Bumpless cropping method, device and equipment for 3D GS scene editing and medium

Abstract

The application provides a no-burr clipping method and device for 3DGS scene editing, equipment and medium, comprising: obtaining a 3DGS scene model to be clipped composed of a plurality of 3D Gaussian kernels; defining at least one clipping plane of the 3DGS scene model; for each 3D Gaussian kernel in the 3DGS scene model, if the spatial distribution represented by the covariance of the 3D Gaussian kernel intersects with the clipping plane, performing a distribution-level plane clipping operation to split the 3D Gaussian kernel into two sub-Gaussian kernels, which are respectively located in the positive side and negative side of the probability distribution part of the clipping plane; performing a rejection processing on the sub-Gaussian kernel located in the rejection side, and performing a retention processing on the sub-Gaussian kernel located in the retention side, and outputting the clipped 3DGS sub-model. The density jump and visual burr are effectively avoided, the true sense of "seamless clipping" is realized, and the problems of edge burr, geometric distortion, distribution fracture and the like existing in the existing 3DGS model clipping technology are overcome.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of computer graphics and three-dimensional scene modeling, in particular to a no-burr clipping method, device and equipment and medium for 3DGS scene editing. BACKGROUND

[0002] As a new emerging explicit three-dimensional scene representation method, 3D Gaussian Splatting (3DGS) has shown significant advantages in high-quality, real-time rendering tasks in recent years. In the actual application process, the local editing operation of the 3DGS model is extremely frequent, and "clipping" is one of the basic and key functions. However, the existing 3DGS clipping method generally has serious visual defect problems, especially the "burr", jagged edge, local hole or geometric depression phenomenon at the clipping boundary, which seriously affects the subsequent rendering quality and user experience.

[0003] The current mainstream clipping strategy mainly includes two categories: 1. Center judgment method: each Gaussian kernel is regarded as a point entity, and only the spatial position of the center coordinates relative to the clipping plane is used to make a decision on retention or rejection. This method is simple to calculate but ignores the actual spatial support range of the Gaussian kernel (i.e. the covariance ellipsoid), resulting in a large number of Gaussian kernels near the boundary that cross the clipping plane being roughly truncated, causing density mutations and visual burrs. 2. Hard boundary rejection method: considering the ellipsoid shape of the Gaussian kernel, if it intersects with the clipping plane, the Gaussian kernel is directly deleted. Although it avoids the fuzzy mixing across the boundary, it often leads to local holes and geometric breaks, destroying the distribution continuity of the original scene. In summary, there are problems such as edge burrs, geometric distortion, and distribution break in the existing 3DGS model clipping technology. Therefore, there is an urgent need for a new clipping method that can not only maintain efficient analytical calculation characteristics but also eliminate visual artifacts at the clipping edge. SUMMARY

[0004] Therefore, the purpose of the present application is to provide a no-burr clipping method, device and equipment and medium for 3DGS scene editing, which reconstructs the Gaussian kernel across the clipping plane at the distribution level, generates new sub-Gaussian kernels, not only preserves the spatial continuity of the original Gaussian distribution, but also effectively avoids density jumps and visual burrs, realizes true "seamless clipping", and overcomes the problems of edge burrs, geometric distortion, and distribution break in the existing 3DGS model clipping technology.

[0005] The present application provides a no-burr clipping method for 3DGS scene editing, which comprises:

[0006] Obtaining a 3DGS scene model to be clipped composed of a plurality of 3D Gaussian kernels;

[0007] defining at least one clipping plane of the 3DGS scene model; wherein the clipping plane is an infinitely extended plane with a unit normal vector and a intercept parameter, for dividing the 3DGS scene model into a keep side and a cull side;

[0008] for each 3D Gaussian kernel in the 3DGS scene model, if the spatial distribution represented by the covariance of the 3D Gaussian kernel intersects with the clipping plane, performing a distribution-level plane clipping operation to split the 3D Gaussian kernel into two sub-Gaussian kernels, which are located in the positive side and the negative side of the probability distribution part of the clipping plane, respectively;

[0009] performing cull processing on the sub-Gaussian kernel located in the cull side, and performing keep processing on the sub-Gaussian kernel located in the keep side, and outputting a clipped 3DGS sub-model.

[0010] In one possible implementation, the performing of the distribution-level plane clipping operation to split the 3D Gaussian kernel into two sub-Gaussian kernels comprises:

[0011] performing Cholesky decomposition on the covariance matrix of the current 3D Gaussian kernel to generate a lower triangular matrix;

[0012] constructing an affine transformation based on the lower triangular matrix, mapping the current 3D Gaussian kernel to a standard space, and synchronously projecting the clipping plane to the space;

[0013] calculating the probability mass of the Gaussian distribution in the positive side and the negative side of the clipping plane in the standard space, and solving the statistical parameters of the two-side sub-Gaussian kernels based on the truncated multivariate normal distribution theory;

[0014] transforming the statistical parameters back to the original space to obtain the positions and covariance matrices of the two sub-Gaussian kernels;

[0015] weighting and allocating the transparency according to the probability mass, and inheriting the color information of the original Gaussian kernel to generate the two sub-Gaussian kernels.

[0016] In one possible implementation, the clipping plane is a limited plane region, and the non-stub clipping method further comprises:

[0017] selecting a point on the infinitely extended plane, and constructing two mutually orthogonal unit basis vectors, setting the extension scale along the two unit basis vectors, and determining a rectangular limited plane region;

[0018] detecting whether the 3D Gaussian kernel intersects with the limited plane region, and if yes, performing the distribution-level plane clipping operation on the 3D Gaussian kernel to determine a clipped 3DGS sub-model.

[0019] In one possible implementation, whether the 3D Gaussian kernel intersects with the limited planar region is determined by the following steps:

[0020] A point of orthogonal projection of the center of the 3D Gaussian kernel along the normal of the clipping plane is determined:

[0021] It is determined whether the point of orthogonal projection falls within the limited planar region and whether the spatial distribution represented by the covariance of the 3D Gaussian kernel intersects with the corresponding infinitely extended plane;

[0022] If both are yes, the 3D Gaussian kernel intersects with the limited planar region;

[0023] If either is no, the 3D Gaussian kernel does not intersect with the limited planar region.

