Bone skin rendering method and system based on graph neural network and proxy model

Abstract

The invention relates to the field of model rendering, in particular to a skeleton skin rendering method and system based on a graph neural network and an agent model, and the method comprises the steps: obtaining an adjacent cluster pair of a to-be-rendered skeleton skin model; constructing a triangular cluster graph taking the triangular patch clusters as nodes and the adjacency relation as edges; extracting cluster dynamic characteristics of each triangular patch cluster at the time point of the to-be-tested animation; processing cluster dynamic features through a cluster graph neural network containing local deformation attention and detail precision attention, and extracting deformation strength representation and detail retention value; and generating a rendering strategy according to the geometric error increment and the envelope variation predicted by the multi-branch adaptive proxy model, and performing dynamic optimization rendering on the to-be-rendered skeleton skin model. Compared with the prior art, the rendering precision of the local deformation area is optimized through graph structure modeling and an attention mechanism, adaptive strategy adjustment is achieved in combination with the proxy model, and the rendering efficiency and effect of the complex animation scene are improved.

Description

Technical Field

[0001] The present invention relates to the field of model rendering, and more specifically, to a rendering method and system for skinning based on a graph neural network and a surrogate model. Background Art

[0002] In scenes where the interaction between characters and the scene is more frequent, traditional occlusion culling strategies and frustum culling strategies often only focus on the visibility of the overall bounding volume of the character, lacking a more detailed culling strategy for local key parts with high distortion, and there may be a phenomenon that requires high computational load even at a long distance. For close-up shots that require high-precision rendering, although the existing technology can locally increase the precision based on the bone slots, its parameter scheduling lacks comprehensive consideration of bone pose changes, mesh distortion directions, and coupling degrees of adjacent regions, which is not conducive to the adaptive maintenance and iteration of long-time and multi-frame sequences. Summary of the Invention

[0003] The present invention aims to overcome at least one defect (shortcoming) of the above-mentioned prior art, and provides a rendering method and system for skinning based on a graph neural network and a surrogate model, which are used to optimize the rendering and culling of local key parts with high distortion and improve the rendering effect of the character model.

[0004] According to a first aspect of the present application, there is provided a rendering method for skinning based on a graph neural network and a surrogate model, the rendering method comprising:

[0005] Obtain all adjacent cluster pairs of the bone skinning model to be rendered; each of the adjacent cluster pairs contains two adjacent triangular patch clusters; the triangular patch cluster contains a number of connected triangular patches of the bone skinning model to be rendered;

[0006] Construct a triangular cluster graph according to all the adjacent cluster pairs; the nodes of the triangular cluster graph are triangular patch clusters, and the edges of the triangular cluster graph are the adjacency relationships between adjacent cluster pairs;

[0007] Obtain the displacement information matrix, deformation intensity, and rendering requirements of each triangular patch cluster at the time point of the animation to be measured;

[0008] Obtain the cluster dynamic features of each triangular patch cluster at the time point of the animation to be measured according to the displacement information matrix, deformation intensity, rendering requirements, and triangular cluster graph;

[0009] Process each of the cluster dynamic features by using a pre-trained cluster graph neural network to obtain the deformation intensity representation and detail retention value of each triangular patch cluster; the cluster graph neural network introduces local deformation attention and detail precision attention, and the cluster graph neural network extracts the deformation intensity representation and detail retention value of the cluster dynamic features through local deformation attention and detail precision attention;

[0010] According to the deformation intensity characterization, detail retention value, and displacement information matrix, use the pre-trained multi-branch adaptive proxy model to predict each triangle patch cluster, and obtain the geometric error increment and envelope change amount of each triangle patch cluster at the animation time point to be measured;

[0011] According to the geometric error increment and envelope change amount, obtain the rendering strategy at the animation time point to be measured, and render the skeleton skinning model to be rendered according to the rendering strategy.

[0012] Optionally, the obtaining of all adjacent cluster pairs of the skeleton skinning model to be rendered specifically includes:

[0013] Obtain all triangle patch clusters of the skeleton skinning model to be rendered;

[0014] According to the skeleton skinning model to be rendered, obtain the vertex coordinate set of each of the triangle patch clusters;

[0015] According to the vertex coordinate set, obtain the triangle patch clusters with a common boundary or a common vertex;

[0016] According to the triangle patch clusters with a common boundary or a common vertex, obtain the adjacent cluster pairs.

[0017] Optionally, the constructing of a triangle cluster graph according to all the adjacent cluster pairs specifically includes:

[0018] Use each of the triangle patch clusters as a node, use the adjacency relationship between the two triangle patch clusters in the adjacent cluster pairs as an edge, and set corresponding edge weights for each edge, and construct the triangle cluster graph according to each node and each edge with set edge weights;

[0019] The calculation of the edge weights includes:

[0020] Obtain the bone joint positions of each triangle patch cluster corresponding to the skeleton skinning model to be rendered;

[0021] According to the bone joint positions, obtain the joint correlation degree between the two triangle patch clusters in each of the adjacent cluster pairs;

[0022] According to the vertex coordinate set, obtain the vertex overlap degree between the two triangle patch clusters in each of the adjacent cluster pairs; the vertex coordinate set represents the set of vertices of the skeleton skinning model included in each triangular patch of the corresponding triangle patch cluster;

[0023] Obtain the dynamic distortion degree between the two triangle patch clusters in each of the adjacent cluster pairs; the dynamic distortion degree represents the bone deformation trend between the two triangle patch clusters;

[0024] Obtain the edge weights between the two triangular patch clusters in each of the adjacent cluster pairs according to the joint correlation, vertex overlap degree, and the dynamic distortion degree; the obtaining of the edge weights is expressed as:

[0025]

[0026] In the formula, W ij represents the adjacency weight between the triangular patch cluster C i and the triangular patch cluster C j ; A ij is the joint correlation degree, O ij is the vertex overlap degree, D ij is the dynamic distortion degree; α, β, and γ represent adjustment coefficients.

[0027] Optionally, the obtaining of the cluster dynamic features of each triangular patch cluster at the to-be-tested animation time point according to the displacement information matrix, deformation intensity, rendering requirement, and triangular cluster graph specifically includes:

[0028] The displacement information matrix includes a rotation matrix and a translation matrix;

[0029] Obtain the initial dynamic features of each triangular patch cluster at the to-be-tested animation time point according to the rotation matrix, translation matrix, deformation intensity, and rendering requirement; the obtaining of the initial dynamic features is expressed as:

[0030]

[0031] In the formula, t represents the to-be-tested animation time point; X i (t) represents the initial dynamic features of the triangular patch cluster C i ; Ri(t), Ti(t), Si(t), and Di(t) respectively represent the rotation matrix, translation matrix, deformation intensity, and rendering requirement of the triangular patch cluster i at the to-be-tested animation time point t; respectively represent the values after normalizing or numerically transforming Ri(t), Ti(t), Si(t), and Di(t);

[0032] Obtain the edge weights of each edge in the triangular cluster graph, and obtain the cluster dynamic features of each triangular patch cluster according to the edge weights and the initial dynamic features; the obtaining of the cluster dynamic features is expressed as:

[0033]

[0034] In the formula, X i ′ (t) represents the cluster dynamic features of the triangular patch cluster C i at the to-be-tested animation time point t; Represents the triangle face cluster C i The set of adjacent triangle face clusters; ω ij Represents the triangle face cluster C in the adjacent cluster pair i With the triangle patch cluster C j δ represents the coupling balance coefficient, which is used to control the balance of features transferred from the corresponding triangle cluster to the adjacent triangle cluster.

[0035] Optionally, the network structure of the cluster graph neural network is constructed according to each node and each edge of the triangular cluster graph; the network structure of the cluster graph neural network includes a multi-layer graph convolution unit and an attention mechanism unit; the attention mechanism unit sets the local deformation attention and detail accuracy attention;

[0036] The attention coefficient of the local deformation attention is:

[0037]

[0038] In the formula, a L (i,j) is the triangle face cluster C in the adjacent cluster pair i With the triangle patch cluster C j The attention coefficient of the local deformation attention; T ij Indicates the triangle face cluster C in the adjacent cluster pair at the animation time point to be tested i With the triangle patch cluster C j The bone distortion coupling strength between them; Δθi represents the triangle face cluster C i With the triangle patch cluster C j The angle difference of the corresponding bone joint in the direction of bone distortion; τ is the distortion strength judgment threshold, α L , β L are the adjustment coefficients respectively;

[0039] The attention coefficient of the detail precision attention is:

[0040]

[0041] In the formula, a D (i) is the triangular face cluster C i The attention coefficient of detail precision attention in the cluster graph neural network aggregation process, Di(t) represents the triangle face cluster C at the animation time point to be tested i , Vi(t) represents the triangle face cluster C at the animation time point to be tested i The current visible contour area; η is the offset coefficient, ζ, Ui(t) are the adjustment coefficients.

