 3D Gaussian model based on a space deformation field, wherein the method comprises the steps of obtaining an original 3DGS model to be corrected, wherein the original 3DGS model comprises an original center position of each Gaussian sphere; registering the original model with a preset deformation standard model through a non-rigid registration algorithm to generate a space deformation graph containing control nodes and a corresponding affine transformation matrix. And for each Gaussian sphere, searching K nearest neighbor control nodes in the deformation graph according to the original center position of the Gaussian sphere, and calculating a normalized interpolation weight according to the distance. The weight is utilized to carry out weighted interpolation on the transformation of each control node, and then the center position of each Gaussian ball is updated efficiently; and performing rasterization rendering based on the updated original center position to obtain the corrected 3DGS model accurately aligned with the deformation standard model. The method is completely executed in the reasoning stage, retraining is not needed, and the requirement of quick, accurate and instant correction of the model in the engineering site can be met.

Inventors

ZHU YUANBIAO
ZU YANAN

Assignees

绘见科技(深圳)有限公司

Dates

Publication Date: 20260512
Application Date: 20260203

Claims (10)

1. The method for non-rigid correction of the 3D Gaussian model based on the spatial deformation field is characterized by comprising the following steps of: Acquiring an original 3DGS model to be rectified, wherein the original 3DGS model comprises the attribute of each Gaussian sphere, and the attribute comprises an original center position; Registering the original 3DGS model with a preset deformation standard model through a non-rigid registration algorithm to obtain a space deformation graph, wherein the space deformation graph comprises control nodes and a corresponding affine transformation matrix; obtaining K nearest neighbor control nodes of the original center position of each Gaussian sphere in the space deformation graph, and determining the normalized interpolation weight of each Gaussian sphere according to the distance between the original center position and each nearest neighbor control node; Updating the original center position of each Gaussian sphere by using the normalized interpolation weight; And performing rasterization rendering based on the updated original center position of the Gaussian sphere to obtain a corrected 3DGS model.
2. The method for non-rigid rectification of a 3D gaussian model based on a spatial deformation field according to claim 1, wherein the attributes further comprise an original rotation quaternion and an original covariance matrix, the rasterization rendering is performed based on the updated original center position, and a rectified 3DGS model is obtained, which specifically comprises: Updating an original rotation quaternion and an original covariance matrix of each Gaussian sphere by using the normalized interpolation weight; And performing rasterization rendering based on the updated original center position of the Gaussian sphere, the original rotation quaternion and the original covariance matrix to obtain a corrected 3DGS model.
3. The method for non-rigid rectification of a 3D gaussian model based on a spatial deformation field according to claim 2, wherein updating the original rotation quaternion and the original covariance matrix of each gaussian sphere by using the normalized interpolation weights specifically comprises: extracting rotation components from affine transformation matrixes of each control node, converting the rotation components into quaternions, and carrying out weighted interpolation on the quaternions based on the normalized interpolation weights to obtain local space distortion rotation quantity; Determining a final rotation quaternion of each Gaussian sphere according to the local space distortion rotation quantity and the original rotation quaternion of the Gaussian sphere; Determining a jacobian matrix of the spatial deformation map at a center position of the Gaussian sphere; And transforming the original covariance matrix of the Gaussian balls according to the jacobian matrix, and determining the final covariance matrix of each Gaussian ball.
4. The method for non-rigid rectification of a 3D gaussian model based on a spatial deformation field according to claim 3, wherein said determining a jacobian matrix of said spatial deformation map at an original center position of a gaussian sphere specifically comprises: taking the original center position of the Gaussian sphere as input and the final center position of the Gaussian sphere as output, and constructing a deformation function of the space deformation graph at the original center position of the Gaussian sphere; determining the partial derivative of the deformation function by a finite difference method; And arranging the partial derivatives according to a preset sequence, and determining a jacobian matrix of the spatial deformation map at the original center position of the Gaussian sphere.
5. The method for non-rigid rectification of a 3D gaussian model based on a spatial deformation field according to claim 1, wherein updating the original center position of each gaussian sphere by using the normalized interpolation weights specifically comprises: using the normalized interpolation weights according to Updating the original center position of each Gaussian ball, determining the final center position of each Gaussian ball, wherein, For the final center position of the ith gaussian sphere, In order to normalize the interpolation weights, For the radiation conversion matrix corresponding to the j nearest neighbor control node, Is the original center position of the ith Gaussian ball.
6. The method for non-rigid rectification of a 3D gaussian model based on a spatial deformation field according to claim 3, wherein said determining a final rotation quaternion for each gaussian sphere based on the local spatial distortion rotation and the original rotation quaternion for the gaussian sphere specifically comprises: According to A final rotational quaternion for each gaussian sphere is determined, wherein, For the final rotational quaternion of the ith gaussian sphere, For the amount of local spatial twist rotation, Is the original rotation quaternion of the ith gaussian sphere.
7. The method for non-rigid rectification of a 3D gaussian model based on a spatial deformation field according to claim 3, wherein said transforming the original covariance matrix of the gaussian balls according to the jacobian matrix, determining the final covariance matrix of each gaussian ball, specifically comprises: According to A final covariance matrix for each gaussian sphere is determined, wherein, For the final covariance matrix of the ith gaussian sphere, In the form of a jacobian matrix, For the original covariance matrix of the ith gaussian sphere, Is the transpose of the jacobian matrix.
8. The method for non-rigid rectification of a 3D gaussian model based on a spatial deformation field according to claim 1, wherein determining a normalized interpolation weight of each gaussian sphere according to the distance between the original center position and each nearest neighbor control node position specifically comprises: According to The distance between the original center position of each gaussian sphere and the control node position is determined, wherein, As the distance between the original center position of the ith gaussian ball and the position of the jth control node, Is the original center position of the ith gaussian sphere, The position of the j-th control node; According to An interpolation weight for each gaussian sphere is determined, wherein, For the purpose of the interpolation weights, Is a very small positive value; According to A normalized interpolation weight for each gaussian sphere is determined, wherein, In order to normalize the interpolation weights, Is the sum of the difference weights of all K nearest neighbor control nodes of the ith Gaussian sphere.
9. A 3D gaussian model non-rigid correction system based on a spatial deformation field, the system comprising: The system comprises an original 3DGS model acquisition module, a correction module and a correction module, wherein the original 3DGS model acquisition module is used for acquiring an original 3DGS model to be corrected, the original 3DGS model comprises the attribute of each Gaussian sphere, and the attribute comprises an original center position; The space deformation graph acquisition module is used for registering the original 3DGS model with a preset deformation standard model through a non-rigid registration algorithm to acquire a space deformation graph, wherein the space deformation graph comprises control nodes and a corresponding affine transformation matrix; The normalized interpolation weight determining module is used for acquiring K nearest neighbor control nodes of the original center position of each Gaussian sphere in the space deformation graph and determining the normalized interpolation weight of each Gaussian sphere according to the distance between the original center position and each nearest neighbor control node; The attribute updating module is used for updating the original center position of each Gaussian ball by utilizing the normalized interpolation weight; and the corrected 3DGS model acquisition module is used for carrying out rasterization rendering based on the updated original center position of the Gaussian sphere to acquire the corrected 3DGS model.
10. A computer readable storage medium storing a computer program which, when executed by a processor, causes the processor to perform the steps of the method of any one of claims 1 to 8.