[0024] In one possible implementation, for a closed geometric body composed of a plurality of clipping planes, the non-stub clipping method further comprises:

[0025] The closed geometric body is decomposed into a plurality of limited clipping planes facing the interior; wherein the normal vector of each limited clipping plane points to the interior of the closed geometric body;

[0026] Each of the limited clipping planes is arranged in a preset clipping order, and the distributed-level plane clipping operation is sequentially performed on the 3D GS scene model to determine the 3D GS sub-model limited in the closed geometric body.

[0027] In one possible implementation, the distributed-level plane clipping operation is sequentially performed on the 3D GS scene model in the preset clipping order to determine the 3D GS sub-model limited in the closed geometric body, comprising:

[0028] Each 3D Gaussian kernel in the 3D GS scene model is taken as input, and a distributed-level plane clipping operation is performed according to the geometric relationship between the covariance matrix of the 3D Gaussian kernel and the limited clipping plane to generate an intermediate 3D GS subset after this clipping;

[0029] The intermediate 3D GS subset output by the current clipping step is taken as the input of the distributed-level plane clipping operation of the next limited clipping plane, and the next limited clipping plane is continuously processed until the distributed-level plane clipping operation of each limited clipping plane is completed, and the 3D GS sub-model limited in the closed geometric body is determined.

[0030] In one possible implementation, after the distributed-level plane clipping operation of the limited clipping plane is completed, the non-stub clipping method further comprises:

[0031] The spatial distribution characteristics of the remaining 3D Gaussian kernel set are dynamically updated;

[0032] Based on the updated distribution characteristics, the processing priority of the subsequent limited clipping plane is re-evaluated to determine an evaluation result, and the clipping order is adjusted according to the evaluation result.

[0033] The embodiments of the present application also provide a no-burr clipping device for 3DGS scene editing, which comprises:

[0034] An acquisition module is configured to acquire a 3DGS scene model to be clipped, which is composed of a plurality of 3D Gaussian kernels.

[0035] A definition module is configured to define at least one clipping plane of the 3DGS scene model, wherein the clipping plane is an infinitely extended plane with a unit normal vector and an intercept parameter, and is used to divide the 3DGS scene model into a reserved side and a rejected side.

[0036] A clipping operation module is configured to, for each 3D Gaussian kernel in the 3DGS scene model, if a spatial distribution represented by a covariance of the 3D Gaussian kernel intersects the clipping plane, perform a distribution-level plane clipping operation to split the 3D Gaussian kernel into two sub-Gaussian kernels, which are located in the probability distribution parts on the positive side and the negative side of the clipping plane, respectively.

[0037] A determination module is configured to perform rejection processing on the sub-Gaussian kernel located on the rejected side, perform reservation processing on the sub-Gaussian kernel located on the reserved side, and output a clipped 3DGS sub-model.

[0038] The embodiments of the present application also provide an electronic device, which comprises a processor, a memory and a bus, the memory stores machine readable instructions executable by the processor, when the electronic device is running, the processor and the memory communicate through the bus, and the machine readable instructions are executed by the processor to perform the steps of the no-burr clipping method for 3DGS scene editing as described above.

[0039] The embodiments of the present application also provide a computer readable storage medium, which stores a computer program, and the computer program is executed by the processor to perform the steps of the no-burr clipping method for 3DGS scene editing as described above.

[0040] The method, device, equipment and medium for 3DGS scene editing provided by the embodiment of the application, the method for 3DGS scene editing without burr cutting comprises the following steps: obtaining a 3DGS scene model to be cut composed of a plurality of 3D Gaussian kernels; defining at least one cutting plane of the 3DGS scene model; wherein the cutting plane is an infinitely extended plane with a unit normal vector and an intercept parameter, and is used to divide the 3DGS scene model into a reserved side and a rejected side; for each 3D Gaussian kernel in the 3DGS scene model, if the spatial distribution represented by the covariance of the 3D Gaussian kernel intersects with the cutting plane, a distribution level plane cutting operation is performed to split the 3D Gaussian kernel into two sub Gaussian kernels, which are located in the positive side and the negative side of the probability distribution part of the cutting plane respectively; the sub Gaussian kernel located in the rejected side is subjected to rejection processing, and the sub Gaussian kernel located in the reserved side is subjected to reservation processing, and a cut 3DGS sub model is output. The Gaussian kernel across the cutting plane is reconstructed at the distribution level to generate a new sub Gaussian kernel, which not only retains the spatial continuity of the original Gaussian distribution, but also effectively avoids the density jump and visual burr, realizes the true sense of "seamless cutting", and overcomes the problems of edge burr, geometric distortion, distribution rupture and the like in the existing 3DGS model cutting technology.

[0041] In order to make the above objectives, characteristics and advantages of the present application more apparent, clear and easy to understand, the following preferred embodiments are specifically described below, and the accompanying drawings are referred to for detailed description. BRIEF DESCRIPTION OF DRAWINGS

[0042] In order to more clearly illustrate the technical solutions of the embodiments of the present application, the following will briefly introduce the drawings needed to be used in the embodiments. It should be understood that the following drawings only show some embodiments of the present application, and therefore should not be regarded as a limitation on the scope. For those skilled in the art, other related drawings can also be obtained without creative labor on the basis of these drawings.

[0043] Figure 1 A flow chart of a 3DGS scene editing method without burr cutting provided by an embodiment of the present application;

[0044] Figure 2 A structure schematic diagram of a 3DGS scene editing device without burr cutting provided by an embodiment of the present application;

[0045] Figure 3 A structure schematic diagram of a 3DGS scene editing device without burr cutting provided by an embodiment of the present application;

[0046] Figure 4 A structure schematic diagram of an electronic device provided by an embodiment of the present application. DETAILED DESCRIPTION

[0047] In order to make the purposes, technical solutions and advantages of the embodiments of the present application clearer, the technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, rather than all the embodiments. The components of the embodiments of the present application described and shown in the drawings herein can be arranged and designed in various different configurations. Therefore, the following detailed description of the embodiments of the present application provided in the drawings is not intended to limit the scope of the claimed present application, but only represents selected embodiments of the present application. Based on the embodiments of the present application, every other embodiment obtained by a person skilled in the art without creative work belongs to the scope of protection of the present application.