[0042] Optionally, according to the deformation intensity characterization, detail retention value, and displacement information matrix, use a pre-trained multi-branch adaptive proxy model to predict each triangular patch cluster, and obtain the geometric error increment and envelope change amount of each triangular patch cluster at the time point of the animation to be measured, specifically including:

[0043] Construct an adaptive input vector of the multi-branch adaptive proxy model according to the deformation intensity characterization, detail retention value, and displacement information matrix;

[0044] Use the pre-trained multi-branch adaptive proxy model to obtain the geometric error increment and envelope change amount of each triangular patch cluster at the time point of the animation to be measured according to the adaptive input vector;

[0045] Among them, the acquisition of the envelope change amount is as follows:

[0046]

[0047] In the formula, ΔB i (t) is the envelope change amount of the triangular patch cluster C i at the time point t of the animation to be measured, I i (t) is the adaptive input vector of the triangular patch cluster C i corresponding to the time point t of the animation to be measured, B i (t) represents the deformation intensity characterization of the triangular patch cluster C i at the time point t of the animation to be measured; α1 is the power function gain, and β1 is the denominator adjustment coefficient;

[0048] The acquisition of the geometric error increment is as follows:

[0049]

[0050] In the formula, ΔE i (t) is the geometric error increment of the triangular patch cluster C i at the time point t of the animation to be measured, D i (t) represents the rendering requirement of the triangular patch cluster C i at the time point t of the animation to be measured; α2 is the power function gain, and β2 is the denominator adjustment coefficient.

[0051] Optionally, according to the geometric error increment and envelope change amount, obtain the rendering strategy at the time point of the animation to be measured, specifically including:

[0052] Obtain the pixel contribution degree of each triangular patch cluster at the time point of the animation to be measured; the pixel contribution degree represents the screen coverage ratio of the corresponding triangular patch cluster at the time point of the animation to be measured;

[0053] Based on the geometric error increment and pixel contribution degree, obtain the level-of-detail selection index of each triangular patch cluster at the animation time point to be measured, expressed as:

[0054]

[0055] In the formula, L i (t) is the level-of-detail selection index of triangular patch cluster C i at the animation time point t to be measured, ΔE i (t) is the geometric error increment of triangular patch cluster C i at the animation time point t to be measured, P i (t) is the pixel contribution degree of triangular patch cluster C i at the animation time point t to be measured; is the exponential adjustment coefficient, and μ is the pixel contribution threshold;

[0056] Based on the envelope change amount and pixel contribution degree, obtain the visibility metric of each triangular patch cluster in the occlusion culling process during rendering at the animation time point to be measured, expressed as:

[0057]

[0058] In the formula, Θ i (t) is the visibility metric of triangular patch cluster C i at the animation time point t to be measured, ΔB i (t) is the envelope change amount of triangular patch cluster C i at the animation time point t to be measured, and σ is the power decay coefficient;

[0059] Respectively, based on the level-of-detail selection index and visibility metric, formulate the level-of-detail rendering strategy and visibility rendering strategy for the corresponding triangular patch cluster;

[0060] The level-of-detail rendering strategy is:

[0061] When the level-of-detail selection index is greater than or equal to the preset tolerance threshold, increase the rendering detail level of the corresponding triangular patch cluster; when the level-of-detail selection index is less than the preset tolerance threshold, keep or decrease the rendering detail level of the corresponding triangular patch cluster;

[0062] The visibility rendering strategy is:

[0063] When the visibility metric is less than or equal to the visibility threshold, determine that the corresponding triangular patch cluster is culled at the animation time point to be measured; when the visibility is independently greater than the visibility threshold, obtain a new envelope based on the corresponding envelope change amount and output it.

[0064] Optionally, before obtaining the rendering strategy at the time point of the animation to be measured according to the geometric error increment and the envelope change amount, the following steps are further included:

[0065] According to the geometric error increment and the envelope change amount, correct the attention coefficients of the local deformation attention and the detail precision attention, specifically including:

[0066] According to the geometric error increment and the envelope change amount, obtain the feedback input metric of each triangular patch cluster at the time point of the animation to be measured, expressed as:

[0067]

[0068] In the formula, M i (t) is the feedback input metric of the triangular patch cluster C i at the time point t of the animation to be measured, ΔB i (t) is the envelope change amount of the triangular patch cluster C i at the time point t of the animation to be measured, ΔE i (t) is the geometric error increment of the triangular patch cluster C i at the time point t of the animation to be measured; γ1 is the weight coefficient;

[0069] According to the feedback input metric, correct the attention coefficients of the local deformation attention and the detail precision attention, expressed as:

[0070]

[0071] In the formula, aL(i,j) and a Lr (i,j) are the attention coefficients of the local deformation attention before and after correction of the triangular patch cluster C i and the triangular patch cluster C j in the adjacent cluster pair; a D (i) and a Dr (i) are the attention coefficients of the detail precision attention before and after correction of the triangular patch cluster C i ; κ1, κ2 are used as update amplitude adjustments, and ρ is the power decay coefficient;

[0072] Feed back the attention coefficients of the corrected local deformation attention and detail precision attention to the cluster graph neural network.

[0073] Optionally, before obtaining the displacement information matrix, deformation intensity, and rendering requirements of each triangular patch cluster at the time point of the animation to be measured, the following steps are further included:

[0074] Periodically obtain the character pose changes and character rendering requirements of the skeleton skinned model to be rendered;

[0075] According to the character pose change and character rendering requirements, adjust the frequencies of obtaining the character pose change and character rendering requirements, and determine whether to execute the steps of obtaining the time points of the animation to be measured and subsequent steps according to the adjusted frequencies; specifically including:

[0076] According to the cycle, respectively obtain the character pose vectors of the current time point and the previous time point of the bone skinned model to be rendered, and obtain the character pose change at the current time point according to the character pose vectors of the current time point and the previous time point;

[0077] Obtain the character rendering requirements at the current time point of the bone skinned model to be rendered;

[0078] Construct a monitoring function according to the character pose change and character rendering requirements, expressed as:

[0079]

[0080] In the formula, ω(t) represents the monitoring function, Δθ r (t) is the character pose change, d(t) is the character rendering requirement; η1, η2 are monitoring coefficients, ∈1, ∈2 are monitoring thresholds;

[0081] Obtain the target adjustment frequency according to the monitoring function, expressed as:

[0082]

[0083] In the formula, f u (t) is the target adjustment frequency, f max is the preset maximum adjustment frequency, f min is the preset minimum adjustment frequency;

[0084] When the target adjustment frequency is greater than or equal to the preset frequency threshold, execute the steps of obtaining the time points of the animation to be measured and subsequent steps; when the target adjustment frequency is less than the preset frequency threshold, maintain the rendering strategy of the previous time point.

[0085] According to the second aspect of the present application, a rendering system for bone skinning based on a graph neural network and a surrogate model is provided. The rendering system includes:

[0086] An adjacent cluster pair acquisition module, configured to acquire all adjacent cluster pairs of the bone skinned model to be rendered; each of the adjacent cluster pairs includes two adjacent triangular patch clusters; the triangular patch clusters include several connected triangular patches of the bone skinned model to be rendered;

[0087] A cluster graph construction module, configured to construct a triangular cluster graph according to all the adjacent cluster pairs; the nodes of the triangular cluster graph are triangular patch clusters, and the edges of the triangular cluster graph are the adjacency relationships between adjacent cluster pairs;

[0088] A cluster information acquisition module, configured to acquire a displacement information matrix, a deformation intensity, and a rendering requirement of each triangular patch cluster at a time point of an animation to be measured;

[0089] A dynamic feature acquisition module, configured to acquire the cluster dynamic features of each triangular patch cluster at the time point of the animation to be measured according to the displacement information matrix, the deformation intensity, the rendering requirement, and the triangular cluster map;

[0090] A neural network processing module, configured to process each of the cluster dynamic features by using a pre-trained cluster map neural network to obtain a deformation intensity representation and a detail retention value of each triangular patch cluster; the cluster map neural network introduces local deformation attention and detail precision attention, and the cluster map neural network extracts the deformation intensity representation and the detail retention value of the cluster dynamic features through the local deformation attention and the detail precision attention;

[0091] An attitude prediction module, configured to predict each triangular patch cluster by using a pre-trained multi-branch adaptive proxy model according to the deformation intensity representation, the detail retention value, and the displacement information matrix to obtain a geometric error increment and an envelope change amount of each triangular patch cluster at the time point of the animation to be measured;

[0092] An attitude rendering module, configured to obtain a rendering strategy at the time point of the animation to be measured according to the geometric error increment and the envelope change amount, and render the to-be-rendered skeleton skin model according to the rendering strategy.

[0093] According to a third aspect of the present application, there is provided an electronic device, including:

[0094] A memory, configured to store one or more computer programs;

[0095] A processor, when the one or more computer programs are executed by the processor, implementing the rendering system of the skeleton skin based on the graph neural network and the proxy model described in the first aspect above.

[0096] According to a fourth aspect of the present application, there is provided a computer-readable storage medium, where the computer-readable storage medium stores computer instructions, and the computer instructions are used to implement the rendering system of the skeleton skin based on the graph neural network and the proxy model described in the first aspect above when executed by a processor.

[0097] Based on any of the above aspects, the rendering method, system, electronic device, and computer storage medium for skeletal skinning based on graph neural network and surrogate model provided by the embodiments of the present application divide adjacent cluster pairs and construct a triangular cluster graph, decompose the skeletal skinning model to be rendered into locally associated cluster units, and effectively achieve precise control of highly deformable joint regions. At the same time, a cluster graph neural network is constructed by combining local deformation attention and detail accuracy attention, enabling the cluster graph neural network to more attentively obtain highly distorted local detail features, thereby improving the accuracy of deformation modeling and focusing on the high-detail demand regions in the foreground. After identifying highly distorted and high-precision clusters, a multi-branch adaptive surrogate model is also used to predict the envelope change amount and geometric error increment for each cluster in real time, which helps to obtain a more refined error estimate in highly deformable regions and provides more accurate envelope boundary information for occlusion culling and frustum culling. Finally, a specific rendering strategy is formulated based on the envelope change amount and geometric error increment to optimize the rendering and culling of the skeletal skinning model to be rendered. BRIEF DESCRIPTION OF THE DRAWINGS

[0098] In order to more clearly illustrate the technical solutions in the embodiments of the present application, the following will briefly introduce the drawings required for the description of the embodiments. Obviously, the drawings in the following description are only some embodiments of the present application. For those of ordinary skill in the art, other drawings can be obtained based on these drawings without creative efforts.

