Self-supervised clothes simulation method based on multilayer message passing and geometric collision of graph neural network

Abstract

The invention discloses a self-supervised clothes simulation method based on multilayer message passing and geometric collision of a graph neural network. The method comprises the following steps: making a human body action sequence data set; by learning the graph neural network, reasonable dynamic motion of various clothes types and versions can be predicted, the local dynamic graph neural network is learned in a mode independent of the clothes types, a layered message passing mechanism for efficiently capturing long-distance information is achieved, and a loss function is defined by utilizing physical constraint. Collision post-processing in a reasoning stage is avoided based on a geometric collision strategy, and a realistic cloth simulation effect is obtained. The system can assist in game 3D cloth simulation, virtual reality digital twinborn character construction and the like.

Description

Technical Field

[0001] This invention relates to the field of computer graphics, and in particular to a self-supervised clothing simulation method based on multi-layer information transmission and geometric collision of graph neural networks. Background Technology

[0002] This section provides only background information relevant to this disclosure and is not necessarily prior art.

[0003] With the rapid development of virtual reality technology, clothing simulation has become a very popular research direction in the many application areas of computer graphics. For example, virtual try-on allows users to experience the effect of clothing on their body in real time without going to a physical store, and to make personalized choices based on their personal preferences. However, how to perform real-time physical simulation of fabric remains an unsolved problem. Traditional physical simulation methods based on solving implicit Euler equations have very high computational costs and cannot meet real-time requirements. Although PBD / XPBD solvers commonly used in industry have fast solution speeds, they are not physically realistic. The most advanced data-driven 3D clothing deformation modeling methods currently use supervised training strategies. These methods require large datasets, usually acquired through professional multi-camera capture systems. Although they can capture clothing deformation effects with high accuracy, they are limited by expensive capture equipment and reliance on complex capture systems with strict environmental constraints, making them difficult for non-professionals to use. Previous unsupervised learning methods required the use of linear skinning to reduce spatial complexity, resulting in fabric movement that clings to the body, leading to poor dynamic effects and greatly limiting the application scope of virtual try-on. Summary of the Invention

[0004] To address the shortcomings of existing technologies, this invention aims to provide a self-supervised clothing simulation method based on graph neural networks, multi-layer message passing, and geometric collision. Utilizing graph neural networks, hierarchical message passing, and a physically defined loss function, the resulting clothing simulation effect is smoother and better meets visual requirements. This solves the problems of high computational complexity and difficulty in achieving real-time performance in traditional methods, high equipment requirements and labor costs in data-driven methods, and poor dynamic effects in skinning-based solutions.

[0005] To achieve the above objectives, the present invention provides the following solution:

[0006] A self-supervised clothing simulation method based on multi-layer message passing and geometric collision in graph neural networks includes:

[0007] Step 1: Select the axis-angle rotation data of key human skeletal points and the human body shape parameters as input to the model;

[0008] Step 2: Construct a graph model of the clothing mesh, and enhance it by adding connection edges between the clothing and the nearest human node;

[0009] Step 3: Convert the graph model obtained in Step 2 into a hierarchical graph structure to enable rapid information propagation, and process it with a hierarchical message-passing graph neural network to update the features of nodes and edges;

[0010] Step 4: Based on the clothing node features obtained after information transmission in Step 3, input them into the decoder multilayer perceptron to predict the acceleration value of each clothing vertex. Then, according to the time step, the velocity is updated using numerical integration based on the fundamental laws of kinematics, and the displacement is updated using the velocity. Finally, the position of the clothing vertex is calculated. The time step can be 0.0333s.

[0011] Step 5: Construct a loss function based on physical and geometric perception, and construct a geometric collision constraint loss term to solve the clipping problem that occurs in previous solutions;

[0012] Step 6: For the loss function defined in Step 5, iteratively optimize the neural network parameters using the Adam optimizer. During training, expand the feature vector of the graph neural network using dynamic random material parameters until the model reaches convergence.

[0013] In some embodiments, step 1 includes the following steps:

[0014] Step 1-1: Based on the type of clothing to be simulated, filter out the key human skeletal points that affect the deformation of the clothing and exclude skeletal points that have no effect; for example, a top will be affected by bones such as the pelvis, chest, neck, and arms, but will not be affected by bones such as the knees and ankles.