[0048] Firstly, the application scenarios applicable to the present application are introduced. The present application can be applied to the technical field of computer graphics and three-dimensional scene modeling.

[0049] It is found through research that the prior art lacks sufficient consideration of the statistical properties and spatial continuity of the Gaussian kernel inside when processing 3DGS model clipping, and fails to achieve natural transition of the boundary region from the distribution level, and there are problems such as edge burr, geometric distortion, and distribution fracture. Therefore, there is an urgent need for a new clipping method that can not only maintain efficient analytical calculation characteristics, but also eliminate clipping edge visual artifacts.

[0050] Based on this, the embodiments of the present application provide a burr-free clipping method for 3DGS scene editing, which reconstructs the Gaussian kernel across the clipping plane at the distribution level to generate new sub-Gaussian kernels, not only preserving the spatial continuity of the original Gaussian distribution, but also effectively avoiding density jumps and visual burrs, achieving true "seamless clipping" and overcoming the problems of edge burr, geometric distortion, and distribution fracture in existing 3DGS model clipping techniques.

[0051] Please refer to Figure 1 , Figure 1 A flowchart of a burr-free clipping method for 3DGS scene editing provided by the embodiments of the present application. As shown in Figure 1 The burr-free clipping method provided by the embodiments of the present application includes:

[0052] S101: Obtain a 3DGS scene model to be clipped composed of a plurality of 3D Gaussian kernels.

[0053] It should be noted that the 3DGS scene model is an explicit scene representation structure composed of a large number of discrete 3D Gaussian kernels (Gaussian Splat), each of which serves as a basic rendering unit and contains attributes such as spatial position, scale parameter, rotation matrix, opacity, and spherical harmonic function coefficient, which collectively encode the geometric morphology and appearance information of the scene. The covariance matrix of each 3D Gaussian kernel can be calculated by combining the scale parameter and the rotation matrix.

[0054] S102: At least one clipping plane of the 3DGS scene model is defined; wherein the clipping plane is an infinitely extended plane with a unit normal vector and an intercept parameter, used to divide the 3DGS scene model into a reserved side and a rejected side.

[0055] In this step, at least one clipping plane of the 3DGS scene model is defined; wherein the clipping plane is an infinitely extended plane with a unit normal vector and an intercept parameter, used to divide the 3DGS scene model into a reserved side and a rejected side.

[0056] It should be noted that the direction of the unit normal vector is conventionally defined as the positive direction pointing to the reserved side, i.e., the side pointed by the normal vector is the space region to be reserved. By adjusting the values of the unit normal vector and the intercept parameter, the user or the system can flexibly set the spatial pose and position of the clipping plane, thereby realizing clipping operations at any angle and position.

[0057] S103: For each 3D Gaussian kernel in the 3DGS scene model, if the spatial distribution represented by the covariance of the 3D Gaussian kernel intersects with the clipping plane, a distribution-level plane clipping operation is performed to split the 3D Gaussian kernel into two sub-Gaussian kernels, located in the positive and negative probability distribution parts of the clipping plane, respectively.

[0058] In this step, for each 3D Gaussian kernel in the 3DGS scene model, if the spatial distribution represented by the covariance of the 3D Gaussian kernel intersects with the clipping plane, a distribution-level plane clipping operation is performed to split the 3D Gaussian kernel into two sub-Gaussian kernels, located in the positive and negative probability distribution parts of the clipping plane, respectively.

[0059] Here, when the aforementioned infinitely extended clipping plane is introduced, the spatial relationship between each 3D Gaussian kernel and the plane needs to be determined. In particular, if any of the following conditions is met: the center point of the 3D Gaussian kernel is located on the rejected side, but its covariance ellipsoid extends to the reserved side; or the center point is located on the reserved side, but its probability mass extends significantly to the rejected side; then it is considered that the spatial distribution of the Gaussian kernel intersects with the clipping plane, i.e., its probability density function has non-zero contribution on both sides of the plane.

[0060] For such a cross-boundary Gaussian kernel, the present application proposes a distribution-level plane clipping operation (also called "razor operation"), instead of simply retaining or deleting it as a whole, based on the statistical truncation theory of multi-dimensional Gaussian distribution on the half-space, the original distribution is analytically split into two mutually exclusive sub-distributions along the clipping plane, corresponding to the probability mass part of the positive side (retention side) and the negative side (rejection side) of the plane respectively.

[0061] In one possible implementation, the distribution-level plane clipping operation splits the 3D Gaussian kernel into two sub-Gaussian kernels, including:

[0062] A: Cholesky decomposition is performed on the covariance matrix of the current 3D Gaussian kernel to generate a lower triangular matrix.

[0063] Here, the internal region of the Gaussian kernel is defined as:

[0064]

[0065] wherein, is the internal region (ellipsoid) of a single Gaussian kernel in three-dimensional space, is an arbitrary point in three-dimensional space, is the center position of the 3D Gaussian kernel, is a scaled covariance matrix, wherein , Σ is the covariance matrix of the Gaussian kernel, describing its distribution shape in space, is the dimension.

[0066] Here, the clipping plane is: , is the unit normal vector, x is an arbitrary point on the plane, is the intercept parameter of the plane.

[0067] wherein, for cholesky decomposition is performed to obtain a lower triangular matrix S。

[0068] B: Based on the lower triangular matrix, an affine transformation is constructed to map the current 3D Gaussian kernel to the standard space, and the clipping plane is projected to the space at the same time.

[0069] Here, the affine transformation is defined as follows:

[0070]

[0071] wherein, the point in the original space is transformed to space, after transformation, the center of the Gaussian kernel is located at the origin of the y space, Sis a lower triangular matrix.

[0072] Here, the transformation parameters of the clipping plane are: , . Wherein, is the normal vector n , the transformed vector, is the unit normal vector of the transformed plane, is the intercept parameter of the plane, and α is the directed distance from the transformed plane to the origin. Determine whether the clipping is effective. If |α| < 1, the plane intersects the ellipsoid and can be divided. If |α| ≥ 1, the plane does not intersect the ellipsoid and cannot be divided.