[0099] Figure 1 FIG. is a schematic application scenario diagram of the rendering method provided in this embodiment.

[0100] Figure 2 FIG. is a flowchart of the steps of the rendering method provided in this embodiment.

[0101] Figure 3 FIG. is a flowchart of the steps for obtaining adjacent cluster pairs provided in this embodiment.

[0102] Figure 4 FIG. is a flowchart of the steps for obtaining the edge weights of the triangular cluster graph provided in this embodiment.

[0103] Figure 5 FIG. is a flowchart of the steps for obtaining the dynamic features of clusters provided in this embodiment.

[0104] Figure 6 FIG. is a flowchart of the steps for obtaining the geometric error increment and envelope change amount provided in this embodiment.

[0105] Figure 7 FIG. is a flowchart of the steps for obtaining the rendering strategy provided in this embodiment.

[0106] Figure 8Flowchart of the step for correcting the attention coefficients of local deformation attention and detail precision attention provided in this embodiment.

[0107] Figure 9 Flowchart of the step for obtaining the character pose change and adjusting the acquisition frequency of the character rendering requirement provided in this embodiment.

[0108] Figure 10 Schematic diagram of the functional modules of the rendering system provided in this embodiment.

[0109] Figure 11 Schematic diagram of the structure of the electronic device provided in this embodiment. Detailed implementation manners

[0110] The accompanying drawings of this application are only for illustrative purposes and should not be construed as a limitation to this application. To better illustrate the following embodiments, some components in the drawings will be omitted, enlarged or reduced, which do not represent the dimensions of the actual products; for those skilled in the art, it is understandable that some well-known structures and their descriptions in the drawings may be omitted.

[0111] In order to enable those skilled in the art to better understand the solution of this application, the technical solutions in the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings in the embodiments of this application. Obviously, the described embodiments are only a part of the embodiments of this application, rather than all of the embodiments. Based on the embodiments in this application, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the scope of protection of this application.

[0112] It should be noted that the terms "first", "second", etc. in the description and claims of this application and the above accompanying drawings are used to distinguish similar objects and do not necessarily need to describe a specific order or sequence. It should be understood that such used data can be interchanged under appropriate circumstances so that the embodiments of this application described here can be implemented in an order other than those illustrated or described here. In addition, the terms "include" and "have" and any variations thereof are intended to cover non-exclusive inclusion. For example, a process, method, system, product or device that includes a series of steps or units does not necessarily have to be limited to those steps or units clearly listed, but may include other steps or units not clearly listed or inherent to these process, method, product or device.

[0113] In the current field of game and animation production, the commonly adopted method for character driving is skeletal skinning. By configuring transformation matrices for each bone joint and then interpolating the mesh vertices according to weights, smooth deformation of the character's joints in different poses can be achieved. To better realize the real-time rendering of characters, Epic Games has introduced a real-time geometry rendering technology - Nanite in Unreal Engine 5 (UE5). The core of this technology is to allow users or developers to directly use film-level high-polygon models without relying on traditional manual optimization processes such as decimation and LOD (Level of Detail) grading, while maintaining high performance in real-time rendering.

[0114] Although existing character animation rendering technologies such as Nanite described above can achieve smooth visual effects in most regular action scenes using skeletal skinning, in more complex cases of joint twisting and large deformations, it is often difficult to balance the deformation accuracy of local areas of the character. To meet the diverse motion requirements, most common hierarchical processing methods focus on overall mesh accuracy switching, and the perception and real-time update of local high-frequency details often rely on manual or static parameter settings and cannot be flexibly adjusted in a dynamic environment. As the scene scale and the number of characters increase, using simple distance determination or unified threshold determination methods is likely to result in problems such as insufficient deformation or redundant calculations in some key action areas.

[0115] At the same time, in scenes where the interaction between the character and the scene is more frequent, traditional occlusion culling and frustum culling strategies often only focus on the visibility of the overall bounding volume of the character and lack a more refined culling strategy for highly distorted local key parts, which may result in the phenomenon of high computational load being occupied even at a long distance. For close-up shots that require high-precision rendering, although the existing technology can be achieved by locally magnifying the accuracy based on bone slots, its parameter scheduling lacks comprehensive consideration of bone pose changes, mesh distortion directions, and coupling degrees of adjacent regions, which is not conducive to the adaptive maintenance and iteration of long-term and multi-frame sequences.

[0116] This embodiment provides a technical solution that can solve the above problems. The following will describe the specific implementation manners of the present application in detail with reference to the accompanying drawings.

[0117] Exemplarily, it is a schematic diagram of an application scenario of a rendering method for skeletal skinning based on a graph neural network and a proxy model provided by an embodiment of the present application. As Figure 1 shown, the application scenario at least includes a server 100 and a terminal 200 that can communicate with the server 100.

[0118] It is understandable that the server 100 can be an independent electronic device or a cluster composed of multiple electronic devices; the terminal 200 can be a smart phone terminal, a personal computer, a tablet computer, a vehicle-mounted terminal, etc., but is not limited thereto.

[0119] In an implementable manner, the server 100 and the terminal 200 can respectively execute the rendering method of skeletal skinning based on graph neural network and proxy model provided by the embodiments of the present application. Alternatively, optionally, part of the rendering method of skeletal skinning based on graph neural network and proxy model provided by the embodiments of the present application is executed in the server 100, and part is executed in the terminal 200.

[0120] As Figure 2 shown, the present embodiment provides a rendering method of skeletal skinning based on graph neural network and proxy model, which may include the following steps:

[0121] S1: Obtain all adjacent cluster pairs of the skeletal skinning model to be rendered;

[0122] Specifically, as Figure 3 shown, the specific steps of step S1 may include the following steps:

[0123] S11: Obtain all triangular patch clusters of the skeletal skinning model to be rendered;

[0124] In this embodiment, each of the adjacent cluster pairs contains two adjacent triangular patch clusters; the triangular patch cluster contains several triangular patches with a connectivity relationship of the skeletal skinning model to be rendered;

[0125] Specifically, the triangular patch mainly refers to the basic geometric unit forming the skeletal skinning model to be rendered, which can be understood as a planar geometric region formed by three vertices linked pairwise, and the skeletal skinning model to be rendered is formed by splicing a large number of the triangular patches.

[0126] In this embodiment, the triangular patch cluster can be denoted as C = {C1, C2,..., C k}, C k represents the triangular patch cluster numbered k; by forming triangular patch clusters with connected triangular patches, the skeletal influence ranges corresponding to the triangular patches inside each formed triangular patch cluster are relatively consistent. At the same time, numbering the divided triangular patch clusters can facilitate sequential processing in subsequent steps.

[0127] S12: According to the skeletal skinning model to be rendered, obtain the vertex coordinate set of each of the triangular patch clusters;

[0128] As described above, the vertex coordinate set of the triangular patch cluster is the set of vertex coordinates of the three vertices of each triangular patch in the triangular patch cluster. The coordinates of the vertices can be obtained by constructing a reference coordinate system in the to-be-rendered skinned model or the scene where the to-be-rendered skinned model is located, which will not be elaborated further here.

[0129] S13: Obtain a triangular patch cluster with a common boundary or a common vertex according to the vertex coordinate set;

[0130] It can be understood that, through specific vertex coordinates, it can be determined whether two corresponding triangular patch clusters have a common boundary or a common vertex. And since there is a connectivity relationship between the triangular patches in the triangular patch cluster, it can be considered that if two triangular patch clusters have a common boundary or a common vertex, it means that the two triangular patch clusters are adjacent to each other.

[0131] S14: Obtain the adjacent cluster pairs according to the triangular patch cluster with a common boundary or a common vertex.

[0132] After determining the adjacency relationship of each triangular patch cluster in step S13, all the adjacent cluster pairs in all the triangular patch clusters can be obtained according to the adjacency relationship.

[0133] In one implementation, the triangular patch cluster can be divided according to the small-scale cluster standard required in Nanite. For the convenience of description and understanding, the following description of this embodiment is based on the benchmark of Nanite technology. It should be noted that the rendering method of this embodiment is not limited to being applied in Nanite technology. It can be understood that the rendering method of this embodiment can be applied to various technologies for animating character rendering using skinned bones.

[0134] S2: Construct a triangular cluster graph according to all the adjacent cluster pairs; wherein, the nodes of the triangular cluster graph are triangular patch clusters, and the edges of the triangular cluster graph are the adjacency relationships between adjacent cluster pairs;

[0135] The construction of the triangular cluster graph can be specifically as follows:

[0136] Taking each triangular patch cluster as a node, taking the adjacency relationship between the two triangular patch clusters in the adjacent cluster pair as an edge, and setting corresponding edge weights for each edge, and constructing the triangular cluster graph according to each node and each edge with set edge weights;

[0137] Specifically, as Figure 4 shown, the obtaining of the edge weights can specifically include the following steps:

[0138] S21: Obtain the bone joint positions corresponding to each triangular patch cluster in the to-be-rendered skinned model;

[0139] In this embodiment, the acquisition of the bone joint positions can be as follows:

[0140] Extract the weight coefficients between each vertex in each vertex coordinate set and each bone of the bone skinning model to be rendered, which can be denoted as W(v, b), where v is the vertex number and b is the bone number. By extracting the weight coefficients between the vertex and the bone, the mapping relationship between the vertex and the bone is obtained.

