Three-dimensional model editing method and device, storage medium and program product

Abstract

The invention discloses a three-dimensional model editing method and device, a storage medium and a program product, and relates to the technical field of image processing, and the method comprises the steps: carrying out the preliminary non-rigid editing of sparse anchor points, thereby obtaining the anchor points after non-rigid editing, and enabling the anchor points after non-rigid editing to represent the characteristics of a target state, then, the random Gaussian points are initialized based on the non-rigid edited anchor points, more Gaussian points can be obtained after densification, the appearance of the target object can be reflected more accurately, and the obtained dense Gaussian points of the target can accurately describe the appearance characteristics of the target object, so that the accuracy of the target object is improved. Therefore, the dense three-dimensional model generated by the target dense Gaussian point is more accurate, the dense three-dimensional model is the target three-dimensional model at the moment, and the anchor point after non-rigid editing reflects from the source state to the target state, so that the target object in the target three-dimensional model is changed from the source state to the target state, and the target object in the target three-dimensional model is changed from the target state to the target state. Therefore, the dynamic change is reflected.

Description

Technical Field

[0001] The present application relates to the field of image processing technology, and in particular to a three-dimensional model editing method, device, storage medium and program product. Background Art

[0002] Nowadays, the construction of three-dimensional models is widely used in film production and games, and building high-quality three-dimensional models has become a research goal for researchers.

[0003] In the related art, the objects in the source 3D model are edited by text-driven editing to obtain the edited image, thereby obtaining the final 3D model. However, when the objects in the source 3D model are edited by text-driven editing in the related art, only the color thereof can be replaced. In the face of complex dynamic application scenarios, the requirements for dynamic changes cannot be met, and the quality and accuracy of the 3D model are low. Summary of the invention

[0004] The present application provides a three-dimensional model editing method, device, storage medium and program product to at least solve the problem that the related technology cannot meet the needs of dynamic application scenarios and the quality and accuracy of the three-dimensional model are low.

[0005] In a first aspect, the present application provides a three-dimensional model editing method, comprising:

[0006] Sparse anchor points are obtained from a source three-dimensional model; the source three-dimensional model includes a target object; the target object is in a source state; the sparse anchor points are anchor points that describe geometric features of the target object;

[0007] Perform preliminary non-rigid editing based on the sparse anchor points to obtain non-rigidly edited anchor points; the non-rigidly edited anchor points are anchor points of the target object that form a target state from a source state after non-rigid editing;

[0008] Initializing random Gaussian points based on the anchor points after non-rigid editing to obtain initial Gaussian points; the number of random Gaussian points is greater than the number of sparse anchor points; the random Gaussian points are part of the source Gaussian points in the source three-dimensional model;

[0009] The initial Gaussian points are densified to determine target densed Gaussian points that describe the appearance features of the target object; the densified three-dimensional model generated based on the target densed Gaussian points is the target three-dimensional model, and the target object in the target three-dimensional model forms a target state from a source state after performing a target action.

[0010] In a second aspect, the present application also provides an editing device, comprising: a memory for storing a computer program; and a processor for implementing the steps of any one of the above-mentioned three-dimensional model editing methods when executing the computer program.

[0011] In a third aspect, the present application further provides a computer-readable storage medium, in which a computer program is stored, wherein the computer program, when executed by a processor, implements the steps of any of the above-mentioned three-dimensional model editing methods.

[0012] In a fourth aspect, the present application also provides a computer program product, including a computer program, which implements the steps of any of the above-mentioned three-dimensional model editing methods when executed by a processor.

[0013] The present application provides a three-dimensional model editing method, which first obtains sparse anchor points from a source three-dimensional model, including a target object in the source three-dimensional model. The sparse anchor points can describe the geometric features of the target object, and then the sparse anchor points are preliminarily non-rigidly edited to obtain the non-rigidly edited anchor points. It can be seen that the non-rigidly edited anchor points at this time characterize the characteristics of the target state, and then the random Gaussian points are initialized based on the non-rigidly edited anchor points to obtain initial Gaussian points, and then the initial Gaussian points are densified. Since more Gaussian points can be obtained due to densification, the appearance of the target object can be more accurately reflected. When a certain density is reached, After the target densified Gaussian points are obtained, the target densified Gaussian points can accurately describe the appearance characteristics of the target object, and then the dense three-dimensional model generated by the target densified Gaussian points is more accurate. At this time, the dense three-dimensional model is the target three-dimensional model. In the target three-dimensional model, the target densified Gaussian points are obtained by densifying the initial Gaussian points, and the initial Gaussian points are obtained based on the anchor points after non-rigid editing, so that the target object in the target three-dimensional model is formed from the source state to the target state, and then the dynamic change of the three-dimensional model is reflected, so that this scheme can realize the change in the state of the target object and can be used in replicating dynamic application scenarios. It should be noted that the number of random Gaussian points is greater than the number of sparse anchor points, thereby ensuring the number of initial Gaussian points, effectively controlling the number of Gaussian points after densification, and improving the accuracy of the target three-dimensional model. In addition, in the present application, sparse anchor points can describe the geometric features of the target object, and target densified Gaussian points can describe the appearance features of the target object. In the present application, the two features are combined to achieve decoupling of geometry and appearance, and the geometry and appearance of the target object are accurately characterized in a sparse geometry-dense appearance manner. Among them, the use of sparse anchor points facilitates geometric transformation after preliminary non-rigid editing. At the same time, the target densified Gaussian points have improved quality due to their density. In the present application, densification is performed on the basis of the anchor points after non-rigid editing to obtain target dense Gaussian points that characterize the appearance features of the target object, which can accurately capture the geometric features and appearance features of the target object, making the target three-dimensional model more accurate. BRIEF DESCRIPTION OF THE DRAWINGS

[0014] In order to more clearly illustrate the embodiments of the present application, the following is a brief introduction to the drawings required for use in the embodiments. Obviously, the drawings described below are only some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0015] Figure 1 A schematic diagram of a 3D model editing scene provided for this application;

[0016] Figure 2 A schematic diagram of a three-dimensional model editing method provided in Example 1;

[0017] Figure 3 A schematic diagram of a flow chart of a three-dimensional model editing method provided in Example 2;

[0018] Figure 4 A schematic flow chart of a three-dimensional model editing method provided in Example 4;

[0019] Figure 5 A schematic diagram of a three-dimensional model editing method provided in Example 5;

[0020] Figure 6 A schematic flow chart of a three-dimensional model editing method provided in Example 6;

[0021] Figure 7 A schematic flow chart of a three-dimensional model editing method provided in Example 7;

[0022] Figure 8 A schematic flow chart of a three-dimensional model editing method provided in Example 10;

[0023] Fig. 9 A schematic flow chart of a three-dimensional model editing method provided in Example 12;

[0024] Fig.10 A schematic diagram of the structure of a three-dimensional model editing device provided in Example 13;

[0025] Fig.11 This is a schematic diagram of the structure of the editing device provided in Example 13. DETAILED DESCRIPTION

[0026] The following will be combined with the drawings in the embodiments of the present application to clearly and completely describe the technical solutions in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, not all of the embodiments. Based on the embodiments in the present application, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of this application.

[0027] It should be noted that, in the description of this application, the terms "include", "comprise" or any other variant thereof are intended to cover non-exclusive inclusion, so that a process, method, article or device including a series of elements includes not only those elements, but also includes other elements not explicitly listed, or also includes elements inherent to such process, method, article or device. The terms "first", "second", etc. in this application are used to distinguish similar objects, and are not used to describe a specific order or sequence.

[0028] In order to enable those skilled in the art to better understand the present application, the present application is further described in detail below in conjunction with the accompanying drawings and specific implementation methods.

[0029] In conjunction with the specific application environment architecture or the specific hardware architecture on which the execution of the three-dimensional model editing method depends, the specific application environment architecture or the specific hardware architecture is described herein.

[0030] Nowadays, the construction of three-dimensional models is widely used in film production and games, and building high-quality three-dimensional models has become a research goal for researchers.

[0031] In the related art, the objects in the source 3D model are edited by text-driven editing to obtain the target 3D model. However, when the objects in the source 3D model are edited by text-driven editing in the related art, only the color thereof can be replaced, and in the face of complex dynamic application scenarios, the demand for dynamic changes cannot be met.