[0073] C: Calculate the probability mass of the Gaussian distribution on the positive side and the negative side of the clipping plane in the standard space, and solve the statistical parameters of the two-sided sub-Gaussian kernel based on the truncated multivariate normal distribution theory.

[0074] Here, the conditional expectation and conditional covariance of the retained side are calculated in the standard space using the truncated expectation and covariance formula of the standard multivariate normal distribution on the half space, combined with the cumulative distribution function and probability density function of the one-dimensional standard normal distribution.

[0075] D: Transform the statistical parameters back to the original space to obtain the position and covariance matrix of the two sub-Gaussian kernels; weight the transparency according to the probability mass and inherit the color information of the original Gaussian kernel to generate two sub-Gaussian kernels.

[0076] Here, the above statistical parameters are restored to the original space through inverse affine transformation to obtain the new mean and new covariance matrix of the retained side and the rejected side sub-Gaussian kernel. According to the probability mass proportion of each side, the opacity of the sub-Gaussian kernel is scaled to maintain visual consistency during rendering. Finally, the original Gaussian kernel is split into two logically independent sub-Gaussian kernels: one on the positive side (retained side) of the clipping plane, participating in subsequent rendering and editing; the other on the negative side (rejected side), which can be discarded or temporarily stored according to application requirements.

[0077] Wherein, the ellipsoid parameters of the positive side of the divided plane are calculated:

[0078] Solve the following equation group:

[0079]

[0080] , , , , .

[0081] Here, ​is the offset of the center of the ellipsoid after segmentation in the positive direction of the plane, is the scaling ratio of the ellipsoid after segmentation, is the center position of the plane after segmentation, is the covariance matrix of the plane after segmentation, Q is a scaled covariance matrix.

[0082] where the ellipsoid parameters of the plane negative side after segmentation are calculated as

[0083] solving the following equations:

[0084]

[0085] is obtained , is , .

[0086] Here, is the offset of the center of the ellipsoid after segmentation in the negative direction of the plane, is the scaling ratio of the ellipsoid after segmentation, is the center position of the plane after segmentation, is the covariance matrix of the plane after segmentation.

[0087] where the Gaussian distribution function of the Gaussian kernel is as follows:

[0088]

[0089] is , , so , . is the sub-Gaussian kernel of the plane positive side, is the sub-Gaussian kernel of the plane negative side, is the transparency of the sub-Gaussian kernel of the plane positive side, is the transparency of the sub-Gaussian kernel of the plane negative side.

[0090] ​It should be noted that the existing method usually regards the Gaussian kernel as a point or a simple ellipsoid, and only according to the center position or the spatial surrounding relationship to carry out binary retention / deletion, resulting in density mutation and holes in the boundary area, and then causing jagged edges and visual noise in the rendering result. The present application introduces a "shaving operation", carries out statistical modeling and distribution reconstruction on the Gaussian kernel across the clipping plane in the half space, accurately calculates the new mean, covariance matrix and transparency of the retained side, so that the clipped Gaussian distribution still maintains smooth transition at the boundary, fundamentally avoiding the visual burr problem caused by distribution truncation, and significantly improving the visual integrity and realism of the clipping area.

[0091] S103: The sub-Gaussian kernels located on the rejection side are subjected to rejection processing, and the sub-Gaussian kernels located on the retention side are subjected to retention processing, and a clipped 3D GS sub-model is output.

[0092] In this step, all the retained side sub-Gaussian kernels are integrated according to the original index structure or the new organization mode to form a structurally complete clipped 3D GS sub-model. The sub-model inherits the attribute format (such as position, scale, rotation, transparency, spherical harmonic coefficient, etc.) of the original model, can be directly connected to the standard 3D GS rendering pipeline for real-time visualization, and has no visual burr, no geometric hole or mutation phenomenon in the clipping edge area.

[0093] It should be noted that the retained side subset includes all sub-Gaussian kernels located on the positive side of the clipping plane, and the rejection side subset includes all sub-Gaussian kernels located on the negative side of the clipping plane.

[0094] Here, all the sub-Gaussian kernels belonging to the rejection side are removed from the current model. The rejection operation is a logical deletion, and the parameter information is not retained in the final output. Since the spatial regions represented by these sub-kernels do not belong to the target editing range, their deletion will not affect the integrity of the retained area, but will help to reduce the amount of redundant data and improve the efficiency of subsequent rendering and editing. All sub-Gaussian kernels belonging to the retained side, including the complete Gaussian kernels that are originally located on the retained side and do not participate in splitting, and the truncated sub-kernels generated by the distribution level clipping operation are all retained in the result model. Among them, the sub-Gaussian kernels reconstructed by splitting have updated their mean, covariance matrix and opacity parameters to accurately reflect their probability distribution characteristics in the retained half space, ensuring the continuity and naturalness of color, density and geometric shape at the boundary.

[0095] In one possible implementation, for the clipping plane being a limited plane region, the burr-free clipping method further comprises:

[0096] (1) Select a point on an infinitely extended plane, construct two mutually orthogonal unit basis vectors, set the extension scale along the two unit basis vectors, and determine a rectangular limited plane region.

[0097] Here, an optional point on an infinite plane , on the basis of which two unit orthogonal bases and are constructed, and scales extending on the two orthogonal bases and are constructed, a rectangular finite plane region is constructed as .

[0098] It should be noted that the finite plane region can be rectangular, elliptical, or other plane shapes.

[0099] (2) detecting whether the 3D Gaussian kernel intersects with the finite plane region, and if so, performing a distribution-level plane clipping operation on the 3D Gaussian kernel to determine a clipped 3D GS sub-model.

[0100] In this way, whether the 3D Gaussian kernel intersects with the finite plane region is detected, and if so, a distribution-level plane clipping operation is performed on the 3D Gaussian kernel to determine a clipped 3D GS sub-model.

[0101] In one possible implementation, whether the 3D Gaussian kernel intersects with the finite plane region is determined by the following steps:

[0102] A normal projection point of the center of the 3D Gaussian kernel along the normal of the clipping plane is determined; it is judged whether the normal projection point falls within the finite plane region, and it is judged whether the spatial distribution represented by the covariance of the 3D Gaussian kernel intersects with the corresponding infinite plane; if both are true, the 3D Gaussian kernel intersects with the finite plane region; if either is false, the 3D Gaussian kernel does not intersect with the finite plane region.