[0141] According to the weight coefficients, obtain the bones corresponding to each vertex of each triangular patch cluster; then obtain the hierarchical relationship of the bones of the bone skinning model to be rendered. Based on this hierarchical relationship, the vertex and its corresponding bone, determine the bone joint positions where each triangular patch cluster is located. It can be understood that the hierarchical relationship can represent the connection relationship of each bone of the bone skinning model to be rendered, and the bone joint position represents the specific connection position between the bones. Since the bone joint position is the main position where the bone skinning model to be rendered generates distorted deformation, obtaining the bone joint position can effectively reflect the distorted deformation situation of each triangular patch cluster in the subsequent calculation process.

[0142] S22: According to the bone joint positions, obtain the joint correlation degree between the two triangular patch clusters in each adjacent cluster pair;

[0143] S23: According to the vertex coordinate set, obtain the vertex overlap degree between the two triangular patch clusters in each adjacent cluster pair;

[0144] It can be understood that in this embodiment, the joint correlation degree represents the overlapping situation of the influence ranges of the corresponding triangular patch clusters in the bone space, which can be understood as the overlapping situation of the bone joints corresponding to the corresponding triangular patch clusters; the vertex overlap degree represents the overlapping situation of the vertices between the vertex coordinate sets of the corresponding triangular patch clusters.

[0145] S24: Obtain the dynamic distortion degree between the two triangular patch clusters in each adjacent cluster pair; the dynamic distortion degree represents the bone deformation trend between the two triangular patch clusters;

[0146] In this embodiment, the dynamic distortion degree is used to measure the coupling trend of the deformation of the bones of the corresponding two triangular patch clusters during strong torsion.

[0147] S25: According to the joint correlation degree, vertex overlap degree and the dynamic distortion degree, obtain the edge weight between the two triangular patch clusters in each adjacent cluster pair; the acquisition of the edge weight is expressed as:

[0148]

[0149] In the formula, W ijRepresents the triangular patch cluster C in the adjacent cluster pair i And the triangular patch cluster C j The adjacent weight between them, A ij Is the joint correlation degree, O ij Is the vertex overlap degree, D ij Is the dynamic distortion degree; α, β, γ represent adjustment coefficients, γ is used to adjust the influence of the dynamic distortion degree, β is used to adjust the influence of the vertex overlap degree, and α is used to adjust the overall influence of the joint correlation degree, vertex overlap degree and dynamic distortion degree;

[0150] In this embodiment, in practical applications, the A ij , O ij , D ij Are all mapped to [0, 1] in a normalized manner, And (O ij ) 2 Reflects the quadratic sensitivity to the joint influence and the cluster overlap degree; (γ + D ij )2 is mainly for amplification processing when the bones are greatly deformed; α, β, γ are used to balance the influences of the joint correlation degree, vertex overlap degree and dynamic distortion degree in three aspects.

[0151] S3: Obtain the displacement information matrix, deformation intensity and rendering requirements of each triangular patch cluster at the time point of the animation to be measured;

[0152] In this embodiment, the time point of the animation to be measured represents the time point when the bone skinned model to be rendered needs to be rendered. It can be understood that the time point of the animation to be measured can be any time point in the animation time sequence when the bone skinned model to be rendered is performing an animation action. Therefore, a time point of the animation to be measured corresponds to a set of displacement information matrix, deformation intensity and rendering requirements of the bone skinned model to be rendered.

[0153] Among them, the displacement information matrix includes a rotation matrix and a translation matrix; it can be understood that the rotation matrix represents the distortion situation of the corresponding triangular patch cluster relative to the previous time point, and the translation matrix represents the displacement situation of the corresponding triangular patch cluster relative to the previous time point. The rotation matrix and translation matrix of the triangular patch cluster can be obtained through the rotation matrix and translation matrix of the corresponding bone joint respectively.

[0154] The deformation intensity represents the deformation degree of the local bone joint corresponding to the triangular patch cluster, and can be calculated through the rotation matrix and translation matrix of the corresponding bone joint.

[0155] After obtaining the rotation matrix and translation matrix of the bone joint, as well as the local deformation degree of the bone joint, they can be stored in a preset data buffer for convenient subsequent call and calculation.

[0156] The rendering requirement represents the required rendering detail accuracy of the corresponding triangle patch cluster at the time point of the animation to be measured, and the rendering requirement can be set according to the distance between the corresponding triangle patch cluster and the set camera.

[0157] S4: Obtain the cluster dynamic features of each triangle patch cluster at the time point of the animation to be measured according to the displacement information matrix, the deformation intensity, the rendering requirement, and the triangle cluster map;

[0158] In this embodiment, as Figure 5 shown, the obtaining of the cluster dynamic features may specifically include the following steps:

[0159] S41: Obtain the initial dynamic features of each triangle patch cluster at the time point of the animation to be measured according to the rotation matrix, the translation matrix, the deformation intensity, and the rendering requirement;

[0160] In this embodiment, the obtaining of the initial dynamic features may be expressed as:

[0161]

[0162] In the formula, t represents the time point of the animation to be measured; X i (t) represents the initial dynamic features of the triangle patch cluster C at the time point t of the animation to be measured i ; Ri(t), Ti(t), Si(t), and Di(t) respectively represent the rotation matrix, the translation matrix, the deformation intensity, and the rendering requirement of the triangle patch cluster i at the time point t of the animation to be measured; respectively represent the values after normalizing or numerically transforming Ri(t), Ti(t), Si(t), and Di(t);

[0163] S42: Obtain the edge weights of each edge in the triangle cluster map, and obtain the cluster dynamic features of each triangle patch cluster according to the edge weights and the initial dynamic features;

[0164] In this embodiment, the obtaining of the cluster dynamic features may be expressed as:

[0165]

[0166] In the formula, X i ′ (t) represents the cluster dynamic features of the triangle patch cluster C at the time point t of the animation to be measured i ; represents the set of triangle patch clusters adjacent to the triangle patch cluster C i ; ω ij represents the triangle patch cluster C in the adjacent cluster pair i and the triangle patch cluster C jThe adjacency weight between them; δ represents the coupling balance coefficient, which is used to control the balance of the corresponding triangular patch cluster transferring features to the adjacent triangular patch cluster.

[0167] S5: Process the dynamic features of each of the clusters by using a pre-trained cluster graph neural network to obtain the deformation intensity representation and detail retention value of each triangular patch cluster;

[0168] In this embodiment, the local deformation attention and detail precision attention are introduced into the cluster graph neural network. The cluster graph neural network extracts the deformation intensity representation and detail retention value of the cluster dynamic features through the local deformation attention and detail precision attention;

[0169] It can be understood that the network structure of the cluster graph neural network is set according to the graph structure of the triangular cluster graph, that is, according to the set of each node, the set of each edge, and the set of the edge weights corresponding to the edges in the triangular cluster graph, the network structure of the cluster graph neural network is constructed; the network structure of the cluster graph neural network includes multiple graph convolutional units and an attention mechanism unit; the local deformation attention and detail precision attention are set in the attention mechanism unit.

[0170] In this embodiment, the attention coefficient of the local deformation attention is:

[0171]

[0172] In the formula, a L (i,j) is the attention coefficient of the local deformation attention of the triangular patch cluster C i in the adjacent cluster pair and the triangular patch cluster C j ; T j represents the bone twisting coupling strength between the triangular patch cluster C i and the triangular patch cluster C j in the adjacent cluster pair at the time point of the animation to be measured; Δθi represents the angle difference between the bone joints corresponding to the triangular patch cluster C i and the triangular patch cluster C j in the bone twisting direction; τ is the twisting strength determination threshold, and α L , β L are adjustment coefficients respectively;

[0173] It can be understood that in this embodiment, for the area near the highly deformed bone joint, the local deformation attention is set to measure the assignment coefficient when the corresponding triangular patch cluster receives feature information from the adjacent triangular patch cluster. By introducing the angle difference Δθi between the bone joints in the bone twisting direction and the bone twisting coupling strength T ij , it can effectively ensure that the local area strongly affected by the bone joint obtains a higher weight.

[0174] The attention coefficient of the detailed precision attention is as follows:

[0175]

[0176] where a D (i) is the attention coefficient of the detailed precision attention during the aggregation process in the cluster graph neural network. Di(t) represents the rendering requirement of the triangular patch cluster C i at the time point of the animation to be measured. Vi(t) represents the current visible contour area of the triangular patch cluster C i at the time point of the animation to be measured; η is the offset coefficient, and ζ, Ui(t) are adjustment coefficients; i

[0177] It can be understood that the larger the value of the attention coefficient of the detailed precision attention, the higher the attention that the corresponding triangular patch cluster should receive during the aggregation process of the cluster graph neural network. The sum of the rendering requirement Di(t) and the adjustment coefficient Ui(t) improves the attention in scenarios with high-precision requirements, and the current visible contour area Vi(t) is used to suppress large areas with low correlation.

[0178] After obtaining the attention coefficients of the local deformation attention and the detailed precision attention, during the update stage of each node of the cluster graph neural network, the attention coefficient aL(i,j) of the local distortion attention is multiplied by the attention coefficient aD(i) of the detailed visibility attention as the final message passing coefficient, which is used to correct the information intensity i transmitted from the neighbor nodes of the triangular patch cluster C i to the triangular patch cluster C

[0179] It can be understood that the training of the cluster graph neural network can refer to the above description, construct a corresponding triangular cluster graph for training, and train the cluster graph neural network through the triangular cluster graph for training. The training can be implemented in the conventional training method of the graph neural network, which will not be elaborated further here.