[0032] In order to solve the defects of the related art, the inventor of this solution designed a new solution after creative research. The present invention provides a three-dimensional model editing method. In order to solve the problem that the related technology cannot meet the dynamic change requirements, the present invention performs preliminary non-rigid editing on the sparse anchor points to obtain the non-rigidly edited anchor points. The non-rigidly edited anchor points reflect the anchor points of the target object forming the target state from the source state. Then, based on the non-rigidly edited anchor points, further initialization and densification can be performed to obtain the target densified Gaussian points. The densified three-dimensional model generated by the target densified Gaussian points is the target three-dimensional model. The target object in the target three-dimensional model forms the target state from the source state after executing the target action. Therefore, the state of the target object in the target three-dimensional model changes, thereby realizing non-rigid editing in the present application and meeting the dynamic change requirements. In addition, in order to solve the problem that the three-dimensional model quality is poor and the accuracy is low in the related technology, in the present application, sparse anchor points are obtained from the source three-dimensional model. The sparse anchor points can describe the geometric features of the target object. Then, the non-rigidly edited anchor points obtained by the preliminary non-rigid editing of the sparse anchor points are initialized to obtain the initial Gaussian points. The initial Gaussian points are densified again, so that more can be obtained. Gaussian points are obtained in the densification process, and more densified Gaussian points are obtained until the target densified Gaussian points are obtained. The target densified Gaussian points improve the density. Since the target densified Gaussian points can describe the appearance characteristics of the target object, the target 3D model generated based on the target densified Gaussian points is more accurate in appearance. At the same time, since the target densified Gaussian points are obtained on the basis of sparse anchor points, the target 3D model is more accurate in geometry. Therefore, this scheme realizes the 3D model with decoupled geometry and appearance, which specifically adopts sparse geometry-dense The target object is decomposed into sparse anchor points and target densified Gaussian points obtained by densification in an appearance-based manner. The sparse anchor points can realize non-rigid editing, and the target densified Gaussian points improve the rendering quality due to their density. At the same time, this scheme uses sparse anchor points as basic control points and uses densification to optimize geometry and appearance, thereby realizing non-rigid editing. In summary, this scheme can improve the accuracy and quality of the target 3D model in the process of forming a target state from a source state.

[0033] Figure 1 A schematic diagram of a 3D model editing scene provided for this application, such as Figure 1 As shown, the specific application scenario of the present application includes an editing device 101.

[0034] The editing device 101 may be a computer or a server, which is not limited here.

[0035] The editing device 101 includes a three-dimensional model editing apparatus.

[0036] In this application scenario, the editing device 101 obtains sparse anchor points and random Gaussian points from the source three-dimensional model, and then performs preliminary non-rigid editing on the sparse anchor points to obtain non-rigidly edited anchor points.

[0037] Furthermore, the editing device 101 initializes the random Gaussian points based on the anchor points after non-rigid editing to obtain initial Gaussian points.

[0038] Furthermore, the editing device 101 densifies the initial Gaussian points and determines target densified Gaussian points, wherein the densified three-dimensional model generated by the target densified Gaussian points is the target three-dimensional model.

[0039] In this application, the source three-dimensional model is the source state, and the sparse anchor points are preliminarily non-rigidly edited, thereby realizing the target action of the target object, forming the target state from the source state, and obtaining the anchor points after non-rigid editing. The densified three-dimensional model generated based on the dense target densified Gaussian points is the target three-dimensional model. The target object in the target three-dimensional model forms the target state from the source state, so dynamic changes are realized. In addition, the construction of the target three-dimensional model is realized by using the sparse geometry-dense appearance method. In the process of constructing the target three-dimensional model, the geometric features and appearance features of the target object are fully considered, so that the target object in the target three-dimensional model not only changes in state, but also the target three-dimensional model is more accurate in geometry and appearance, and has better quality.

[0040] In the present application, the target three-dimensional model is generated using target densified Gaussian points and optimized anchor points, wherein the optimized anchor points are obtained by keeping the positions of the anchor points after non-rigid editing unchanged and optimizing their properties, and thus the optimized anchor points can describe the geometric features of the target object in the target state, and the dense target densified Gaussian points can describe the appearance features of the target object, combining the geometric features and the appearance features to generate a target three-dimensional model, thereby making the target three-dimensional model more accurate in terms of combined features and appearance features, and having higher quality.

[0041] The embodiment of the present application provides a three-dimensional model editing method, and the method is described in detail in conjunction with the execution flow of the three-dimensional model editing method.

[0042] Embodiment 1

[0043] The execution body of Embodiments 1 to 12 of the present application is a three-dimensional model editing device (referred to as editing device), which is located in the editing equipment.

[0044] Figure 2 A schematic diagram of a three-dimensional model editing method provided in Example 1. Figure 2 As shown, including:

[0045] S201, obtaining sparse anchor points from a source three-dimensional model; the source three-dimensional model includes a target object; the target object is in a source state; the sparse anchor points are anchor points that describe geometric features of the target object.

[0046] The source three-dimensional model refers to a three-dimensional model of a target object to be edited.

[0047] The source state refers to the posture of the target object in the source 3D model or source image. For example, if the source 3D model includes a standing cat, the source state is "standing".

[0048] In one approach, sparse anchor points may be obtained from a source 3D model via a furthest distance sampling mechanism.

[0049] S202, performing preliminary non-rigid editing based on the sparse anchor points to obtain non-rigidly edited anchor points; the non-rigidly edited anchor points are anchor points of the target object that are formed from the source state to the target state after non-rigid editing.

[0050] Among them, the preliminary non-rigid editing is to perform rough non-rigid editing on the sparse anchor points, and a non-rigid transformation field is constructed through the preliminary non-rigid editing.

[0051] In one approach, preliminary non-rigid editing is performed based on sparse anchor points and a multi-layer perception model, and finally non-rigidly edited anchor points are obtained.

[0052] S203, initializing random Gaussian points based on the anchor points after non-rigid editing to obtain initial Gaussian points; the random Gaussian points are part of the source Gaussian points in the source three-dimensional model.

[0053] Among them, the number of random Gaussian points is greater than the number of sparse anchor points. Since the number of random Gaussian points is greater than the number of sparse anchor points, the number of initial Gaussian points has a certain advantage, which effectively controls the number of target densified Gaussian points and improves density.

[0054] The source three-dimensional model includes a plurality of source Gaussian points. For example, assuming that there are 1000 source Gaussian points, a random mechanism is used to determine a plurality of source Gaussian points from the 1000 source Gaussian points as random Gaussian points.

[0055] It should be noted that the random Gaussian points used in this application are Gaussian points in the source three-dimensional model, not completely random Gaussian points, which can improve the controllability of initialization of random Gaussian points and ensure the stability of the initial Gaussian points.

[0056] S204, densifying the initial Gaussian points to determine target densed Gaussian points that describe the appearance features of the target object; the densified three-dimensional model generated based on the target densed Gaussian points is the target three-dimensional model, and the target object in the target three-dimensional model forms a target state from a source state after performing a target action.

[0057] In one approach, the initial Gaussian points may be densified by gradually densifying using a Gaussian point replication method and an iterative densification method, thereby obtaining target densified Gaussian points.

[0058] It should be noted that the densified three-dimensional model generated based on the target densified Gaussian is the target three-dimensional model, and the target object in the target three-dimensional model at this time is in the target state.

[0059] The present embodiment provides a 3D model editing method. In the present embodiment, sparse anchor points are first obtained from a source 3D model, where the source 3D model includes a target object. The sparse anchor points can describe the geometric features of the target object. Then, the sparse anchor points are preliminarily non-rigidly edited to obtain non-rigidly edited anchor points. It can be seen that the non-rigidly edited anchor points represent the characteristics of the target state. Then, random Gaussian points are initialized based on the non-rigidly edited anchor points to obtain initial Gaussian points. Then, the initial Gaussian points are densified. Since more Gaussian points can be obtained by densification, the appearance of the target object can be more accurately reflected. When a certain density is reached, the initial Gaussian points are initialized to obtain initial Gaussian points. After a certain degree, the target densified Gaussian points are obtained, which can accurately describe the appearance characteristics of the target object, and then the dense three-dimensional model generated by the target dense Gaussian points is more accurate. At this time, the dense three-dimensional model is the target three-dimensional model. In the target three-dimensional model, since the target dense Gaussian points are obtained by densifying the initial Gaussian points, and the initial Gaussian points are obtained based on the anchor points after non-rigid editing, the target object in the target three-dimensional model is formed from the source state to the target state, and then the dynamic change of the three-dimensional model is reflected, so that this scheme can realize the change in the state of the target object and can be used in complex dynamic application scenarios. It should be noted that the number of random Gaussian points is greater than the number of sparse anchor points, thereby ensuring the number of initial Gaussian points, effectively controlling the number of Gaussian points after densification, and improving the accuracy of the target three-dimensional model. In addition, in this embodiment, sparse anchor points can describe the geometric features of the target object, and target densified Gaussian points can describe the appearance features of the target object. In this embodiment, the two features are combined to achieve decoupling of geometry and appearance, and the geometry and appearance of the target object are accurately characterized in a sparse geometry-dense appearance manner, wherein the use of sparse anchor points facilitates geometric transformation after preliminary non-rigid editing, and at the same time, the target densified Gaussian points have improved quality due to their density; in this embodiment, densification is performed on the basis of the anchor points after non-rigid editing to obtain target densified Gaussian points that characterize the appearance features of the target object, which can accurately capture the geometric features and appearance features of the target object, making the target three-dimensional model more accurate.

[0060] Embodiment 2

[0061] This embodiment is a further refinement of the first embodiment. This embodiment is an optional way to obtain sparse anchor points from a source three-dimensional model.

[0062] Figure 3 A schematic diagram of a three-dimensional model editing method provided in Example 2. Figure 3 As shown, specifically including:

[0063] S301, determining a plurality of source Gaussian points included in a source three-dimensional model.

[0064] Specifically, a three-dimensional Gaussian is used to scan the source three-dimensional model to obtain a plurality of source Gaussian points included therein.

[0065] S302, calculating a first distance; the first distance is the distance between source Gaussian points calculated based on the three-dimensional coordinates of the source Gaussian points.

[0066] Specifically, the distance between each source Gaussian point is calculated.

[0067] Specifically, the distance between any two source Gaussian points is calculated respectively.