[0103] Here, if the above two conditions are true at the same time, i.e., (1) the projection point falls within the finite region; (2) the spatial distribution of the Gaussian kernel intersects with the infinite plane, it is determined that the 3D Gaussian kernel intersects with the finite plane region and should participate in the subsequent distribution-level plane clipping operation (i.e., "razor operation"); if either condition is not true, it is determined that the Gaussian kernel does not intersect with the finite plane region, and the clipping process is skipped, remaining unchanged or directly retained.

[0104] In the present application, this double determination mechanism ensures that the clipping operation only acts on the Gaussian kernel that is truly in the user-specified local region and crosses the boundary, avoiding invalid calculation and interference outside the boundary, significantly improving editing accuracy and running efficiency. At the same time, since the continuous distribution characteristics of the Gaussian kernel across the boundary are preserved, even in the finite region, a high-quality clipping effect is achieved visually without burrs and geometrically continuously.

[0105] ​In one possible implementation, the non-penetrating clipping method further comprises, for a closed geometry composed of a plurality of clipping planes:

[0106] a: decomposing the closed geometry into a plurality of interior-facing limited clipping planes; wherein the normal vector of each limited clipping plane points to the interior of the closed geometry.

[0107] Here, a three-dimensional closed geometry is set, which can be an axis-aligned bounding box, an oriented bounding box, or other convex polyhedron surrounded by planes (such as a cube, prism, etc.). Each outer surface boundary of the geometry is decomposed into a corresponding limited clipping plane. Each limited clipping plane includes a unit normal vector, a intercept parameter, a base point, and two mutually orthogonal unit tangent vectors. Thus, each limited clipping plane retains the spatial division capability of an infinite plane, and its influence range is limited by local area constraints, forming an editing primitive with directionality and spatial selectivity.

[0108] b: arranging each limited clipping plane according to a preset clipping order, and sequentially performing distributed-level plane clipping operations on the 3DGS scene model to determine a 3DGS sub-model confined within the closed geometry.

[0109] Here, each limited clipping plane is arranged according to a preset clipping order, and distributed-level plane clipping operations are sequentially performed on the 3DGS scene model to determine a 3DGS sub-model confined within the closed geometry.

[0110] It should be noted that the preset clipping order includes a face-first order: first processing a pair of relatively parallel faces, and then sequentially processing planes in other directions; and a user-defined order: dynamically adjusting the clipping order according to the interactive intent to optimize the calculation efficiency or match a specific editing process.

[0111] In one possible implementation, the sequentially performing distributed-level plane clipping operations on the 3DGS scene model according to the preset clipping order to determine a 3DGS sub-model confined within the closed geometry comprises:

[0112] Taking each 3D Gaussian kernel in the 3DGS scene model as input, performing distributed-level plane clipping operations according to the geometric relationship between the covariance matrix of the 3D Gaussian kernel and the limited clipping plane to generate an intermediate 3DGS subset after this clipping; taking the intermediate 3DGS subset output by the current clipping step as input for the distributed-level plane clipping operation of the next limited clipping plane, and continuing to process the next limited clipping plane until the distributed-level plane clipping operation of each limited clipping plane is completed, to determine a 3DGS sub-model confined within the closed geometry.

[0113] In specific embodiments, the initial input is a complete original 3D GS scene model. For the k-th limited clipping plane, the following process is performed: taking the intermediate 3D GS subset output by the previous round as input, judging for each 3D Gaussian kernel whether it satisfies two conditions: (i) the spatial distribution represented by its covariance intersects with the limited clipping plane; (ii) its center is within the limited rectangular region along the normal of the limited clipping plane; if both conditions are satisfied, performing the distribution-level plane clipping operation (i.e. "razor operation") on the Gaussian kernel as described above, reconstructing its new mean value, new covariance and opacity on the retained side (i.e. inside the geometric body); deleting the part on the rejected side, retaining the updated sub-kernel, and outputting the intermediate 3D GS subset after this round of processing. Apply the above process to all limited clipping planes in turn to form a chain clipping process: after multiple rounds of iterative clipping, the final output model is a 3D GS sub-model strictly limited inside the closed geometric body. This sub-model has the following characteristics: all retained Gaussian kernels are located inside or near the boundary, and are reconstructed at the distribution level to ensure edge continuity; there is no "burr", void or density mutation phenomenon at the boundary caused by traditional methods; it is fully compatible with the standard 3D GS rendering pipeline, supporting real-time visualization and subsequent editing operations.

[0114] In one possible implementation, after completing the distribution-level plane clipping operation of the limited clipping plane, the burr-free clipping method further comprises:

[0115] I: dynamically updating the spatial distribution characteristics of the remaining 3D Gaussian kernel set.

[0116] Here, after each round of limited clipping plane processing is completed, the system automatically performs spatial statistical analysis on the current retained 3D GS sub-model to extract its latest geometric and density distribution characteristics. Mainly including: spatial barycenter, principal direction, local density distribution atlas and boundary proximity, as the basis for subsequent decision-making.

[0117] II: re-evaluating the processing priority of the subsequent limited clipping plane based on the updated distribution characteristics, determining the evaluation result, and adjusting the clipping order according to the evaluation result.

[0118] It should be noted that the evaluation criteria include: large intersection priority principle: preferentially processing the plane with the largest overlap area with the current Gaussian kernel distribution. For example, if there are a large number of Gaussian kernels across the border near a certain plane, it should be processed as soon as possible to reduce the subsequent redundant data amount; structure alignment priority principle: align the cropping order with the principal direction. For example, first crop the two opposite faces along the longest principal axis direction, which can quickly converge to the main structure; visual impact maximization principle: preferentially removing the external region containing the most visual energy (such as high opacity, strong color signal), so that the preview result tends to be stable faster; load balancing principle: avoid continuous processing of adjacent planes leading to local calculation intensive, and appropriately stagger planes in different directions to balance the GPU resource occupation. Each plane is assigned a dynamic priority score, which integrates the above factors and generates an evaluation result through weighted fusion.