[0180] S6: According to the deformation intensity characterization, the detail retention value, and the displacement information matrix, use the pre-trained multi-branch adaptive proxy model to predict each triangular patch cluster, and obtain the geometric error increment and the envelope change amount of each triangular patch cluster at the time point of the animation to be measured;

[0181] ​In this embodiment, the multi-branch adaptive proxy model includes an initial layer network, a gating function, and multiple branch networks; among them, the initial layer network is used to extract features from the data input to the adaptive proxy model, and the branch networks are used to process the extracted features; the gating function is used to select the corresponding branch network for processing according to a preset rule for the extracted features; the multiple branch networks are stratified according to the prediction accuracy requirements of high-distortion and high-detail demand scenarios, and can be divided into shallow networks and deep networks.

[0182] In this embodiment, each branch network contains different numbers of neuron layers and weight parameters, and is used to hierarchically process scenarios of mild deformation and extreme deformation of the corresponding triangular patch clusters, that is, the setting of the branch network aims to balance the computational cost in low-deformation scenarios and the prediction accuracy in high-deformation scenarios.

[0183] As described above, the gating function is used to determine the branch network entered by each triangular patch cluster at the time point of the animation to be measured, and specifically may include:

[0184] Concatenate the rotation moment and translation matrix of the triangular patch cluster to obtain the bone pose deformation amount;

[0185] Compare the bone pose deformation amount and the detail retention value of the corresponding triangular patch cluster with a preset first gating threshold and a second gating threshold respectively, and select the corresponding branch network according to the comparison result.

[0186] Exemplarily, if the bone pose deformation amount is greater than the first gating threshold and the detail retention value is greater than the second gating threshold, select a deep branch network to improve the prediction accuracy for high-distortion and high-detail demand scenarios; if the bone pose deformation amount is less than or equal to the first gating threshold and the detail retention value is less than or equal to the second gating threshold, select a shallow branch network for processing to reduce the computational cost.

[0187] Specifically, as Figure 6 shown, the acquisition of the geometric error increment and the envelope change amount of each triangular patch cluster may specifically include the following steps:

[0188] S61: Construct an adaptive input vector of the multi-branch adaptive proxy model according to the deformation intensity representation, detail retention value, and displacement information matrix;

[0189] In this embodiment, the adaptive input vector is obtained by concatenating the deformation intensity representation, detail retention value, and displacement information matrix.

[0190] S62: Using the pre-trained multi-branch adaptive proxy model, obtain the geometric error increment and the envelope change amount of each triangular patch cluster at the animation time point to be measured according to the adaptive input vector; the envelope change amount represents the change amount of the main outer envelope size of the corresponding triangular patch cluster at the animation time point to be measured, and the geometric error increment represents the deviation degree of the geometric accuracy of the corresponding triangular patch cluster at the animation time point to be measured; the envelope change amount is used for subsequent determination of specific occlusion culling and frustum culling, and the geometric error increment is used to assist in the accuracy determination of Nanite when switching between different levels of detail;

[0191] Among them, the acquisition of the envelope change amount is as follows:

[0192]

[0193] In the formula, ΔB i (t) is the envelope change amount of the triangular patch cluster C i at the animation time point t to be measured, I i (t) is the adaptive input vector of the triangular patch cluster C i at the animation time point t to be measured, B i (t) represents the deformation intensity characterization of the triangular patch cluster C i at the animation time point t to be measured; α1 is the power function gain, and β1 is the denominator adjustment coefficient; it can be understood that the exponential function is used in the numerator part to improve the envelope estimation of the high-deformation region, and the denominator part amplifies Zi(t) through the power function to suppress the ineffective update in the low-deformation scene;

[0194] The acquisition of the geometric error increment is as follows:

[0195]

[0196] In the formula, ΔE i (t) is the geometric error increment of the triangular patch cluster C i at the animation time point t to be measured, D i (t) represents the rendering requirement of the triangular patch cluster C i at the animation time point t to be measured; α2 is the power function gain, and β2 is the denominator adjustment coefficient. Among them, the power function exponent improves the response strength of the high-demand cluster in terms of geometric error, and the denominator term combines the square of the rendering requirement D i (t) to suppress unnecessary misjudgment and prevent the network from making excessive corrections to low-detail demand segments.

[0197] After obtaining the geometric error increment and the envelope change amount of each triangular patch cluster, the set increment and the envelope change amount of each triangular patch cluster are summarized in real time as the basis for the subsequent rendering strategy.

[0198] S7: Obtain the rendering strategy at the animation time point to be measured according to the geometric error increment and the change amount of the envelope, and render the bone skin model to be rendered according to the rendering strategy.

[0199] In this embodiment, as Figure 7 shown, the obtaining of the rendering strategy may specifically include the following steps:

[0200] S71: Obtain the pixel contribution degree of each triangle patch cluster at the animation time point to be measured;

[0201] Specifically, the pixel contribution degree represents the screen coverage ratio of the corresponding triangle patch cluster at the animation time point to be measured, and the pixel contribution degree can be obtained from the rendering viewport of the rendering engine;

[0202] S72: Obtain the level-of-detail selection index of each triangle patch cluster at the animation time point to be measured according to the geometric error increment and the pixel contribution degree;

[0203] In this embodiment, the obtaining of the level-of-detail selection index may be expressed as:

[0204]

[0205] In the formula, L i (t) is the level-of-detail selection index of triangle patch cluster C i at the animation time point t to be measured, ΔE i (t) is the geometric error increment of triangle patch cluster C i at the animation time point t to be measured, P i (t) is the pixel contribution degree of triangle patch cluster C i at the animation time point t to be measured; is the exponential adjustment coefficient, and μ is the pixel contribution threshold. It can be understood that the square of the geometric error increment ΔEi(t) is adopted in the numerator part to amplify the influence of high-error clusters on the level-of-detail selection. The denominator part combines the exponential comparison between the pixel contribution degree P i (t) and the threshold μ. When the pixel contribution degree P i (t) exceeds μ, the exponential term reduces the denominator, prompting the level-of-detail selection index Li(t) to rise, thereby triggering a finer level of detail.

[0206] S73: Obtain the visibility metric of each triangle patch cluster during the rendering occlusion culling process at the animation time point to be measured according to the change amount of the envelope and the pixel contribution degree;

[0207] In this embodiment, the obtaining of the visibility metric may be expressed as:

[0208]

[0209] where Θ i (t) is the visibility measure of the triangle patch cluster C i at the animation time point t to be measured, and ΔB i (t) is the change amount of the bounding volume of the triangle patch cluster C i at the animation time point t to be measured, and σ is the power decay coefficient. It can be understood that the numerator part uses the change amount ΔBi(t) of the bounding volume of the triangle patch cluster. When the change amount ΔBi(t) of the bounding volume increases, Θi(t) shows an upward trend; the denominator part uses the power of the pixel contribution degree P i (t) to decay and suppress the over-occlusion judgment, ensuring that the triangle patch cluster in the high pixel contribution degree scenario will not be underestimated.

[0210] S74: Respectively formulate the level-of-detail rendering strategy and visibility rendering strategy for the corresponding triangle patch cluster according to the level-of-detail selection index and visibility measure;

[0211] Specifically, in this embodiment, the level-of-detail rendering strategy is:

[0212] When the level-of-detail selection index is greater than or equal to the preset tolerance threshold, increase the rendering detail level of the corresponding triangle patch cluster; when the level-of-detail selection index is less than the preset tolerance threshold, keep or reduce the rendering detail level of the corresponding triangle patch cluster;

[0213] The visibility rendering strategy is:

[0214] When the visibility measure is less than or equal to the visibility threshold, it is determined that the corresponding triangle patch cluster is culled at the animation time point to be measured; when the visibility is greater than the visibility threshold, a new bounding volume is obtained according to the corresponding change amount of the bounding volume and output.

[0215] Finally, according to the rendering strategy obtained in step S74, determine the rendering detail level and visibility of each triangle patch cluster of the bone skinned model to be rendered, and use the rendering detail level and visibility as the rendering reference for the bone skinned model to be rendered at the next time point corresponding to the animation time point to be measured.

[0216] It can be considered that the rendering method in this embodiment takes the animation time point to be measured as the current time point, obtains the triangle cluster map and the character pose situation of the bone skinned model to be rendered at the current time point, and according to the triangle cluster map and the character pose situation, predicts the rendering detail level and visibility at the next time point through the constructed cluster map neural network and multi-branch adaptive proxy model, so that when the bone skinned model to be rendered is rendered at the next time point, it can adjust the rendering details of the corresponding triangle patch cluster according to the predicted rendering detail level, and at the same time, according to the predicted visibility, perform occlusion culling and frustum culling on the corresponding triangle patch cluster.

[0217] In an implementation manner of this embodiment, before executing step S7, the following steps may also be executed first:

[0218] According to the geometric error increment and the envelope change amount, correct the attention coefficients of the local deformation attention and the detail precision attention.

[0219] Specifically, as Figure 8 shown, the correction of the attention coefficients of the local deformation attention and the detail precision attention may specifically include:

[0220] A1: According to the geometric error increment and the envelope change amount, obtain the feedback input metric of each triangular patch cluster at the animation time point to be measured;

[0221] In this embodiment, the obtaining of the feedback input metric may be expressed as:

[0222]

[0223] In the formula, M i (t) is the feedback input metric of the triangular patch cluster C i at the animation time point t to be measured, ΔB i (t) is the envelope change amount of the triangular patch cluster C i at the animation time point t to be measured, ΔE i (t) is the geometric error increment of the triangular patch cluster C i at the animation time point t to be measured; γ1 is a weight coefficient; it can be understood that the feedback input metric M i (t) expands as the envelope change amount ΔB i (t) increases. When the geometric error increment ΔEi(t) rises, in the integral term, (ΔE i (t)) 2 makes Mi(t) show non-linear accumulation, and combines exp(-τ2) to attenuate the large-range envelope.