[0068] When calculating the first distance, the three-dimensional coordinates of each source Gaussian point are used for calculation.

[0069] S303, determining the source Gaussian points corresponding to the first distance that satisfies the preset farthest distance sampling mechanism as sparse anchor points; the number of sparse anchor points is less than the number of source Gaussian points; the sparse anchor points are greater than or equal to the first preset number.

[0070] Specifically, according to the preset farthest distance sampling mechanism, a first distance satisfying the preset farthest distance sampling mechanism is determined from the calculated multiple first distances, and then a source Gaussian point corresponding to the first distance is determined, and then the source Gaussian point is determined as a sparse anchor point.

[0071] Exemplarily, assume that there are five source Gaussian points, namely source Gaussian point 1, source Gaussian point 2, source Gaussian point 3, source Gaussian point 4 and source Gaussian point 5, and then calculate the first distance, specifically: calculate the distance between source Gaussian point 1 and source Gaussian point 2, expressed as L12. Similarly, calculate the remaining first distances according to the above method, which are L13, L14, L15, L23, L24, L25, L34, L35 and L45. Determine the first distance that meets the preset longest distance sampling mechanism from the above multiple first distances, assumed to be L13 and L14, and then determine the source Gaussian points corresponding to L13 and L14 as sparse anchor points.

[0072] It should be noted that the number of sparse anchor points also needs to meet certain requirements to ensure the accuracy of the preliminary non-rigid editing of the sparse anchor points. In this embodiment, the sparse anchor points are greater than or equal to a first preset number, wherein the first preset number is a preset number of sparse anchor points.

[0073] It should be noted that in this embodiment, the obtained sparse anchor points are used as basic control points. Since the sparse anchor points are anchor points that describe the geometric features of the target object, based on the basic control points, the geometric features of the target object can be guaranteed in the subsequent process, thereby improving accuracy.

[0074] The present embodiment provides a three-dimensional model editing method. In the present embodiment, a first distance and a preset maximum distance are used to adopt a mechanism to accurately determine sparse anchor points. In addition, in the present embodiment, the sparse anchor points are ensured to be greater than or equal to a first preset number, thereby ensuring that in the subsequent target three-dimensional model generation, the accuracy of the target three-dimensional model is improved.

[0075] Embodiment 3

[0076] This embodiment is a further refinement of any of the above embodiments. This embodiment is an optional method for performing preliminary non-rigid editing based on sparse anchor points to obtain anchor points after non-rigid editing, specifically including:

[0077] The DOF transformation algorithm is iteratively optimized to perform preliminary non-rigid editing on the sparse anchor points to obtain the anchor points after non-rigid editing; the DOF transformation algorithm is used to obtain the target rotation matrix and the target translation vector, and the anchor points after non-rigid editing are obtained based on the target rotation matrix and the target translation vector. The DOF transformation algorithm is iteratively optimized in the preliminary non-rigid editing to obtain the target rotation matrix and the target translation vector.

[0078] Specifically, the degree of freedom transformation algorithm in the multi-layer perception model is used to perform preliminary non-rigid editing of sparse anchor points, and the parameters included in the degree of freedom transformation algorithm are iteratively optimized so that the editing loss value of the edited three-dimensional model generated based on the rotation matrix and translation vector output by the optimized degree of freedom transformation algorithm meets the preset editing loss condition, and the iteratively optimized anchor points corresponding to the edited three-dimensional model that meets the preset editing loss condition are determined as the final non-rigid edited anchor points. At the same time, the rotation matrix and translation vector that meet the preset editing loss condition are determined as the target rotation matrix and target translation vector, wherein the edited three-dimensional model is generated based on the non-rigid edited anchor points, and wherein the non-rigid edited anchor points are generated based on the non-rigid editing anchor point algorithm.

[0079] The degree of freedom transformation algorithm is shown in formula (1):

[0080] (1)

[0081] in, represents the rotation matrix, ; represents the translation vector, , represents sparse anchor points, i.e., anchor points before preliminary non-rigid editing, represents the degrees of freedom transformation algorithm, Represents the parameters in the degrees of freedom transformation algorithm.

[0082] The non-rigid editing anchor point algorithm is shown in formula (2):

[0083] (2)

[0084] in, Represents the anchor point for iterative optimization.

[0085] For example, the sparse anchor points are initially non-rigidly edited. At the first iteration, the parameters in the degree of freedom transformation algorithm are is the initial value (denoted as ),Will Input to In, get ,Will Input into the non-rigid editing anchor point algorithm, and get , based on If the editing loss value of the edited 3D model generated based on the source 3D model does not meet the preset editing loss condition, iterative optimization is performed based on the sparse anchor points, where: is a sparse anchor point. In the second iteration, the parameter From the initial value Optimized for , and then Input into the second iterative optimization degree of freedom transformation algorithm, and get ; At this time, Input into the non-rigid editing anchor point algorithm, and get ,based on The edit loss value of the generated edited 3D model does not meet the preset edit loss condition. Since the preset edit loss condition is not met, the parameters are further adjusted. Iterative optimization, in the third iteration, the parameters from Optimized for ,Will Input into the third iterative optimization degree of freedom transformation algorithm, and get , at this time will Input into the non-rigid editing anchor point algorithm, and get , assuming that based on The editing loss value of the generated edited 3D model meets the preset editing loss condition, and the iteration ends. At this time, the target rotation matrix (i.e. ) and the target translation vector (i.e. ), at this time, the anchor point after non-rigid editing is , it can be seen that is obtained based on the target rotation matrix and the target translation vector; it should be noted that if in the third iteration, based on The editing loss value of the generated edited 3D model does not meet the preset editing loss condition, and the iterative optimization parameters are continued. , until the editing loss value of the edited three-dimensional model generated based on the iteratively optimized anchor point meets the preset editing loss condition, the iteration ends, and the iteratively optimized anchor point corresponding to the preset editing loss condition is determined as the anchor point after non-rigid editing.

[0086] Among them, the preset editing loss condition is a pre-set loss condition for performing preliminary non-rigid editing on sparse anchor points, and the preset editing loss condition is a loss condition considered at a standard resolution.

[0087] Among them, based on The generated edited 3D model can be based on The edited 3D model is generated based on the edited 3D model of the previous iteration, wherein the previous iteration is the first iteration.

[0088] The present embodiment provides a three-dimensional model editing method. In the present embodiment, a degree of freedom transformation algorithm is used to obtain a target rotation matrix and a target translation vector, so that preliminary non-rigid editing can be performed based on the target rotation matrix and the target translation vector, and then the anchor point after non-rigid editing is obtained. In the present embodiment, the degree of freedom transformation algorithm can be iteratively optimized to obtain a better target rotation matrix and target translation vector, so that a better anchor point after non-rigid editing is obtained, thereby improving accuracy.

[0089] Embodiment 4

[0090] This embodiment is a further refinement of any of the above embodiments. This embodiment is an optional way to initialize random Gaussian points based on anchor points after non-rigid editing to obtain initial Gaussian points.

[0091] Figure 4 A schematic diagram of a three-dimensional model editing method provided in Example 4. Figure 4 As shown, including:

[0092] S401, for any random Gaussian point, obtaining the random three-dimensional coordinates of the random Gaussian point.

[0093] The positions corresponding to the random Gaussian points are represented by random three-dimensional coordinates.

[0094] S402, determining K-value anchor points that satisfy a preset K-value nearest neighbor method from a plurality of non-rigidly edited anchor points; the K-value anchor points are a second preset number; the second preset number is less than or equal to the first preset number.

[0095] Specifically, a second preset number of non-rigidly edited anchor points with the closest K values ​​may be determined as K value anchor points.

[0096] Exemplarily, assuming that the first preset number of sparse anchor points is 10 and the second preset number is 4, the first 4 non-rigidly edited anchor points with the closest K values ​​are determined as K value anchor points.

[0097] S403, calculating a second distance; the second distance is the distance between the random three-dimensional coordinates and the three-dimensional coordinates of each K value anchor point.

[0098] S404: Determine the second distance as the weight corresponding to the anchor point after non-rigid editing.

[0099] S405, inputting the weight and K value anchor point into a preset initialization algorithm to output an initial Gaussian point after random Gaussian point initialization.

[0100] Among them, the initial three-dimensional Gaussian points can be used It indicates that, ,in , , represents the position of the initial center point, represents the covariance matrix of the initial three-dimensional Gaussian points, represents the color of the initial three-dimensional Gaussian points, Indicates the transparency of the initial 3D Gaussian points, Represents the number of initial 3D Gaussian points.

[0101] Among them, the preset initialization algorithm is shown in formula (3): (3)

[0103]

[0104]

[0105]

[0106]

[0107] in, as well as These are the anchor points after non-rigid editing. The center point position, covariance matrix, color and transparency in . Among them, is the weight corresponding to the i-th initial three-dimensional Gaussian point. is the second preset number. Assuming the second preset number is 4, then K=4.

[0108] Further, as well as Input into the preset initialization algorithm, and then obtain the initial value of each initial Gaussian point, that is, .