[0119] Here, the system generates an updated execution sequence by dynamically reordering the limited cropping planes to be processed according to the newly calculated priority score. This process can be triggered after each cropping, or a threshold can be set to start only when the distribution changes significantly, to balance flexibility and overhead. The adjusted cropping order not only improves the overall processing efficiency (such as removing irrelevant regions earlier, reducing invalid projection judgment), but also better matches the user's editing intention. For example, in interactive editing, if the user stretches a bounding box in a certain direction, the system can immediately identify and preferentially process the new boundary plane in the expansion direction, achieving an "intention-aware" intelligent response.

[0120] In this application, the invention takes infinite planes as the basic cropping unit, and further expands to limited planes and rectangular bounding boxes and other practical cropping modes. By defining the base point, orthogonal direction and scale parameter, a rectangular cropping region of any direction and range can be constructed; at the same time, it supports face-by-face iterative cropping of axis-aligned or non-axis-aligned bounding boxes, meeting the user's demand for free-form region extraction. The architecture design is modular, logical and clear, easy to integrate into various 3D editing tools, with strong scalability and engineering practicality.

[0121] The embodiment of the present application provides a non-burr clipping method for 3DGS scene editing, and the non-burr clipping method comprises the following steps: acquiring a 3DGS scene model to be clipped which is composed of a plurality of 3D Gaussian kernels; defining at least one clipping plane of the 3DGS scene model; wherein the clipping plane is an infinitely extended plane, has a unit normal vector and an intercept parameter, and is used for dividing the 3DGS scene model into a reserved side and a rejected side; for each 3D Gaussian kernel in the 3DGS scene model, if the spatial distribution represented by the covariance of the 3D Gaussian kernel intersects with the clipping plane, then a distribution-level plane clipping operation is performed to split the 3D Gaussian kernel into two sub-Gaussian kernels which are located in the positive side and the negative side of the probability distribution part of the clipping plane respectively; the sub-Gaussian kernel located in the rejected side is subjected to a rejection processing, and the sub-Gaussian kernel located in the reserved side is subjected to a reservation processing, and a clipped 3DGS sub-model is output. The Gaussian kernel across the clipping plane is subjected to a distribution-level reconstruction to generate a new sub-Gaussian kernel, which not only reserves the spatial continuity of the original Gaussian distribution, but also effectively avoids the density jump and visual burr, realizes the true sense of "seamless clipping", and overcomes the problems of edge burr, geometric distortion, distribution rupture and the like in the existing 3DGS model clipping technology.

[0122] Please refer to Figure 2 、 Figure 3 , Figure 2 Figure 1 is a structural schematic diagram of a non-burr clipping device for 3DGS scene editing provided by the embodiment of the present application; Figure 3 Figure 2 is a structural schematic diagram of a non-burr clipping device for 3DGS scene editing provided by the embodiment of the present application. As shown in the figure, Figure 2 The non-burr clipping device for 3DGS scene editing 200 comprises:

[0123] The acquisition module 210 is used for acquiring a 3DGS scene model to be clipped which is composed of a plurality of 3D Gaussian kernels;

[0124] The definition module 220 is used for defining at least one clipping plane of the 3DGS scene model; wherein the clipping plane is an infinitely extended plane, has a unit normal vector and an intercept parameter, and is used for dividing the 3DGS scene model into a reserved side and a rejected side;

[0125] The clipping operation module 230 is used for, for each 3D Gaussian kernel in the 3DGS scene model, if the spatial distribution represented by the covariance of the 3D Gaussian kernel intersects with the clipping plane, then a distribution-level plane clipping operation is performed to split the 3D Gaussian kernel into two sub-Gaussian kernels which are located in the positive side and the negative side of the probability distribution part of the clipping plane respectively;

[0126] The determining module 240 is configured to perform rejection processing on the sub-Gaussian kernel located on the rejection side and perform retention processing on the sub-Gaussian kernel located on the retention side, and output a cropped 3D GS sub-model.

[0127] Further, the cropping operation module 230 is configured to split the 3D Gaussian kernel into two sub-Gaussian kernels by performing the distribution-level plane cropping operation.

[0128] Cholesky decomposition is performed on the covariance matrix of the current 3D Gaussian kernel to generate a lower triangular matrix.

[0129] Based on the lower triangular matrix, an affine transformation is constructed to map the current 3D Gaussian kernel to a standard space, and the cropping plane is projected to the standard space.

[0130] The probability mass of the Gaussian distribution on the positive side and the negative side of the cropping plane is calculated in the standard space, and the statistical parameters of the two sub-Gaussian kernels are solved based on the truncated multivariate normal distribution theory.

[0131] The statistical parameters are inversely transformed back to the original space to obtain the positions and covariance matrices of the two sub-Gaussian kernels.

[0132] The transparency is weighted and distributed according to the probability mass, and the color information of the original Gaussian kernel is inherited to generate two sub-Gaussian kernels.

[0133] Further, the cropping operation module 230 is configured to determine whether the 3D Gaussian kernel intersects with the limited plane region.

[0134] A point is selected on the infinitely extended plane, and two mutually orthogonal unit basis vectors are constructed. The extension scale along the two unit basis vectors is set to determine a rectangular limited plane region.

[0135] It is detected whether the 3D Gaussian kernel intersects with the limited plane region. If yes, the distribution-level plane cropping operation is performed on the 3D Gaussian kernel to determine a cropped 3D GS sub-model.

[0136] Further, the cropping operation module 230 determines whether the 3D Gaussian kernel intersects with the limited plane region by the following steps:

[0137] The orthogonal projection point of the center of the 3D Gaussian kernel along the normal of the cropping plane is determined.

[0138] It is judged whether the orthogonal projection point falls within the limited plane region, and whether the spatial distribution represented by the covariance of the 3D Gaussian kernel intersects with the corresponding infinitely extended plane.

[0139] If both are yes, the 3D Gaussian kernel intersects with the limited plane region.

[0140] If either is false, the 3D Gaussian kernel does not intersect the finite planar region.

[0141] Further, the clipping operation module 230 is configured to divide the closed geometry composed of a plurality of clipping planes into a plurality of interior-facing finite clipping planes, wherein a normal vector of each finite clipping plane points to the interior of the closed geometry.