[0224] A2: Correct the attention coefficients of the local deformation attention and the detail precision attention according to the feedback input metric;

[0225] In this embodiment, the correction may be specifically expressed as:

[0226]

[0227] In the formula, aL(i,j) and a Lr (i,j) are respectively the attention coefficients of the local deformation attention before and after correction of the triangular patch cluster C i and the triangular patch cluster C j in the adjacent cluster pair; a D (i) and aDr (i) Respectively, for the triangular patch cluster C i The attention coefficients of the detail precision attention before and after correction; κ1 and κ2 are used as the update amplitude adjustment, and ρ is the power decay coefficient;

[0228] A3: Feed back the attention coefficients of the corrected local deformation attention and detail precision attention to the cluster graph neural network.

[0229] The cluster graph neural network updates the attention coefficients of the original local deformation attention and detail precision attention according to the attention coefficients of the corrected local deformation attention and detail precision attention. The updated cluster graph neural network is used to predict the deformation intensity representation and detail retention value of each triangular patch cluster at the next animation time point to be measured.

[0230] By obtaining the feedback input metric to update the local deformation attention and detail precision attention in the cluster graph neural network, the cluster graph neural network can use the information of the previous moment to predict the information of the next moment.

[0231] It can be understood that when the pose change of the bone skinning model to be rendered is large, it means that there is more pose change information contained in a short time range, and higher dynamic rendering is required. Therefore, it is necessary to shorten the time interval for obtaining the pose change of the bone skinning model to be rendered; on the contrary, when the pose change of the bone skinning model to be rendered is small, it means that there is less pose change information contained in a long time range, and lower dynamic rendering is required. Therefore, it is necessary to increase or maintain the time interval for obtaining the pose change of the bone skinning model to be rendered.

[0232] Therefore, in an implementation manner of this embodiment, before obtaining the displacement information matrix, deformation intensity, and rendering requirement of each triangular patch cluster at the time point of the animation to be measured, it may further include adjusting the acquisition period of the bone skinning model information to be rendered, such as Figure 9 shown, specifically including:

[0233] B1: Periodically obtain the character pose change and character rendering requirement of the bone skinning model to be rendered;

[0234] It can be understood that the period represents the frequency of obtaining the displacement information matrix, deformation intensity, and rendering requirement of the bone skinning model to be rendered.

[0235] B2: According to the character pose change and character rendering requirement, adjust the frequency of obtaining the character pose change and character rendering requirement, and determine whether to execute the steps of obtaining the time point of the animation to be measured and subsequent steps according to the adjusted frequency; specifically including:

[0236] B3: Obtain the character pose vectors of the to-be-rendered skinned model at the current time point and the previous time point respectively according to the period, and obtain the character pose change at the current time point according to the character pose vectors at the current time point and the previous time point;

[0237] B4: Obtain the character rendering requirement of the to-be-rendered skinned model at the current time point;

[0238] B5: Construct a monitoring function according to the character pose change and the character rendering requirement, expressed as:

[0239]

[0240] In the formula, ω(t) represents the monitoring function, Δθ r (t) is the character pose change, d(t) is the character rendering requirement; η1, η2 are monitoring coefficients, and ∈1, ∈2 are monitoring thresholds;

[0241] It can be understood that in a dynamic rendering scene, the character pose change and the character rendering requirement (distance from the camera) of the to-be-rendered skinned model have a greater impact on the switching of the level of detail and the rendering quality. Therefore, a piecewise monitoring function is constructed to comprehensively judge the character pose change and the character rendering requirement.

[0242] B6: Obtain the target adjustment frequency according to the monitoring function, expressed as:

[0243]

[0244] In the formula, f u (t) is the target adjustment frequency, f max is the preset maximum adjustment frequency, f min is the preset minimum adjustment frequency;

[0245] B7: When the target adjustment frequency is greater than or equal to the preset frequency threshold, execute the steps at and after the to-be-measured animation time point; when the target adjustment frequency is less than the preset frequency threshold, maintain the rendering strategy of the previous time point.

[0246] It can be understood that for dynamic scenes with large changes in character poses and high rendering requirements, it is necessary to increase the acquisition frequency to ensure that in the current scene, the deformation accuracy of each local area of the bone skinning model to be rendered can be better reflected. Therefore, when the target adjustment frequency is greater than or equal to the preset frequency threshold, the current time point is used as the time point of the animation to be measured, and the above steps S3 - S7 are executed to obtain a specific rendering strategy to render the bone skinning model to be rendered. On the contrary, when the target adjustment frequency is less than the preset frequency threshold and it is considered that no significant character pose change is detected, the acquisition frequency is maintained or reduced to save computing resources and optimize rendering performance.

[0247] Simply put, for variable motion scenes, by increasing the acquisition frequency, the attention to the bone skinning model to be rendered is increased, and the dynamic rendering effect is improved; for motion scenes with little change, by reducing the acquisition frequency, the attention to the bone skinning model to be rendered is reduced to save computing resources. As Figure 10 shown, the embodiment of the present application also provides a rendering system for bone skinning based on a graph neural network and a proxy model. Optionally, the rendering system may include:

[0248] An adjacent cluster pair acquisition module 11, configured to acquire all adjacent cluster pairs of the bone skinning model to be rendered;

[0249] In this embodiment, the cluster pair acquisition module 11 can be used to execute Figure 2 the step S1 shown, and the specific description of the cluster pair acquisition module 11 can refer to the description of the step S1.

[0250] A cluster graph construction module 12, configured to construct a triangular cluster graph according to all the adjacent cluster pairs;

[0251] In this embodiment, the cluster graph construction module 12 can be used to execute Figure 2 the step S2 shown, and the specific description of the cluster graph construction module 12 can refer to the description of the step S2.

[0252] A cluster information acquisition module 13, configured to acquire the displacement information matrix, deformation intensity, and rendering requirements of each triangular patch cluster at the time point of the animation to be measured;

[0253] In this embodiment, the cluster information acquisition module 13 can be used to execute Figure 2 the step S3 shown, and the specific description of the cluster information acquisition module 13 can refer to the description of the step S3.

[0254] The dynamic feature acquisition module 14 is configured to obtain the cluster dynamic features of each triangular patch cluster at the time point of the animation to be measured according to the displacement information matrix, the deformation intensity, the rendering requirement, and the triangular cluster graph;

[0255] In this embodiment, the dynamic feature acquisition module 14 can be used to execute Figure 2 Step S4 shown in the figure. For the specific description of the dynamic feature acquisition module 14, reference can be made to the description of step S4.

[0256] The neural network processing module 15 is configured to process each of the cluster dynamic features by using a pre-trained cluster graph neural network to obtain the deformation intensity representation and the detail retention value of each triangular patch cluster;

[0257] In this embodiment, the neural network processing module 15 can be used to execute Figure 2 Step S5 shown in the figure. For the specific description of the neural network processing module 15, reference can be made to the description of step S5.

[0258] The pose prediction module 16 is configured to predict each triangular patch cluster by using a pre-trained multi-branch adaptive proxy model according to the deformation intensity representation, the detail retention value, and the displacement information matrix to obtain the geometric error increment and the envelope change amount of each triangular patch cluster at the time point of the animation to be measured;

[0259] In this embodiment, the pose prediction module 16 can be used to execute Figure 2 Step S6 shown in the figure. For the specific description of the pose prediction module 16, reference can be made to the description of step S6.

[0260] The pose rendering module 17 obtains a rendering strategy at the time point of the animation to be measured according to the geometric error increment and the envelope change amount, and renders the skeleton skin model to be rendered according to the rendering strategy.

[0261] In this embodiment, the pose rendering module 17 can be used to execute Figure 2 Step S7 shown in the figure. For the specific description of the pose rendering module 17, reference can be made to the description of step S7.

[0262] In one implementation, the neural network processing module 15 can also be used to correct the attention coefficients of the local deformation attention and the detail precision attention according to the geometric error increment and the envelope change amount. For the specific description, reference can be made to Figure 8 The description of steps A1 - A3.

[0263] In one embodiment, the rendering system may further include a frequency adjustment module 18, which is used to adjust the acquisition frequencies of the character pose changes and character rendering requirements of the skeleton skinning model to be rendered;

[0264] The frequency adjustment module 18 can be used to execute Figure 9 the steps B1 - B7 shown. For the specific description of the frequency adjustment module 18, reference can be made to the description of the steps B1 - B7.

[0265] It can be understood that the above device embodiments and the above method embodiments can correspond to each other. Similar descriptions of the device embodiments can refer to the method embodiments. To avoid repetition, they will not be elaborated here. The rendering system for skeleton skinning based on a graph neural network and a surrogate model provided by the embodiments of the present application can execute the rendering method for skeleton skinning based on a graph neural network and a surrogate model provided by any embodiment of the present application, and has the corresponding functional modules and beneficial effects for executing the method. The functional modules of the rendering system for skeleton skinning based on a graph neural network and a surrogate model can be implemented in the form of hardware, can be implemented by instructions in the form of software, and can also be implemented by a combination of hardware and software modules.