[0109] The present embodiment provides a three-dimensional model editing method. In the present embodiment, for any random Gaussian point, a second distance is calculated based on its random three-dimensional coordinates and the three-dimensional coordinates of a K-value anchor point, and the second distance is determined as the corresponding weight after non-rigid editing, and then a preset initialization algorithm is used to calculate the initial Gaussian point after the random Gaussian point is initialized, and the assignment of the initial Gaussian point is completed. In the present embodiment, the initialization of each random Gaussian point can be achieved; in addition, the present embodiment uses a K-value anchor point that satisfies the preset K-value nearest neighbor method for initialization, so the accuracy and stability of the initial Gaussian point are ensured.

[0110] Embodiment 5

[0111] This embodiment is a further refinement of any of the above embodiments. This embodiment densifies the initial Gaussian points to determine a further refinement of target densified Gaussian points that describe the appearance features of the target object.

[0112] Figure 5 A schematic diagram of a three-dimensional model editing method provided in Example 5. Figure 5 As shown, specifically including:

[0113] S501, performing a first iterative densification on the initial Gaussian points to generate a first iterative densified three-dimensional model.

[0114] It should be noted that the first-iteration densified 3D model can be generated based on the first-iteration densified Gaussian points and the optimized anchor points on the basis of the initial edited 3D model, wherein the initial edited 3D model is consistent with the source 3D model. Alternatively, the first-iteration densified 3D model can be generated based on the first-iteration densified Gaussian points and the optimized anchor points on the basis of the edited 3D model when the preset editing loss condition is satisfied.

[0115] S502, determining a density loss result of the three-dimensional model after densification in the first iteration.

[0116] The densification loss result refers to whether the densification loss value of the three-dimensional model after densification meets the preset densification loss condition. The densification loss result includes meeting the preset densification loss condition or not meeting the preset densification loss condition.

[0117] S503 , in response to the dense loss result satisfying a preset densification loss condition, determining the initial Gaussian point as a target densified Gaussian point describing the appearance features of the target object.

[0118] The preset densification loss condition is a pre-set loss condition for densification.

[0119] S504, in response to the densification loss result not satisfying the preset densification loss condition, performing at least one iterative densification based on the Gaussian points after the first iterative densification, until the densified loss result of the iteratively optimized densified three-dimensional model satisfies the preset densification loss condition, and the densified Gaussian points corresponding to the preset densification loss condition are determined as the target densified Gaussian points that describe the appearance features of the target object.

[0120] It should be noted that if the dense loss result does not meet the preset densification loss condition, iterative densification will continue on the basis of the first iterative densification Gaussian points, and each iterative densification will continue on the basis of the densified Gaussian points generated in the previous iteration, and a dense three-dimensional model will be generated based on the densified Gaussian points. The current dense three-dimensional model is generated based on the current iterative dense Gaussian points on the basis of the previous iterative dense three-dimensional model, until the dense loss result of the iteratively optimized dense three-dimensional model meets the preset densification loss condition, and the iteration ends. Among them, the current iterative dense Gaussian points are iteratively densified based on the densified Gaussian points generated in the previous iteration.

[0121] Furthermore, at the end of the iteration, the densified Gaussian points corresponding to the preset densification loss conditions are determined as the target densified Gaussian points that describe the appearance features of the target object.

[0122] Exemplarily, assume that two more iterative densifications are required. In the second iterative densification, iterative densification is performed on the basis of the Gaussian points after the first iterative densification to obtain the Gaussian points after the second iterative densification. A densified three-dimensional model of the second iterative densification is generated based on the Gaussian points after the second iterative densification, and a dense loss result of the densified three-dimensional model of the second iterative densification is determined. If the dense loss result meets the preset densification loss condition, the Gaussian points after the second iterative densification are determined as the target dense Gaussian points that describe the appearance characteristics of the target object. If the dense loss result does not meet the preset densification loss condition, iterative densification continues to be performed based on the Gaussian points after the second iterative densification until the dense loss result of the iteratively optimized dense three-dimensional model meets the preset densification loss condition, and the iteration ends.

[0123] The present embodiment provides a three-dimensional model editing method. In the present embodiment, the initial Gaussian points are firstly densified for the first time to generate a densified three-dimensional model of the first iteration, and then the corresponding densified loss result is determined. Furthermore, if the densified loss result satisfies the preset densification condition, the iteration is terminated and the initial Gaussian point is determined as the target densified Gaussian point; if the densified loss result does not satisfy the preset densification condition, the iterative densification is continued until it is determined that the densified loss result of the three-dimensional model satisfies the preset densified loss condition, and the densified Gaussian point corresponding to the preset densified loss condition is determined as the target densified Gaussian point; in the present embodiment, the target densified Gaussian point can describe the appearance characteristics of the target object, thereby making the appearance of the densified three-dimensional model at this time more accurate.

[0124] This embodiment also includes:

[0125] The densified three-dimensional model corresponding to the preset densification loss condition is determined as the target three-dimensional model; the target three-dimensional model is generated by the target densified Gaussian points and the optimized anchor points; the optimized anchor points are generated by optimizing the properties of the non-rigidly edited anchor points while keeping the positions of the non-rigidly edited anchor points unchanged.

[0126] It should be noted that any densified three-dimensional model is generated by the densified Gaussian points and the optimized anchor points. If it is a target three-dimensional model, it is generated by the target densified Gaussian points and the optimized anchor points.

[0127] Among them, the optimized anchor points are generated based on the optimization of the anchor points after non-rigid editing. It should be noted that the optimized anchor points can also be optimized within the current iteration number with each iteration to generate the optimized anchor points corresponding to the current iteration number. Thus, the target three-dimensional model can be generated by the target densified Gaussian points and the optimized anchor points corresponding to the last iteration number.

[0128] Specifically, in response to the densification loss result satisfying a preset densification loss condition, the densified three-dimensional model of the first iteration is determined as the target three-dimensional model.

[0129] The present embodiment provides a three-dimensional model editing method. In the present embodiment, a densified three-dimensional model that meets a preset densification loss condition is determined as a target three-dimensional model. In the present embodiment, the target three-dimensional model is generated by target densified Gaussian points and optimized anchor points, wherein the optimized anchor points are anchor points generated by optimizing the properties of the non-rigidly edited anchor points while keeping the positions of the non-rigidly edited anchor points unchanged. In the present embodiment, the target densified Gaussian points can describe the appearance features of the target object, and the optimized anchor points are optimized by the non-rigidly edited anchor points, which are generated by sparse anchor points. Since the sparse anchor points can describe the geometric features of the target object, the optimized anchor points can also describe the geometric features of the target object, and the optimized anchor point properties are more fitted and accurate, so that the target three-dimensional model is more accurate in geometry and appearance.

[0130] Embodiment 6

[0131] This embodiment is a further refinement of any of the above embodiments. This embodiment is an optional way of performing a first-iteration densification on the initial Gaussian points to generate a first-iteration densified three-dimensional model.

[0132] Figure 6 A schematic diagram of a three-dimensional model editing method provided in Example 6. Figure 6 As shown, including:

[0133] S601, determining the position gradients of multiple initial Gaussian points.

[0134] S602: In response to a position gradient of any initial Gaussian point being greater than or equal to a first preset gradient, determining the initial Gaussian point as a Gaussian point to be densified.

[0135] The first preset gradient can be .

[0136] Specifically, the position gradient of each initial Gaussian point is determined, and the magnitude of each position gradient is compared with the first preset gradient. In response to the position gradient of any initial Gaussian point being greater than or equal to the first preset gradient, the initial Gaussian point is determined to be a Gaussian point to be densified.

[0137] Exemplarily, assuming that there are 100 initial Gaussian points, it is determined that the position gradients of 50 of the initial Gaussian points are greater than or equal to the first preset gradient, and these 50 initial Gaussian points are determined as Gaussian points to be densified.

[0138] S603 , performing a first iterative densification on the Gaussian points to be densified based on the reconstruction type of the region where the Gaussian points to be densified are located, so as to generate a first iterative densified three-dimensional model.

[0139] Furthermore, based on the reconstruction type of the region where the Gaussian points to be densified are located, the first iterative densification is performed in a targeted manner, thereby generating a first iterative densified three-dimensional model.

[0140] The present embodiment provides a three-dimensional model editing method. In the present embodiment, a condition is first defined, that is, only the initial Gaussian points whose position gradient is greater than or equal to the first preset gradient are densified, and the initial Gaussian points are determined as the Gaussian points to be densified. Further, the first iterative densification is performed in a targeted manner according to the reconstruction type of the area where the Gaussian points to be densified are located, thereby generating a three-dimensional model after the first iterative densification.

[0141] Embodiment 7

[0142] This embodiment is a further refinement of any of the above embodiments. This embodiment is an optional way of performing a first-iteration densification on the Gaussian points to be densified based on the reconstruction type of the region where the Gaussian points to be densified are located to generate a first-iteration densified three-dimensional model.

[0143] Figure 7 A schematic diagram of a 3D model editing method provided in Example 7. Figure 7 As shown, including:

[0144] S701 , in response to the reconstruction type of the region where the Gaussian points to be densified are located being an under-reconstruction type, performing under-reconstruction iterative densification on the Gaussian points to be densified to determine under-reconstructed dense Gaussian points corresponding to the Gaussian points to be densified.

[0145] According to the above example, assuming that the reconstruction type of the area where 20 of the 50 Gaussian points to be densified are located is an under-reconstruction type, then under-reconstruction iterative densification is performed on these 20 Gaussian points to be densified, and then the under-reconstructed dense Gaussian points corresponding to these 20 Gaussian points to be densified are determined.