[0142] Further, the clipping operation module 230 is configured to divide the closed geometry composed of a plurality of clipping planes into a plurality of interior-facing finite clipping planes, wherein a normal vector of each finite clipping plane points to the interior of the closed geometry.

[0143] Further, the clipping operation module 230 is configured to divide the closed geometry composed of a plurality of clipping planes into a plurality of interior-facing finite clipping planes, wherein a normal vector of each finite clipping plane points to the interior of the closed geometry.

[0144] Further, the clipping operation module 230 is configured to divide the closed geometry composed of a plurality of clipping planes into a plurality of interior-facing finite clipping planes, wherein a normal vector of each finite clipping plane points to the interior of the closed geometry.

[0145] Further, the clipping operation module 230 is configured to divide the closed geometry composed of a plurality of clipping planes into a plurality of interior-facing finite clipping planes, wherein a normal vector of each finite clipping plane points to the interior of the closed geometry.

[0146] Further, the clipping operation module 230 is configured to divide the closed geometry composed of a plurality of clipping planes into a plurality of interior-facing finite clipping planes, wherein a normal vector of each finite clipping plane points to the interior of the closed geometry.

[0147] Further, as shown in FIG. 2, the 3D GS scene editing non-stitching clipping device 200 further comprises an adjustment clipping order module 250, which is configured to: Figure 3

[0148] dynamically update the spatial distribution characteristics of the remaining 3D Gaussian kernel set;

[0149] based on the updated distribution characteristics, re-evaluate the processing priority of the subsequent finite clipping plane, determine an evaluation result, and adjust the clipping order according to the evaluation result.

[0150] ​This application provides a burr-free cropping device for 3DGS scene editing. The burr-free cropping device includes: an acquisition module for acquiring a 3DGS scene model to be cropped, composed of multiple 3D Gaussian kernels; a definition module for defining at least one cropping plane of the 3DGS scene model; wherein the cropping plane is an infinitely extending plane with a unit normal vector and an intercept parameter, used to divide the 3DGS scene model into a retention side and a culling side; a cropping operation module for performing a distribution-level plane cropping operation on each 3D Gaussian kernel in the 3DGS scene model, if the spatial distribution represented by the covariance of the 3D Gaussian kernel intersects with the cropping plane, splitting the 3D Gaussian kernel into two sub-Gaussian kernels, located on the positive and negative probability distribution parts of the cropping plane respectively; and a determination module for performing culling processing on the sub-Gaussian kernels located on the culling side, retaining processing on the sub-Gaussian kernels located on the retention side, and outputting the cropped 3DGS sub-model. By performing distribution-level reconstruction on the Gaussian kernel across the clipping plane to generate new sub-Gaussian kernels, not only is the spatial continuity of the original Gaussian distribution preserved, but density jumps and visual jaggedness are also effectively avoided, achieving true "seamless clipping" and overcoming problems such as edge jaggedness, geometric distortion, and distribution breakage in existing 3DGS model clipping techniques.

[0151] Please see Figure 4 , Figure 4 This is a schematic diagram of the structure of an electronic device provided in an embodiment of this application. Figure 4 As shown, the electronic device 400 includes a processor 410, a memory 420, and a bus 430.

[0152] The memory 420 stores machine-readable instructions executable by the processor 410. When the electronic device 400 is running, the processor 410 communicates with the memory 420 via the bus 430. When the machine-readable instructions are executed by the processor 410, they can perform the operations described above. Figure 1 The steps of the burr-free cropping method for 3DGS scene editing in the method embodiment shown are described in detail in the method embodiment, and will not be repeated here.

[0153] This application also provides a computer-readable storage medium storing a computer program, which, when executed by a processor, can perform the above-described actions. Figure 1 The steps of the burr-free cropping method for 3DGS scene editing in the method embodiment shown are described in detail in the method embodiment, and will not be repeated here.

[0154] Those skilled in the art can clearly understand that, for the convenience and brevity of description, the specific working process of the system, device and unit described above can refer to the corresponding process in the foregoing method embodiments, which will not be repeated here.

[0155] In several embodiments provided in the present application, it should be understood that the disclosed system, device and method can be implemented in other ways. The device embodiments described above are only schematic, for example, the division of the units is only a logical function division, and actual implementation can have another division manner, for example, a plurality of units or components can be combined or integrated into another system, or some features can be ignored or not executed. In addition, the coupling or direct coupling or communication connection between the units shown or discussed can be indirect coupling or communication connection through some communication interfaces, devices or units, and can be electrical, mechanical or other forms.

[0156] The units described as separate components can or can not be physically separate, and the components shown as units can or can not be physical units, that is, they can be located in one place, or can be distributed on a plurality of network units. Some or all of the units can be selected according to actual needs to achieve the purpose of the embodiment scheme.

[0157] In addition, the functional units in each embodiment of the present application can be integrated in one processing unit, or each unit can be physically present separately, or two or more units can be integrated in one unit.

[0158] If the functions are realized in the form of software function units and sold or used as independent products, they can be stored in a nonvolatile computer readable storage medium executable by a processor. Based on this understanding, the technical solutions of the present application or the part of the present application that essentially contributes to the prior art or the part of the technical solutions can be embodied in the form of a software product, which is stored in a storage medium and includes a plurality of instructions for causing a computer device (which can be a personal computer, a server, or a network device, etc.) to execute all or part of the steps of the method described in each embodiment of the present application. The foregoing storage medium includes: a U disk, a mobile hard disk, a read-only memory (ROM), a random access memory (RAM), a magnetic disk or an optical disk, and various program code storage media.

[0159] Finally, it should be noted that the above-described embodiments are merely specific embodiments of the present application, which are used to illustrate the technical solutions of the present application, but not to limit the same. The protection scope of the present application is not limited thereto. Although the present application has been described in detail with reference to the foregoing embodiments, it should be understood by those skilled in the art that any skilled person in the art can still modify or easily think of changes to the technical solutions recorded in the foregoing embodiments, or make equivalent replacements to some of the technical features, within the technical scope disclosed by the present application. The modifications, changes or replacements do not make the corresponding technical solutions deviate from the spirit and scope of the technical solutions of the embodiments of the present application, and should be covered within the protection scope of the present application. Therefore, the protection scope of the present application should be subject to the protection scope of the claims.