Description

3D Gaussian model non-rigid correction method, system and medium based on spatial deformation field Technical Field The invention relates to the technical field of computer graphics, in particular to a non-rigid correction method, a system and a medium for a 3D Gaussian model based on a space deformation field. Background Under the current technical background, the 3DGS (3D Gaussian Splatting, three-dimensional Gaussian splatter) technology can generate an extremely realistic three-dimensional scene, and is widely applied to multiple fields of three-dimensional modeling, virtual reality and the like. However, when processing large-scene industrial scans, the generated 3DGS model, though appearing very realistic, has significant positioning bias in practical applications, and cannot be precisely overlapped with the prefabricated BIM (Building Information Modeling, building information model) or CAD (Computer-AIDED DESIGN, computer aided design) drawing due to accumulated drift errors generated by underlying data sources such as multi-view reconstruction and synchronous positioning and mapping. Such positioning deviations affect the application range and accuracy of the model, and especially the need for real-time delivery becomes more pronounced when high accuracy is required on the engineering site. In the conventional processing method, when a single three-dimensional light field model needs to be moved, the center position of the Gaussian sphere is usually simply adjusted, and the operation is simple but has obvious defects, such as visual problems of 'shutter effect', artifact or blurring and the like of a rendering result if rotation and scaling properties of the Gaussian sphere are not adjusted at the same time. While some correction methods have been proposed to address these problems, these methods typically require retraining the model during the adjustment of the model, which is extremely time consuming, particularly when dealing with large-scale data, which can take hours, which is far from meeting the real-time delivery requirements of the engineering site. In view of the foregoing, the prior art has many shortcomings, and a three-dimensional model adjustment method capable of realizing real-time, high-efficiency and without losing accuracy is needed to meet the demands of practical applications. The invention provides a direct mapping method based on a discrete deformation graph, which aims to provide a scheme capable of rapidly and accurately correcting a 3DGS model and ensuring the anisotropic property and geometric consistency of the model so as to adapt to the strict requirements of large-scale scene scanning and real-time application. Disclosure of Invention Based on the above, it is necessary to provide a method, a system and a medium for non-rigid correction of a 3D gaussian model based on a spatial deformation field. A method for non-rigid rectification of a 3D gaussian model based on a spatial deformation field, the method comprising: And acquiring an original 3DGS model to be rectified, wherein the original 3DGS model comprises the attribute of each Gaussian sphere, and the attribute comprises an original center position. Registering the original 3DGS model with a preset deformation standard model through a non-rigid registration algorithm to obtain a space deformation graph, wherein the space deformation graph comprises control nodes and a corresponding affine transformation matrix. And obtaining K nearest neighbor control nodes of the original center position of each Gaussian sphere in the space deformation graph, and determining the normalized interpolation weight of each Gaussian sphere according to the distance between the original center position and each nearest neighbor control node. And updating the original center position of each Gaussian sphere by using the normalized interpolation weight. And performing rasterization rendering based on the updated original center position of the Gaussian sphere to obtain a corrected 3DGS model. The attribute further comprises an original rotation quaternion and an original covariance matrix, the rasterization rendering is performed based on the updated original center position, and a 3DGS model after deviation correction is obtained, and the method specifically comprises the following steps: And updating the original rotation quaternion and the original covariance matrix of each Gaussian sphere by using the normalized interpolation weight. And performing rasterization rendering based on the updated original center position of the Gaussian sphere, the original rotation quaternion and the original covariance matrix to obtain a corrected 3DGS model. The updating the original rotation quaternion and the original covariance matrix of each Gaussian sphere by using the normalized interpolation weight specifically comprises the following steps: And extracting rotation components from the affine transformation matrix of each control node, convertin