[0146] S702 , in response to the reconstruction type of the region where the Gaussian points to be densified are located being a transitional reconstruction type, performing transitional reconstruction iterative densification on the Gaussian points to be densified to determine transitional reconstruction dense Gaussian points corresponding to the Gaussian points to be densified.

[0147] According to the above exemplary embodiment, for the remaining 30 Gaussian points to be densified, the reconstruction type of the area where they are located is the transition reconstruction type, and then these 30 Gaussian points to be densified are subjected to transition reconstruction iterative densification to determine the transition reconstruction dense Gaussian points corresponding to these 30 Gaussian points to be densified.

[0148] S703, based on each under-reconstructed dense Gaussian point, each transitional reconstructed dense Gaussian point, the initial Gaussian point except the Gaussian point to be densified, and the optimized anchor point, a densified three-dimensional model of the first iteration is generated; each under-reconstructed dense Gaussian point, each transitional reconstructed dense Gaussian point, the initial Gaussian point except the Gaussian point to be densified is the densified Gaussian point of the first iteration.

[0149] The initial Gaussian points other than the Gaussian points to be densified refer to the initial Gaussian points remaining in the initial Gaussian points except the Gaussian points to be densified.

[0150] According to the above exemplary, in the above steps, 20 under-reconstructed dense Gaussian points corresponding to the Gaussian points to be densified are obtained through S701, and 30 transitional reconstructed dense Gaussian points corresponding to the Gaussian points to be densified are obtained through S702, and it is calculated that the initial Gaussian points other than the Gaussian points to be densified are 50. Further, the under-reconstructed dense Gaussian points corresponding to the 20 Gaussian points to be densified, the transitional reconstructed dense Gaussian points corresponding to the 30 Gaussian points to be densified, the 50 initial Gaussian points other than the Gaussian points to be densified, and the optimized anchor points are used to generate the first iteration of the densified three-dimensional model.

[0151] It should be noted that, in this example, the under-reconstructed dense Gaussian points corresponding to the 20 Gaussian points to be densified, the transitional reconstructed dense Gaussian points corresponding to the 30 Gaussian points to be densified, and the 50 initial Gaussian points other than the Gaussian points to be densified are the densified Gaussian points of the first iteration.

[0152] In one approach, the densified three-dimensional model of the first iteration is generated by covering the area where the corresponding Gaussian points to be densified are located based on the source three-dimensional model, using the corresponding under-reconstructed dense Gaussian points, the transitional reconstructed dense Gaussian points, and the initial Gaussian points other than the Gaussian points to be densified corresponding to the first iteration. Specifically, assuming that a Gaussian point to be densified is densified by under-reconstructed iterations to generate a corresponding under-reconstructed dense Gaussian point, the under-reconstructed dense Gaussian point covers the area where the Gaussian points to be densified are located, and thus the densified three-dimensional model can be accurately obtained.

[0153] The present embodiment provides a three-dimensional model editing method. In the present embodiment, iterative densification is performed in a targeted manner according to whether the reconstruction type of the area where the Gaussian points to be densified are located is an under-reconstruction type or a transitional reconstruction type, and then under-reconstructed dense Gaussian points and transitional reconstructed dense Gaussian points are determined. Then, based on each under-reconstructed dense Gaussian point, each transitional reconstructed dense Gaussian point, the initial Gaussian points except the Gaussian points to be densified, and the optimized anchor points, a densified three-dimensional model of the first iteration is generated based on the densified Gaussian points of the first iteration and the optimized anchor points. Since the iterative densification is performed in a targeted manner according to the distinction of reconstruction types, the target three-dimensional model can be made more accurate.

[0154] Embodiment 8

[0155] This embodiment is a further refinement of any of the above embodiments. This embodiment is an optional method of performing under-reconstruction iterative densification on the Gaussian points to be densified to determine the under-reconstructed dense Gaussian points corresponding to the Gaussian points to be densified, specifically including:

[0156] Step 1: When performing under-reconstruction iterative densification, the Gaussian point to be densified is determined as the first Gaussian point, and the Gaussian point to be densified is copied to obtain the corresponding copied Gaussian point; the scale of the first Gaussian point is smaller than the first preset scale.

[0157] The scale of the first Gaussian point is smaller than the first preset scale. It can be understood that the first Gaussian point is a small Gaussian.

[0158] Furthermore, the first Gaussian point is copied to obtain a copied Gaussian point.

[0159] Step 2: The copied Gaussian points are determined as the under-reconstructed dense Gaussian points corresponding to the Gaussian points to be densified.

[0160] According to the above exemplary embodiment, 20 Gaussian points to be densified are determined as small Gaussian points. For any Gaussian point to be densified, the Gaussian point to be densified is copied, and then the copied Gaussian point corresponding to the Gaussian point to be densified is obtained. The copied Gaussian point is determined as the under-reconstructed dense Gaussian point corresponding to the Gaussian point to be densified, and then the under-reconstructed dense Gaussian points corresponding to each of the 20 Gaussian points to be densified can be obtained.

[0161] The present embodiment provides a three-dimensional model editing method. In the present embodiment, for the first Gaussian point, since the scale of the first Gaussian point is smaller than the first preset scale, it is determined that the first Gaussian point is actually a small Gaussian point. Therefore, in order to perform under-reconstruction iterative densification, the first Gaussian point is copied to obtain the copied Gaussian point, and the copied Gaussian point is the under-reconstructed dense Gaussian point corresponding to the Gaussian point to be densified.

[0162] Embodiment 9

[0163] This embodiment is a further refinement of any of the above embodiments. This embodiment is an optional method of performing transition reconstruction iterative densification on the Gaussian points to be densified to determine the transition reconstruction dense Gaussian points corresponding to the Gaussian points to be densified, specifically including:

[0164] Step 1: When performing iterative densification of transition reconstruction, the Gaussian point to be densified is determined as the second Gaussian point, and the Gaussian point to be densified is split to obtain the split Gaussian point; the scale of the split Gaussian point meets the preset splitting scale condition; the second Gaussian point is greater than or equal to the second preset scale; the first preset scale is less than the second preset scale.

[0165] The first preset scale is smaller than the second preset scale. Since the second Gaussian point is greater than or equal to the second preset scale, the second Gaussian point can be understood as a large Gaussian point, and thus the second Gaussian point needs to be split.

[0166] Step 2: Determine the split Gaussian points as transitional reconstructed dense Gaussian points corresponding to the Gaussian points to be densified.

[0167] According to the above exemplary embodiment, 30 Gaussian points to be densified are determined as second Gaussian points. Any second Gaussian point is split to obtain the split Gaussian point, and then the split Gaussian point is determined as the transitional reconstructed dense Gaussian point corresponding to the Gaussian point to be densified.

[0168] The present embodiment provides a three-dimensional model editing method. In the present embodiment, for the second Gaussian point, since the scale of the second Gaussian point is greater than or equal to the second preset scale, it is determined that the second Gaussian point is actually a large Gaussian point. Therefore, in order to perform transition reconstruction iterative densification, the second Gaussian point is split to obtain the split Gaussian point. The split Gaussian point is the transition reconstruction dense Gaussian point corresponding to the Gaussian point to be densified.

[0169] Embodiment 10

[0170] This embodiment is a further refinement of any of the above embodiments. This embodiment is an optional way to determine the density loss result of the three-dimensional model after the densification of the first iteration.

[0171] Figure 8 A schematic flow chart of a three-dimensional model editing method provided in Example 10. Figure 8 As shown, including:

[0172] S801, rendering the densified three-dimensional model of the first iteration at different viewing angles to obtain a first densified rendering image set, wherein the first densified rendering image set includes first densified rendering images at different viewing angles.

[0173] S802, input the first densified rendering image set and the target multi-view image set into a preset densified reconstruction loss function to output a densified loss value corresponding to the first iteration; the target multi-view image set is an image set that forms a target state from a source state after the target object performs a target action under the perspective of the corresponding source image; the target multi-view image set is a plurality of source images under different perspectives and target text prompt information output from a multi-view image editing model that has been trained to convergence.

[0174] Among them, the preset densification reconstruction loss function is shown in formula (3):

[0175] (3)

[0176] in, It refers to the target multi-view image at the i-th viewing angle, and the target multi-view image comes from the target multi-view image set; refers to the first densified rendering image at the i-th viewing angle, and the first densified rendering image comes from the first densified rendering image set; wherein M refers to the number of viewing angles; wherein, Refers to the densification loss value.

[0177] S803 , in response to the densification loss value satisfying a preset densification loss condition, determining that the densification loss result satisfies the preset densification loss condition.

[0178] Specifically, if the output If the preset densification loss condition is met, the densification loss result is determined to meet the preset densification loss condition.

[0179] S804 , in response to the densification loss value not satisfying a preset densification loss condition, determining that the densification loss result does not satisfy the preset densification loss condition.

[0180] If the output If the preset densification loss condition is not met, then the densification loss result is determined as not meeting the preset densification loss condition.

[0181] The present embodiment provides a three-dimensional model editing method. In the present embodiment, a densification loss value is calculated based on a target multi-view image set. Since the target multi-view image set is a plurality of source images at different viewpoints and target text prompt information output from a multi-view image editing model that has been trained to convergence, the target state in the target multi-view image set is more accurate. Therefore, the dense loss result can be accurately determined based on the target multi-view image set.