Claims

1. A no-pen-cropping method for 3D GS scene editing, characterized in that, The burr-free clipping method comprises: acquiring a 3D Gaussian kernel composed of a 3D Gaussian scene model to be clipped; defining at least one clipping plane of the 3D Gaussian scene model; wherein the clipping plane is an infinitely extended plane, has a unit normal vector and an intercept parameter, and is used to divide the 3D Gaussian scene model into a reserved side and a discarded side; for each 3D Gaussian kernel in the 3D Gaussian scene model, if the spatial distribution represented by the covariance of the 3D Gaussian kernel intersects with the clipping plane, performing a distribution-level plane clipping operation to split the 3D Gaussian kernel into two sub-Gaussian kernels, which are located in the positive side and the negative side of the probability distribution part of the clipping plane, respectively; performing a discard processing on the sub-Gaussian kernel located in the discarded side, and performing a reserved processing on the sub-Gaussian kernel located in the reserved side, and outputting a clipped 3D Gaussian sub-model.

2. The burr-free trimming method according to claim 1, characterized by, The distribution-level plane clipping operation comprises: performing a Cholesky decomposition on the covariance matrix of the current 3D Gaussian kernel to generate a lower triangular matrix; constructing an affine transformation based on the lower triangular matrix to map the current 3D Gaussian kernel to a standard space, and synchronously project the clipping plane to the space; calculating the probability mass of the Gaussian distribution on the positive side and the negative side of the clipping plane in the standard space, and solving the statistical parameters of the two-side sub-Gaussian kernels based on the truncated multivariate normal distribution theory; inverse transforming the statistical parameters back to the original space to obtain the positions and covariance matrices of the two sub-Gaussian kernels; weighting and allocating the transparency according to the probability mass, and inheriting the color information of the original Gaussian kernel to generate two sub-Gaussian kernels.

3. The burr-free trimming method of claim 1, wherein For the clipping plane being a limited plane area, the burr-free clipping method further comprises: selecting a point on the infinitely extended plane, and constructing two mutually orthogonal unit basis vectors, setting the extension scale along the two unit basis vectors to determine a rectangular limited plane area; detecting whether the 3D Gaussian kernel intersects with the limited plane area, and if yes, performing a distribution-level plane clipping operation on the 3D Gaussian kernel to determine a clipped 3D Gaussian sub-model.

4. The burr-free trimming method according to claim 3, characterized by The 3D Gaussian kernel is determined to intersect with the limited plane area by the following steps: determining the orthogonal projection point of the center of the 3D Gaussian kernel along the normal of the clipping plane: judging whether the orthogonal projection point falls within the limited plane area, and judging whether the spatial distribution represented by the covariance of the 3D Gaussian kernel intersects with the corresponding infinitely extended plane; if both are yes, the 3D Gaussian kernel intersects with the limited plane area; if any one is no, the 3D Gaussian kernel does not intersect with the limited plane area.

5. The burr-free trimming method of claim 1, wherein For a closed geometric body composed of multiple clipping planes, the burr-free clipping method further comprises: decomposing the closed geometric body into multiple internally facing limited clipping planes; wherein the normal vector of each limited clipping plane points to the interior of the closed geometric body; The 3DGS scene model is sequentially subjected to the distribution-level plane clipping operation according to a preset clipping order, and a 3DGS sub-model confined in the closed geometry is determined.

6. The burr-free trimming method according to claim 5, characterized in that, The distribution-level plane clipping operation is sequentially performed on the 3DGS scene model according to a preset clipping order, and a 3DGS sub-model confined in the closed geometry is determined. Each 3D Gaussian kernel in the 3DGS scene model is taken as input, and a distribution-level plane clipping operation is performed according to the geometric relationship between the covariance matrix of the 3D Gaussian kernel and the limited clipping plane, to generate an intermediate 3DGS subset after this clipping; The intermediate 3DGS subset output by the current clipping step is taken as input for the distribution-level plane clipping operation of the next limited clipping plane, and the next limited clipping plane is processed, until the distribution-level plane clipping operation of each limited clipping plane is completed, and a 3DGS sub-model confined in the closed geometry is determined.

7. The burr-free trimming method according to claim 6, characterized in that, After the distribution-level plane clipping operation of the limited clipping plane is completed, the no-stitch clipping method further comprises: dynamically updating the spatial distribution characteristics of the remaining 3D Gaussian kernel set; re-evaluating the processing priority of the subsequent limited clipping plane based on the updated distribution characteristics, determining an evaluation result, and adjusting the clipping order according to the evaluation result.

8. A device for 3D GS scene editing without burr clipping, characterized in that, The no-stitch clipping device comprises: an acquisition module configured to acquire a 3DGS scene model to be clipped composed of a plurality of 3D Gaussian kernels; a definition module configured to define at least one clipping plane of the 3DGS scene model; wherein the clipping plane is an infinitely extended plane having a unit normal vector and an intercept parameter, and is used to divide the 3DGS scene model into a retained side and a discarded side; a clipping operation module configured to, for each 3D Gaussian kernel in the 3DGS scene model, if the spatial distribution represented by the covariance of the 3D Gaussian kernel intersects the clipping plane, perform a distribution-level plane clipping operation to split the 3D Gaussian kernel into two sub-Gaussian kernels located in the positive and negative probability distribution parts of the clipping plane, respectively; a determination module configured to discard the sub-Gaussian kernel located on the discarded side and retain the sub-Gaussian kernel located on the retained side, and output a clipped 3DGS sub-model.

9. An electronic device, comprising: comprises: a processor, a memory and a bus, the memory storing machine-readable instructions executable by the processor, when the electronic device is running, the processor and the memory communicate through the bus, the machine-readable instructions are executed by the processor to execute the steps of the no-stitch clipping method for 3DGS scene editing as claimed in any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, The computer program is stored on the computer readable storage medium, and the computer program is executed by the processor to execute the steps of the no-stitch clipping method for 3DGS scene editing as claimed in any one of claims 1 to 7.

Images