[0266] Specifically, each step of the method embodiment of the present application can be completed by the integrated logic circuit in the hardware of the processor and / or instructions in the form of software. Combining the steps of the rendering method for skeleton skinning based on a graph neural network and a surrogate model of the embodiments of the present application can be directly embodied as being completed by the hardware - encoded processor, or completed by a combination of the hardware and software modules in the encoded processor. Optionally, the software module can be located in a random access memory, a read - only memory, a programmable read - only memory, a flash memory, an electrically erasable programmable memory, a register, and other storage media. The storage media is located in the memory, and the processor reads the information in the memory and combines its hardware to complete the steps in the above method embodiments.

[0267] The embodiments of the present application provide an electronic device 20, whose structure is as Figure 11 shown. The electronic device 20 can be the server 100 or the terminal 200 shown in this embodiment Figure 1 shown.

[0268] As Figure 11 shown, the electronic device 20 includes a memory 21, a processor 22, a communication module 23, and an input / output interface 24, etc. Optionally, the memory 21, the processor 22, the communication module 23, and the input / output interface 24 can be connected and communicate through a bus 25.

[0269] The memory 21 is used to store one or more computer programs and transfer the code of the computer program to the processor 22; when the one or more computer programs are executed by the processor 22, the rendering method of skeleton skinning based on graph neural network and proxy model in the embodiment of the present application is implemented.

[0270] Optionally, the electronic device 20 can be connected to a network through a communication module 23 to communicate with other devices, such as a terminal or a server, through the network to achieve data interaction. The electronic device 20 can be various forms of digital computers, such as desktop computers, servers, workstations, mainframe computers or other types of computers. The electronic device 20 can also be various forms of mobile terminals, such as smart phones, tablet computers, wearable devices (such as helmets, glasses, watches, etc.) and other similar mobile terminals.

[0271] Optionally, the electronic device 20 can be connected to the required input / output devices, such as a keyboard, a display device, etc., through the input / output interface 24. The electronic device 20 itself can have a display device, and can also be connected to other display devices through the input / output interface 24. Optionally, a storage device, such as a hard disk, can also be connected through the input / output interface 24, so that data in the electronic device 20 can be stored in the storage device, or data in the storage device can be read, and data in the storage device can also be stored in the memory 21. It can be understood that the input / output interface 24 can be a wired interface or a wireless interface. Depending on the actual application scenario, the device connected to the input / output interface 24 can be a component of the electronic device 20, or it can be an external device connected to the electronic device 20 when needed.

[0272] Optionally, the memory 21 may be a volatile memory and / or a non-volatile memory, the volatile memory may be a random access memory, etc., and the non-volatile memory may be a read-only memory, a programmable read-only memory, an erasable programmable read-only memory, an electrically erasable programmable read-only memory or a flash memory, etc.

[0273] Optionally, the computer program stored in the processor 22 may be divided into one or more modules, which are stored in the memory 21 and executed by the processor 22 to complete the method provided in the embodiment. The one or more modules may be a series of computer program instruction segments capable of completing specific functions, and the computer program instruction segments are used to describe the execution process of the computer program in the electronic device 20.

[0274] Optionally, the processor 22 can be various general-purpose and / or special-purpose processing components with processing and computing capabilities. Some examples of the processor 22 include, but are not limited to, a central processing unit, a graphics processing unit, a digital signal processor, various dedicated artificial intelligence computing chips, various processors running machine learning model algorithms, and can also be any suitable controller, microcontroller, processor, etc. The processor 22 executes the various methods and processes of this embodiment. Exemplarily, it is a rendering method of skeletal skinning based on a graph neural network and a surrogate model as in the embodiment of the present application.

[0275] Optionally, the bus 25 can include a path for transmitting information. The bus 25 can be a PCI (Peripheral Component Interconnect) bus or an EISA (Extended Industry Standard Architecture) bus, etc. According to different functions, the bus 25 can be divided into an address bus, a data bus, a control bus, etc.

[0276] In an alternative implementation, the embodiment of the present application also provides a computer storage medium, on which a computer program is stored. When the computer program is executed by a computer, the computer can execute the methods of the above method embodiments. Part or all of the computer program can be loaded and / or installed on the memory 21 of the electronic device 20. When the computer program is executed by the processor 22, one or more steps of a rendering method of skeletal skinning based on a graph neural network and a surrogate model as in the embodiment of the present application can be executed.

[0277] Optionally, the computer-readable storage medium can be a random access memory, a read-only memory, a programmable read-only memory, an erasable programmable read-only memory, an electrically erasable programmable read-only memory, etc.

[0278] Obviously, the above embodiments of the present application are merely examples for clearly illustrating the technical solutions of the present application, rather than limitations on the specific implementation manners of the present application. Any modifications, equivalent replacements, and improvements made within the spirit and principles of the claims of the present application shall be included within the protection scope of the claims of the present application.

Claims

1. A rendering method for skeletal skinning based on graph neural networks and surrogate models, characterized in that The rendering method includes: Obtain all adjacent cluster pairs of the skeletal skinning model to be rendered; each of the adjacent cluster pairs contains two adjacent triangular patch clusters; the triangular patch clusters contain several connected triangular patches of the skeletal skinning model to be rendered; Construct a triangular cluster graph based on all the adjacent cluster pairs; the nodes of the triangular cluster graph are triangular patch clusters, and the edges of the triangular cluster graph are the adjacency relationships between adjacent cluster pairs; Obtain the displacement information matrix, deformation intensity, and rendering requirements of each triangular patch cluster at the time point of the animation to be measured; Obtain the cluster dynamic features of each triangular patch cluster at the time point of the animation to be measured according to the displacement information matrix, deformation intensity, rendering requirements, and triangular cluster graph; Process each of the cluster dynamic features using a pre-trained cluster graph neural network to obtain the deformation intensity representation and detail retention value of each triangular patch cluster; the cluster graph neural network introduces local deformation attention and detail precision attention; Predict each triangular patch cluster using a pre-trained multi-branch adaptive proxy model according to the deformation intensity representation, detail retention value, and displacement information matrix to obtain the geometric error increment and envelope change amount of each triangular patch cluster at the time point of the animation to be measured; Obtain the rendering strategy at the time point of the animation to be measured according to the geometric error increment and envelope change amount, and render the skeletal skinning model to be rendered according to the rendering strategy.

2. The rendering method according to claim 1, characterized in that, The obtaining of all adjacent cluster pairs of the skeletal skinning model to be rendered specifically includes: Obtain all triangular patch clusters of the skeletal skinning model to be rendered; Obtain the vertex coordinate set of each of the triangular patch clusters according to the skeletal skinning model to be rendered; Obtain the triangular patch clusters with a common boundary or common vertex according to the vertex coordinate set; Obtain the adjacent cluster pairs according to the triangular patch clusters with a common boundary or common vertex.

3. The rendering method according to claim 1, wherein The constructing of the triangular cluster graph according to all the adjacent cluster pairs specifically includes: Use each of the triangular patch clusters as a node, use the adjacency relationship between the two triangular patch clusters in the adjacent cluster pair as an edge, and set corresponding edge weights for each edge, and construct the triangular cluster graph according to each node and each edge with set edge weights; The calculation of the edge weights includes: Obtain the skeletal joint positions corresponding to each triangular patch cluster of the skeletal skinning model to be rendered; Obtain the joint correlation degree between the two triangular patch clusters in each of the adjacent cluster pairs according to the skeletal joint positions; Obtain the vertex overlap degree between the two triangular patch clusters in each of the adjacent cluster pairs according to the vertex coordinate set; the vertex coordinate set represents the set of vertices of the skeletal skinning model contained in each triangular patch of the corresponding triangular patch cluster; Obtain the dynamic distortion degree between the two triangular patch clusters in each of the adjacent cluster pairs; the dynamic distortion degree represents the skeletal deformation trend between the two triangular patch clusters; Obtain the edge weights between the two triangular patch clusters in each of the adjacent cluster pairs according to the joint correlation degree, vertex overlap degree, and the dynamic distortion degree; the obtaining of the edge weights is expressed as: Wherein, W ij represents the triangular patch cluster C in the adjacent cluster pair i and the adjacent weight between the triangular patch cluster C j , A ij is the joint correlation degree, O ij is the vertex overlap degree, D ij is the dynamic distortion degree; α, β, and γ represent adjustment coefficients.

4. The rendering method according to claim 1, characterized in that Obtaining the cluster dynamic features of each triangular patch cluster at the time point of the animation to be measured according to the displacement information matrix, deformation intensity, rendering requirements, and triangular cluster graph, specifically including: The displacement information matrix includes a rotation matrix and a translation matrix; According to the rotation matrix, translation matrix, deformation intensity, and rendering requirements, obtaining the initial dynamic features of each triangular patch cluster at the time point of the animation to be measured; the obtaining of the initial dynamic features is manifested as: Wherein, t represents the time point of the animation to be measured; X i (t) represents the initial dynamic feature of the triangular patch cluster C at the time point t of the animation to be measured i ; Ri(t), Ti(t), Si(t), and Di(t) respectively represent the rotation matrix, translation matrix, deformation intensity, and rendering requirement of the triangular patch cluster i at the time point t of the animation to be measured; respectively represent the values after normalizing or numerically transforming Ri(t), Ti(t), Si(t), and Di(t); Obtaining the edge weights of each edge in the triangular cluster graph, and obtaining the cluster dynamic features of each triangular patch cluster according to the edge weights and the initial dynamic features; the obtaining of the cluster dynamic features is manifested as: where X i ′ (t) represents the cluster dynamic feature of the triangular facet cluster C i at the animation time point t to be measured; represents the set of triangular facet clusters adjacent to the triangular facet cluster C i ; ω ij represents the triangular facet cluster C in the adjacent cluster pair i and the triangular facet cluster C j ; δ represents the coupling balance coefficient, which is used to control the balance of feature transfer from the corresponding triangular facet cluster to the adjacent triangular facet cluster.