[0182] Embodiment 11

[0183] This embodiment is a further refinement of any of the above embodiments. This embodiment performs at least one iterative densification based on the Gaussian points after the first iterative densification until the densified loss result of the iteratively optimized three-dimensional model meets the preset densification loss condition, and determines the densified Gaussian points corresponding to the preset densified loss condition as target densified Gaussian points that describe the appearance characteristics of the target object. An optional method includes:

[0184] Step 1: When performing at least one iterative densification on the basis of the Gaussian points after the first iterative densification, in response to the current number of iterations being greater than the first preset number of iterations, the Gaussian points to be reconstructed are determined from the densified Gaussian points corresponding to the current number of iterations; the position gradient of the Gaussian points to be reconstructed is greater than the second preset gradient.

[0185] The second preset gradient is greater than the first preset gradient.

[0186] The first preset number of iterations may be any number, assuming it is 300.

[0187] Exemplarily, if the current iteration number is 301, it is greater than the first preset iteration number, and the Gaussian points to be reconstructed are determined from the densified Gaussian points corresponding to the current iteration number. If the densified Gaussian points corresponding to the current iteration number are 1000, the Gaussian points to be reconstructed are determined from the 1000 densified Gaussian points.

[0188] Step 2: Continue to iteratively densify the reconstructed Gaussian points to optimize the densified three-dimensional model until the densification loss value of the optimized densified three-dimensional model meets the preset densification loss condition at a certain number of iterations, and determine the densified Gaussian points corresponding to the preset densification loss condition as the target densified Gaussian points.

[0189] Furthermore, the reconstructed Gaussian points are further iteratively densified. The iterative densification method is the same as the first iterative densification method mentioned above, and will not be described in detail here.

[0190] The present embodiment provides a three-dimensional model editing method. In the present embodiment, when the number of iterations reaches a certain level, that is, the current number of iterations is greater than the first preset number of times, Gaussian points to be reconstructed with a gradient greater than the second preset gradient are determined from the densified Gaussian points corresponding to the current number of iterations, and the Gaussian points to be reconstructed are iteratively densified, so the total number of Gaussian points can be controlled in the present embodiment.

[0191] Embodiment 12

[0192] The present embodiment is a further refinement of any of the above embodiments. The present embodiment is an optional method of continuing to iteratively densify the Gaussian points to be reconstructed to optimize the densified three-dimensional model until the densification loss value of the optimized densified three-dimensional model meets the preset densification loss condition at a certain number of iterations, and determining the densified Gaussian points corresponding to the preset densification loss condition as the target densified Gaussian points.

[0193] Fig. 9 A schematic flow chart of a three-dimensional model editing method provided in Example 12. Fig. 9 As shown, including:

[0194] S901, when continuing iterative densification, in response to the current number of iterations being greater than the second preset number of iterations, determining Gaussian points to be deleted from the densified Gaussian points corresponding to the current number of iterations; the scale of the Gaussian points to be deleted is greater than the third preset scale; the second preset number of iterations is greater than the first preset number of iterations.

[0195] The Gaussian points to be deleted refer to Gaussian points whose scale is greater than the third preset scale among the densified Gaussian points corresponding to the current number of iterations when the current number of iterations is greater than the second preset number of iterations.

[0196] In this embodiment, when iterative densification continues, if the current iteration number is 1001, where the second preset iteration number is 1000, the current iteration number is greater than the second preset iteration number, and the Gaussian points to be deleted are determined from the densified Gaussian points corresponding to the current iteration number.

[0197] S902, deleting the Gaussian points to be deleted from the densified Gaussian points of the current iteration number, so as to obtain the deleted densified Gaussian points corresponding to the current iteration number.

[0198] S903, continue to iteratively densify the deleted densified Gaussian points to optimize the densified three-dimensional model, until the densification loss value of the optimized densified three-dimensional model meets the preset densification loss condition at a certain number of iterations, and determine the densified Gaussian points corresponding to the preset densification loss condition as the target densified Gaussian points.

[0199] The present embodiment provides a three-dimensional model editing method. In the present embodiment, when the current number of iterations is greater than the second preset number of iterations, Gaussian points to be deleted whose scale is greater than the second preset scale are determined from the densified Gaussian points corresponding to the current number of iterations, and the Gaussian points to be deleted are deleted therefrom. The deleted densified Gaussian points are then iteratively densified to optimize the densified three-dimensional model, until at a certain number of iterations, the densification loss value of the optimized densified three-dimensional model meets the preset densification loss condition, and the corresponding densified Gaussian points at this time are determined as target densified Gaussian points. In the present embodiment, when the number of iterations reaches a higher level, the Gaussian points to be deleted are deleted, which can effectively ensure the total number of Gaussian points, increase the number and density of target densified Gaussian points, and ensure accuracy.

[0200] Through the description of the above implementation methods, those skilled in the art can clearly understand that the method according to the above embodiment can be implemented by means of software plus a necessary general hardware platform, and of course by hardware, but in many cases the former is a better implementation method.

[0201] Embodiment 13

[0202] The embodiment of the present application also provides a three-dimensional model editing device, Fig.10 This is a schematic diagram of the structure of a 3D model editing device provided in Example 13. The 3D model editing device 1000 includes the following modules:

[0203] The acquisition module 1001 is used to obtain sparse anchor points from a source three-dimensional model; the source three-dimensional model includes a target object; the target object is in a source state; the sparse anchor points are anchor points that describe geometric features of the target object;

[0204] The editing module 1002 is used to perform preliminary non-rigid editing based on the sparse anchor points to obtain non-rigidly edited anchor points; the non-rigidly edited anchor points are anchor points of the target object that are formed from the source state to the target state after the non-rigid editing;

[0205] Initialization module 1003, used for initializing random Gaussian points based on the anchor points after non-rigid editing to obtain initial Gaussian points; the random Gaussian points are source Gaussian points in the source three-dimensional model;

[0206] The densification module 1004 is used to densify the initial Gaussian points to determine the target densed Gaussian points that describe the appearance characteristics of the target object; the densified three-dimensional model generated based on the target densed Gaussian points is the target three-dimensional model, and the target object in the target three-dimensional model forms a target state from a source state after performing a target action.

[0207] Optionally, when obtaining sparse anchor points from the source 3D model, the obtaining module 1001 is specifically used to:

[0208] Determine a plurality of source Gaussian points included in the source three-dimensional model;

[0209] Calculate a first distance; the first distance is the distance between each source Gaussian point calculated based on the three-dimensional coordinates of the source Gaussian point;

[0210] The source Gaussian points corresponding to the first distance that satisfies the preset maximum distance sampling mechanism are determined as sparse anchor points; the number of sparse anchor points is less than the number of source Gaussian points; the sparse anchor points are greater than or equal to the first preset number.

[0211] Optionally, the editing module 1002, when performing preliminary non-rigid editing based on the sparse anchor points to obtain non-rigidly edited anchor points, is specifically configured to:

[0212] The sparse anchor points are initially non-rigidly edited by using the degree of freedom transformation algorithm for iterative optimization to obtain the non-rigidly edited anchor points; the degree of freedom transformation algorithm is used to obtain the target rotation matrix and the target translation vector, and the non-rigidly edited anchor points are obtained based on the target rotation matrix and the target translation vector. The initialization module 1003, when initializing the random Gaussian points based on the non-rigidly edited anchor points to obtain the initial Gaussian points, is specifically used for:

[0213] For any random Gaussian point, obtain the random three-dimensional coordinates of the random Gaussian point;

[0214] Determine a K-value anchor point that satisfies a preset K-value nearest neighbor method from a plurality of non-rigidly edited anchor points; the K-value anchor points are a second preset number; the second preset number is less than or equal to the first preset number;

[0215] Calculate the second distance; the second distance is the distance between the random three-dimensional coordinates and the three-dimensional coordinates of each K value anchor point;

[0216] Determine the second distance as the weight corresponding to the anchor point after non-rigid editing;

[0217] The weights and K value anchor points are input into the preset initialization algorithm to output the initial Gaussian points after random Gaussian point initialization.

[0218] Optionally, the densification module 1004, when densifying the initial Gaussian points to determine target densified Gaussian points describing the appearance features of the target object, is specifically configured to:

[0219] Performing a first-iteration densification on the initial Gaussian points to generate a first-iteration densified three-dimensional model;

[0220] Determine the density loss result of the three-dimensional model after densification in the first iteration;

[0221] In response to the dense loss result satisfying a preset densification loss condition, determining the initial Gaussian point as a target densified Gaussian point describing the appearance features of the target object;

[0222] In response to the densification loss result not satisfying the preset densification loss condition, performing at least one iterative densification based on the Gaussian points after the first iterative densification, until the densified loss result of the iteratively optimized densified three-dimensional model satisfies the preset densification loss condition, and determining the densified Gaussian points corresponding to the preset densification loss condition as target densified Gaussian points describing the appearance features of the target object;

[0223] Optionally, this embodiment provides a three-dimensional model editing device, further comprising: a determination module;

[0224] A determination module is used to determine the densified three-dimensional model corresponding to the preset densification loss condition as the target three-dimensional model; the target three-dimensional model is generated by the target densified Gaussian points and the optimized anchor points; the optimized anchor points are generated by optimizing the properties of the non-rigidly edited anchor points while keeping the positions of the non-rigidly edited anchor points unchanged.

[0225] Optionally, the densification module 1004, when performing the first iterative densification on the initial Gaussian points to generate the first iterative densified three-dimensional model, is specifically configured to:

[0226] Determine the position gradients of multiple initial Gaussian points;

[0227] In response to a position gradient of any initial Gaussian point being greater than or equal to a first preset gradient, determining the initial Gaussian point as a Gaussian point to be densified;

[0228] The Gaussian points to be densified are densified for the first time iteratively based on the reconstruction type of the region where the Gaussian points to be densified are located, so as to generate a densified three-dimensional model of the first iterative step.

[0229] Optionally, the densification module 1004, when performing a first-iteration densification on the Gaussian points to be densified based on the reconstruction type of the region where the Gaussian points to be densified are located to generate a first-iteration densified three-dimensional model, is specifically configured to:

[0230] In response to the reconstruction type of the region where the Gaussian points to be densified are located being an under-reconstruction type, performing under-reconstruction iterative densification on the Gaussian points to be densified to determine under-reconstructed dense Gaussian points corresponding to the Gaussian points to be densified;

[0231] In response to the reconstruction type of the region where the Gaussian points to be densified are located being a transitional reconstruction type, performing transitional reconstruction iterative densification on the Gaussian points to be densified to determine transitional reconstruction dense Gaussian points corresponding to the Gaussian points to be densified;

[0232] A densified three-dimensional model of the first iteration is generated based on each under-reconstructed dense Gaussian point, each transitional reconstructed dense Gaussian point, the initial Gaussian point except the Gaussian point to be densified, and the optimized anchor point; each under-reconstructed dense Gaussian point, each transitional reconstructed dense Gaussian point, and the initial Gaussian point except the Gaussian point to be densified are the densified Gaussian points of the first iteration.

[0233] Optionally, the densification module 1004, when performing under-reconstruction iterative densification on the Gaussian points to be densified to determine under-reconstructed dense Gaussian points corresponding to the Gaussian points to be densified, is specifically configured to:

[0234] When performing under-reconstruction iterative densification, the Gaussian point to be densified is determined as the first Gaussian point, and the Gaussian point to be densified is copied to obtain a corresponding copied Gaussian point; the scale of the first Gaussian point is smaller than the first preset scale;

[0235] The copied Gaussian points are determined as under-reconstructed dense Gaussian points corresponding to the Gaussian points to be densified.

[0236] Optionally, the densification module 1004, when performing transition reconstruction iterative densification on the Gaussian points to be densified to determine transition reconstruction dense Gaussian points corresponding to the Gaussian points to be densified, is specifically configured to:

[0237] When performing iterative densification of transition reconstruction, the Gaussian point to be densified is determined as the second Gaussian point, and the Gaussian point to be densified is split to obtain the split Gaussian point; the scale of the split Gaussian point meets the preset splitting scale condition; the second Gaussian point is greater than or equal to the second preset scale; the first preset scale is less than the second preset scale;

[0238] The split Gaussian points are determined as transitional reconstructed dense Gaussian points corresponding to the Gaussian points to be densified.

[0239] Optionally, the densification module 1004, when determining the density loss result of the densified three-dimensional model of the first iteration, is specifically configured to:

[0240] Rendering the densified three-dimensional model of the first iteration at different viewing angles to obtain a first densified rendering image set, wherein the first densified rendering image set includes first densified rendering images at different viewing angles;

[0241] The first densified rendering image set and the target multi-view image set are input into a preset densified reconstruction loss function to output a densified loss value corresponding to the first iteration; the target multi-view image set is an image set formed from a source state to a target state after the target object performs a target action under the perspective of the corresponding source image; the target multi-view image set is a plurality of source images under different perspectives and target text prompt information output from a multi-view image editing model that has been trained to convergence;

[0242] In response to the densification loss value satisfying a preset densification loss condition, determining that the densification loss result satisfies the preset densification loss condition;

[0243] In response to the densification loss value not satisfying a preset densification loss condition, a densification loss result is determined to not satisfy the preset densification loss condition.

[0244] Optionally, the densification module 1004 performs at least one iterative densification based on the Gaussian points after the first iterative densification until the densified loss result of the iteratively optimized three-dimensional model meets the preset densification loss condition, and determines the densified Gaussian points corresponding to the preset densified loss condition as the target densified Gaussian points describing the appearance characteristics of the target object, specifically for:

[0245] When performing at least one iterative densification based on the Gaussian points after the first iterative densification, in response to the current iteration number being greater than the first preset iteration number, determining the Gaussian points to be reconstructed from the densified Gaussian points corresponding to the current iteration number; the position gradient of the Gaussian points to be reconstructed is greater than the second preset gradient;

[0246] The reconstructed Gaussian points are continuously densified iteratively to optimize the densified three-dimensional model until the densification loss value of the optimized densified three-dimensional model meets the preset densification loss condition at a certain number of iterations, and the densified Gaussian points corresponding to the preset densification loss condition are determined as the target densified Gaussian points.

[0247] Optionally, the densification module 1004 continues to iteratively densify the Gaussian points to be reconstructed to optimize the densified three-dimensional model until the densification loss value of the optimized densified three-dimensional model meets a preset densification loss condition at a certain number of iterations, and determines the densified Gaussian points corresponding to the preset densification loss condition as target densified Gaussian points, specifically for:

[0248] When iterative densification continues, in response to the current number of iterations being greater than the second preset number of iterations, determining Gaussian points to be deleted from the densified Gaussian points corresponding to the current number of iterations; the scale of the Gaussian points to be deleted is greater than the third preset scale; the second preset number of iterations is greater than the first preset number of iterations;

[0249] The Gaussian points to be deleted are deleted from the densified Gaussian points of the current iteration number to obtain the deleted densified Gaussian points corresponding to the current iteration number;

[0250] Continue to iteratively densify the deleted densified Gaussian points to optimize the densified three-dimensional model until the densification loss value of the optimized densified three-dimensional model meets the preset densification loss condition at a certain number of iterations, and determine the densified Gaussian points corresponding to the preset densification loss condition as the target densified Gaussian points.

[0251] The description of the features in the embodiment corresponding to the three-dimensional model editing device can refer to the relevant description of the embodiment corresponding to the three-dimensional model editing method, which will not be repeated here.

[0252] The embodiment of the present application also provides an editing device, Fig.11 This is a schematic diagram of the structure of the editing device provided in Example 13. Fig.11 As shown, the editing device 1100 includes a processor 1101 and a memory 1102, wherein the processor 1101, the memory 1102 and the communication component 1103 are connected via a bus 1104. The memory 1102 stores a computer program, and the processor 1101 is configured to run the computer program to execute the steps in any of the above three-dimensional model editing method embodiments.

[0253] An embodiment of the present application further provides a computer-readable storage medium, in which a computer program is stored, wherein the computer program is configured to execute the steps of any one of the above-mentioned three-dimensional model editing method embodiments when running.

[0254] In an exemplary embodiment, the computer-readable storage medium may include, but is not limited to, various media that can store computer programs, such as a USB flash drive, a read-only memory (ROM), a random access memory (RAM), a mobile hard disk, a magnetic disk or an optical disk.

[0255] An embodiment of the present application further provides a computer program product, which includes a computer program. When the computer program is executed by a processor, the steps in any one of the above-mentioned three-dimensional model editing method embodiments are implemented.

[0256] An embodiment of the present application also provides another computer program product, including a non-volatile computer-readable storage medium, the non-volatile computer-readable storage medium storing a computer program, and when the computer program is executed by a processor, the steps in any of the above-mentioned three-dimensional model editing method embodiments are implemented.

[0257] Professionals may further appreciate that the units and algorithm steps of each example described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, computer software, or a combination of the two. In order to clearly illustrate the interchangeability of hardware and software, the composition and steps of each example have been generally described in the above description according to function. Whether these functions are performed in hardware or software depends on the specific application and design constraints of the technical solution. Professionals and technicians may use different methods to implement the described functions for each specific application, but such implementation should not be considered to be beyond the scope of this application.

[0258] The above is a detailed introduction to a three-dimensional model editing method provided by the present application. Specific examples are used in this article to illustrate the principles and implementation methods of the present application. The description of the above embodiments is only used to help understand the method and core idea of ​​the present application. It should be pointed out that for ordinary technicians in this technical field, without departing from the principles of the present application, several improvements and modifications can be made to the present application, and these improvements and modifications also fall within the scope of protection of the claims of the present application.

Claims

1. A three-dimensional model editing method, characterized in that: include: Obtain sparse anchor points from the source 3D model; The source three-dimensional model includes a target object; the target object is in a source state; the sparse anchor points are anchor points that describe geometric features of the target object; Performing preliminary non-rigid editing based on the sparse anchor points to obtain non-rigidly edited anchor points; The anchor point after non-rigid editing is the anchor point of the target object that forms the target state from the source state after non-rigid editing; Initializing random Gaussian points based on the non-rigidly edited anchor points to obtain initial Gaussian points; the random Gaussian points are part of the source Gaussian points in the source three-dimensional model; Densifying the initial Gaussian points to determine target densified Gaussian points that describe appearance features of the target object; The densified three-dimensional model generated based on the target densified Gaussian points is the target three-dimensional model. After the target object in the target three-dimensional model performs the target action, the target state is formed from the source state.