5. The rendering method according to claim 1, characterized in that, The network structure of the cluster graph neural network is constructed according to each node and each edge of the triangular cluster graph; the network structure of the cluster graph neural network includes multiple graph convolution units and an attention mechanism unit; the attention mechanism unit sets the local deformation attention and the detail precision attention; The attention coefficient of the local deformation attention is: where a L (i, j) is the triangular patch cluster C in the adjacent cluster pair i and the local deformation attention attention coefficient of the triangular patch cluster C j ; T ij represents the bone distortion coupling strength between the triangular patch cluster C in the adjacent cluster pair at the animation time point to be measured i and the triangular patch cluster C j ; Δθi represents the angle difference of the bone joint corresponding to the triangular patch cluster C i and the triangular patch cluster C j in the bone distortion direction; τ is the distortion strength determination threshold, and α L , β L are adjustment coefficients respectively; The attention coefficient of the detail precision attention is: Where a D (i) is the triangular patch cluster C i The attention coefficient of the detail precision attention in the aggregation process in the cluster graph neural network, Di(t) represents the rendering requirement of the triangular patch cluster C i at the time point of the animation to be measured, and Vi(t) represents the current visible contour area of the triangular patch cluster C i at the time point of the animation to be measured; η is an offset coefficient, and ζ, Ui(t) are adjustment coefficients.

6. The rendering method according to claim 1, characterized in that, According to the deformation intensity characterization, detail retention value, and displacement information matrix, using the pre-trained multi-branch adaptive proxy model to predict each triangular patch cluster, and obtaining the geometric error increment and the envelope change amount of each triangular patch cluster at the time point of the animation to be measured, specifically including: According to the deformation intensity characterization, detail retention value, and displacement information matrix, constructing an adaptive input vector of the multi-branch adaptive proxy model; Using the pre-trained multi-branch adaptive proxy model, obtaining the geometric error increment and the envelope change amount of each triangular patch cluster at the time point of the animation to be measured according to the adaptive input vector; Among them, the obtaining of the envelope change amount is: where, ΔB i (t) is the envelope change amount of the triangular facet cluster C i at the animation time point t to be measured, I i (t) is the adaptive input vector corresponding to the triangular facet cluster C i at the animation time point t to be measured, B i (t) represents the deformation strength characterization of the triangular facet cluster C i at the animation time point t to be measured; α1 is the power function gain, and β1 is the denominator adjustment coefficient; The obtaining of the geometric error increment is: Where, ΔE i (t) is the geometric error increment of the triangular facet cluster C i at the animation time point t to be measured, and D i (t) represents the rendering requirement of the triangular facet cluster C i at the animation time point t to be measured; α2 is the power function gain, and β2 is the denominator adjustment coefficient.

7. The rendering method according to claim 1, wherein According to the geometric error increment and the envelope change amount, obtaining the rendering strategy at the time point of the animation to be measured, specifically including: Obtaining the pixel contribution degree of each triangular patch cluster at the time point of the animation to be measured; the pixel contribution degree represents the screen coverage ratio of the corresponding triangular patch cluster at the time point of the animation to be measured; According to the geometric error increment and the pixel contribution degree, obtaining the detail level selection index of each triangular patch cluster at the time point of the animation to be measured, expressed as: Where L i (t) is the level - of - detail selection index of the triangle patch cluster C i at the animation time point t to be measured, ΔE i (t) is the geometric error increment of the triangle patch cluster C i at the animation time point t to be measured, P i (t) is the pixel contribution degree of the triangle patch cluster C i at the animation time point t to be measured; is the exponential adjustment coefficient, and μ is the pixel contribution threshold; According to the envelope change amount and the pixel contribution degree, obtaining the visibility metric of each triangular patch cluster in the occlusion culling process during rendering at the time point of the animation to be measured, expressed as: where Θ i (t) is the visibility measure of the triangular patch cluster C i at the animation time point t to be measured, and ΔB i (t) is the change amount of the bounding volume of the triangular patch cluster C i at the animation time point t to be measured, and σ is the power decay coefficient; Respectively according to the detail level selection index and the visibility metric, formulating the detail level rendering strategy and the visibility rendering strategy for the corresponding triangular patch cluster; The detail level rendering strategy is: When the detail level selection index is greater than or equal to the preset tolerance threshold, increasing the rendering detail level of the corresponding triangular patch cluster; when the detail level selection index is less than the preset tolerance threshold, keeping or reducing the rendering detail level of the corresponding triangular patch cluster; The visibility rendering strategy is: When the visibility metric is less than or equal to the visibility threshold, it is determined that the corresponding triangular patch cluster is culled at the time point of the animation to be measured; when the visibility independence is greater than the visibility threshold, a new envelope is obtained based on the corresponding envelope change amount and output.

8. The rendering method according to any one of claims 1-7, characterized in that, Before obtaining the rendering strategy at the time point of the animation to be measured according to the geometric error increment and the envelope change amount, it further includes: Correcting the attention coefficients of the local deformation attention and the detail precision attention according to the geometric error increment and the envelope change amount, specifically including: Obtaining the feedback input metric of each triangular patch cluster at the time point of the animation to be measured according to the geometric error increment and the envelope change amount, which is expressed as: Where, M i (t) is the feedback input metric of the triangular patch cluster C i at the animation time point t to be measured, ΔB i (t) is the change amount of the envelope of the triangular patch cluster C i at the animation time point t to be measured, ΔE i (t) is the geometric error increment of the triangular patch cluster Ci at the animation time point t to be measured; γ1 is the weight coefficient; Correcting the attention coefficients of the local deformation attention and the detail precision attention according to the feedback input metric, which is expressed as: where, aL(i,j) and a Lr (i,j) are the attention coefficients of the local deformation attention before and after correction of the triangular patch cluster C i in the adjacent cluster pair and the triangular patch cluster C j respectively; a D (i) and a Dr (i) are the attention coefficients of the detail accuracy attention before and after correction of the triangular patch cluster C i respectively; κ1 and κ2 are used as the update amplitude adjustment, and ρ is the power decay coefficient; Feeding back the corrected attention coefficients of the local deformation attention and the detail precision attention to the cluster graph neural network.

9. The rendering method according to any one of claims 1-7, characterized in that Before obtaining the displacement information matrix, deformation intensity and rendering requirements of each triangular patch cluster at the time point of the animation to be measured, it further includes: Periodically obtaining the character pose change and character rendering requirements of the skeleton skinned model to be rendered; Adjusting the frequencies of obtaining the character pose change and character rendering requirements according to the character pose change and character rendering requirements, and determining whether to execute the steps of obtaining at the time point of the animation to be measured and subsequent steps according to the adjusted frequencies; specifically including: Obtaining the character pose vectors of the skeleton skinned model to be rendered at the current time point and the previous time point respectively according to the period, and obtaining the character pose change at the current time point according to the character pose vectors at the current time point and the previous time point; Obtaining the character rendering requirements of the skeleton skinned model to be rendered at the current time point; Constructing a monitoring function according to the character pose change and character rendering requirements, which is expressed as: where ω(t) represents the monitoring function, Δθ r (t) is the change in the character's pose, and d(t) is the character rendering requirement; η1 and η2 are monitoring coefficients, and ∈1 and ∈2 are monitoring thresholds; Obtaining the target adjustment frequency according to the monitoring function, which is expressed as: where f u (t) is the target adjustment frequency, f max is the preset maximum adjustment frequency, and f min is the preset minimum adjustment frequency; When the target adjustment frequency is greater than or equal to the preset frequency threshold, execute the steps of obtaining at the time point of the animation to be measured and subsequent steps; when the target adjustment frequency is less than the preset frequency threshold, maintain the rendering strategy of the previous time point.

10. A rendering system for skeletal skinning based on graph neural networks and surrogate models, characterized in that, The rendering system includes: A cluster pair acquisition module for acquiring all adjacent cluster pairs of the skeleton skinned model to be rendered; each of the adjacent cluster pairs includes two adjacent triangular patch clusters; the triangular patch clusters include several connected triangular patches of the skeleton skinned model to be rendered; A cluster graph construction module for constructing a triangular cluster graph according to all the adjacent cluster pairs; the nodes of the triangular cluster graph are triangular patch clusters, and the edges of the triangular cluster graph are the adjacency relationships between adjacent cluster pairs; A cluster information acquisition module for acquiring the displacement information matrix, deformation intensity and rendering requirements of each triangular patch cluster at the time point of the animation to be measured; A dynamic feature acquisition module for acquiring the cluster dynamic features of each triangular patch cluster at the time point of the animation to be measured according to the displacement information matrix, deformation intensity, rendering requirements and triangular cluster graph; A neural network processing module for processing each of the cluster dynamic features using a pre-trained cluster graph neural network to obtain the deformation intensity representation and detail retention value of each triangular patch cluster; the cluster graph neural network introduces local deformation attention and detail precision attention, and the cluster graph neural network extracts the deformation intensity representation and detail retention value of the cluster dynamic features through local deformation attention and detail precision attention; A pose prediction module for predicting each triangular patch cluster using a pre-trained multi-branch adaptive proxy model according to the deformation intensity representation, detail retention value, and displacement information matrix to obtain the geometric error increment and envelope change amount of each triangular patch cluster at the time point of the animation to be measured; A pose rendering module, according to the geometric error increment and envelope change amount, obtains a rendering strategy at the time point of the animation to be measured, and renders the bone skin model to be rendered according to the rendering strategy.

Images