2. The method according to claim 1, characterized in that: The step of obtaining sparse anchor points from the source three-dimensional model comprises: Determine a plurality of source Gaussian points included in the source three-dimensional model; Calculating a first distance; the first distance is the distance between each source Gaussian point calculated based on the three-dimensional coordinates of the source Gaussian point; The source Gaussian points corresponding to the first distance that satisfies the preset farthest distance sampling mechanism are determined as sparse anchor points; the number of the sparse anchor points is less than the number of the source Gaussian points; the sparse anchor points are greater than or equal to a first preset number.

3. The method according to claim 2, characterized in that The performing preliminary non-rigid editing based on the sparse anchor points to obtain non-rigidly edited anchor points includes: The sparse anchor points are initially non-rigidly edited by iterative optimization using a degree of freedom transformation algorithm to obtain non-rigidly edited anchor points; the degree of freedom transformation algorithm is used to obtain a target rotation matrix and a target translation vector, and the non-rigidly edited anchor points are obtained based on the target rotation matrix and the target translation vector.

4. The method according to claim 3, characterized in that Initializing random Gaussian points based on the anchor points after non-rigid editing to obtain initial Gaussian points includes: For any random Gaussian point, obtain the random three-dimensional coordinates of the random Gaussian point; Determine a K-value anchor point that satisfies a preset K-value nearest neighbor method from the plurality of non-rigidly edited anchor points; the K-value anchor points are a second preset number; the second preset number is less than or equal to the first preset number; Calculate a second distance; the second distance is the distance between the random three-dimensional coordinates and the three-dimensional coordinates of each of the K value anchor points; Determining the second distance as a weight corresponding to the anchor point after non-rigid editing; The weight and K value anchor point are input into a preset initialization algorithm to output the initial Gaussian point after the random Gaussian point is initialized.

5. The method according to claim 4, characterized in that The densifying the initial Gaussian points to determine target densified Gaussian points describing the appearance features of the target object includes: Performing a first-iteration densification on the initial Gaussian points to generate a first-iteration densified three-dimensional model; Determining a density loss result of the three-dimensional model after densification of the first iteration; In response to the dense loss result satisfying a preset densification loss condition, determining the initial Gaussian point as a target densified Gaussian point describing an appearance feature of the target object; In response to the densification loss result not satisfying the preset densification loss condition, performing at least one iterative densification based on the Gaussian points after the first iterative densification, until the densified loss result of the iteratively optimized densified three-dimensional model satisfies the preset densification loss condition, and determining the densified Gaussian points corresponding to the preset densification loss condition as target densified Gaussian points describing the appearance features of the target object; The method further comprises: The densified three-dimensional model corresponding to the preset densification loss condition is determined as the target three-dimensional model; the target three-dimensional model is generated by the target densified Gaussian points and the optimized anchor points; the optimized anchor points are anchor points generated by optimizing the properties of the non-rigidly edited anchor points while keeping the positions of the non-rigidly edited anchor points unchanged.

6. The method according to claim 5, characterized in that The first iterative densification of the initial Gaussian points to generate a first iterative densified three-dimensional model includes: Determining the position gradients of a plurality of the initial Gaussian points; In response to the position gradient of any initial Gaussian point being greater than or equal to a first preset gradient, determining the initial Gaussian point as a Gaussian point to be densified; The Gaussian points to be densified are densified for the first time iteratively based on the reconstruction type of the region where the Gaussian points to be densified are located, so as to generate a densified three-dimensional model of the first iteratively.

7. The method according to claim 6, characterized in that The first iterative densification of the Gaussian points to be densified is performed based on the reconstruction type of the region where the Gaussian points to be densified are located, so as to generate a first iterative densified three-dimensional model, including: In response to the reconstruction type of the region where the Gaussian points to be densified are located being an under-reconstruction type, performing under-reconstruction iterative densification on the Gaussian points to be densified to determine under-reconstructed dense Gaussian points corresponding to the Gaussian points to be densified; In response to the reconstruction type of the region where the Gaussian points to be densified are located being a transitional reconstruction type, performing transitional reconstruction iterative densification on the Gaussian points to be densified to determine transitional reconstruction dense Gaussian points corresponding to the Gaussian points to be densified; A densified three-dimensional model of the first iteration is generated based on each of the under-reconstructed dense Gaussian points, each of the transitional reconstructed dense Gaussian points, the initial Gaussian points except the Gaussian points to be densified, and the optimized anchor points; the under-reconstructed dense Gaussian points, each of the transitional reconstructed dense Gaussian points, and the initial Gaussian points except the Gaussian points to be densified are the densified Gaussian points of the first iteration.

8. The method according to claim 7, characterized in that The performing under-reconstruction iterative densification on the Gaussian points to be densified to determine under-reconstructed dense Gaussian points corresponding to the Gaussian points to be densified includes: When performing under-reconstruction iterative densification, the Gaussian point to be densified is determined as a first Gaussian point, and the Gaussian point to be densified is copied to obtain a corresponding copied Gaussian point; the scale of the first Gaussian point is smaller than a first preset scale; The copied Gaussian points are determined as under-reconstructed dense Gaussian points corresponding to the Gaussian points to be densified.

9. The method according to claim 8, characterized in that The performing transition reconstruction iterative densification on the Gaussian points to be densified to determine transition reconstruction dense Gaussian points corresponding to the Gaussian points to be densified, comprises: When performing iterative densification of transition reconstruction, the Gaussian point to be densified is determined as a second Gaussian point, and the Gaussian point to be densified is split to obtain a split Gaussian point; the scale of the split Gaussian point meets a preset splitting scale condition; the second Gaussian point is greater than or equal to a second preset scale; the first preset scale is less than the second preset scale; The split Gaussian points are determined as transitional reconstructed dense Gaussian points corresponding to the Gaussian points to be densified.

10. The method according to claim 5, characterized in that The step of determining a dense loss result of the three-dimensional model after densification of the first iteration includes: Rendering the densified three-dimensional model of the first iteration at different viewing angles to obtain a first densified rendering image set, wherein the first densified rendering image set includes first densified rendering images at different viewing angles; The first densified rendering image set and the target multi-view image set are input into a preset densified reconstruction loss function to output a densified loss value corresponding to the first iteration; the target multi-view image set is an image set formed from the source state to the target state after the target object performs a target action under the perspective of the corresponding source image; the target multi-view image set is a plurality of source images under different perspectives and target text prompt information output from a multi-view image editing model that has been trained to convergence; In response to the densification loss value satisfying a preset densification loss condition, determining that the densification loss result satisfies the preset densification loss condition; In response to the densification loss value not satisfying a preset densification loss condition, determining the densification loss result as not satisfying the preset densification loss condition.

11. The method according to claim 5, characterized in that The method of performing at least one iterative densification based on the Gaussian points after the first iterative densification until the densified loss result of the iteratively optimized three-dimensional model satisfies a preset densification loss condition, and determining the densified Gaussian points corresponding to the preset densified loss condition as target densified Gaussian points describing the appearance features of the target object, comprises: When performing at least one iterative densification on the basis of the Gaussian points after the first iterative densification, in response to the current number of iterations being greater than the first preset number of iterations, determining a Gaussian point to be reconstructed from the densified Gaussian points corresponding to the current number of iterations; the position gradient of the Gaussian point to be reconstructed is greater than the second preset gradient; The Gaussian points to be reconstructed are continuously densified iteratively to optimize the densified three-dimensional model until the densification loss value of the optimized densified three-dimensional model meets the preset densification loss condition at a certain number of iterations, and the densified Gaussian points corresponding to the preset densification loss condition are determined as the target densified Gaussian points.

12. The method according to claim 11, characterized in that The iterative densification of the Gaussian points to be reconstructed is continued to optimize the densified three-dimensional model until the densification loss value of the optimized densified three-dimensional model meets a preset densification loss condition at a certain number of iterations, and the densified Gaussian points corresponding to the preset densification loss condition are determined as target densified Gaussian points, including: When iterative densification continues, in response to the current number of iterations being greater than the second preset number of iterations, determining Gaussian points to be deleted from the densified Gaussian points corresponding to the current number of iterations; the scale of the Gaussian points to be deleted is greater than the third preset scale; and the second preset number of iterations is greater than the first preset number of iterations; Deleting the Gaussian points to be deleted from the densified Gaussian points of the current iteration number to obtain the deleted densified Gaussian points corresponding to the current iteration number; Continue to iteratively densify the deleted densified Gaussian points to optimize the densified three-dimensional model until the densification loss value of the optimized densified three-dimensional model meets the preset densification loss condition at a certain number of iterations, and determine the densified Gaussian points corresponding to the preset densification loss condition as the target densified Gaussian points.

13. An editing device, characterized in that: include: Memory for storing computer programs; A processor, configured to implement the steps of the three-dimensional model editing method according to any one of claims 1 to 12 when executing the computer program.

14. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores a computer program, wherein the computer program, when executed by a processor, implements the steps of the three-dimensional model editing method according to any one of claims 1 to 12.

15. A computer program product comprising a computer program, characterized in that When the computer program is executed by a processor, the steps of the three-dimensional model editing method according to any one of claims 1 to 12 are implemented.

Images