[0015] Steps 1-2: Collect motion data of key human skeletal points after filtering, and record the local rotation information θ of each skeletal point in each frame using axis-angle rotation. Human motion is specifically realized as the local rotation of human skeletal points. Therefore, this model learns the mapping relationship from skeletal point rotation to clothing deformation. The rotation from skeletal points is expressed using axis-angle rotation, which is more intuitive and has fewer parameters compared to rotation matrices, Euler angles, and other forms of expression.

[0016] Steps 1-3: Determine the body shape parameters that characterize the height, weight, and build of the human body, and set the body shape parameter β to cover different body shape characteristics;

[0017] Steps 1-4: Integrate the axis-angle rotation data obtained in Step 1-2 with the human body shape parameters determined in Step 1-3 to form the input data set for the subsequent model.

[0018] In some embodiments, step 2 includes the following steps:

[0019] Step 2-1: Normalize the clothing mesh and the corresponding human body mesh to be simulated using AABB bounding boxes respectively;

[0020] Step 2-2: Construct a graph model using the vertices and edges of the clothing mesh, where each vertex and edge corresponds to a feature vector of the graph network. and Here, i and j represent vertex indices. The vertex feature vector contains the type (clothing / human body), current state variables (velocity, normal vector), and physical properties (mass). The edge feature vector records the relative positional relationship between its two endpoints in the current state.

[0021] Steps 2-3: Find the nearest vertex on the human body model for each clothing vertex. If the distance is less than the threshold ε, connect the two vertices and add a new edge. This indicates that the apex of the human body will affect the movement of the clothing.

[0022] In some embodiments, step 3 includes the following steps:

[0023] Step 3-1: Coarsen the input graph using a recursive process. First, divide the nodes of the input graph into progressively coarser sets, so that the distance between nodes (i.e. the number of edges in the shortest path between two nodes) increases accordingly. Then, generate a new set of coarsened edges for each partition, which is divided into three layers: L0, L1, and L2. L0 represents the finest layer, and L2 represents the coarsest layer.

[0024] Step 3-2: The nested hierarchical graph structure used supports simultaneous information processing at multiple levels, with each level L of the graph equipped with an independent set of edge feature vectors. And node features It is then shared across all levels.

[0025] Step 3-3: At the beginning of the message passing step, first use node features to update the edge features of all levels. Then, the node features are updated using edge features. f represents the neural network mapping

[0026] Based on graph neural networks and a hierarchical message passing mechanism, the system learns the dynamic deformation of clothing through spatial topology modeling. This method utilizes a hierarchical graph structure to synchronously process multi-scale vertex interactions, effectively capturing long-range mechanical coupling relationships from local folds to global drapes through iterative message passing between layer levels. To overcome the generalization limitations of traditional temporal models, randomly initialized node feature vectors are used, enabling the model to have strong generalization capabilities for unseen clothing styles, material parameters, and topological structures. The spatial topology-based neural network establishes an end-to-end mapping from human motion and shape parameters to fabric dynamic deformation, stably handling long-sequence simulations without relying on temporal recursive structures.

[0027] In some embodiments, step 4 includes the following steps:

[0028] Step 4-1: Pass the node features after message transmission to the decoder multilayer perceptron to predict the acceleration of each clothing vertex. Based on the set time step, use the laws of kinematics and the obtained acceleration to calculate the displacement of the clothing vertex.

[0029] Step 4-2: Based on the actual spatial position of the vertices of the clothing after deformation, update the normal information of the clothing, map the clothing model that has completed deformation calculation and normal update back to its original size, and spatially align it with the corresponding human body model.

[0030] In some embodiments, step 5 includes the following steps:

[0031] Step 5-1: Construct a loss function L based on physical constraints, whereby the physical constraints include dynamic loss. Gravity constraint loss Internal force constraint loss and collision damage ;

[0032] Step 5-2: Construct the geometric collision loss term To address the clipping problem that may occur during the reasoning phase;

[0033] Step 5-3: Add the physical constraint loss term from Step 5-1 to the geometric collision loss term from Step 5-2 to form the complete loss function. The goal is to avoid collision post-processing during the inference phase by using the neural network as an implicit integrator to predict the cloth model in each frame at its lowest energy point.

[0034] In some embodiments, step 5-1 defines the physical constraints as follows:

[0035] The dynamics definition defines the loss as inertial energy. Where x is the motion state of the clothing vertex in the current frame, It represents the motion state of the previous frame, and M is the mass matrix of the clothing. The purpose is to prevent the velocity of the clothing particles from changing continuously over time.

[0036] The external constraint loss is defined as gravitational potential energy. , where m is the mass of each garment vertex and g is the gravitational acceleration, simulating the effect of gravity on the movement of the garment;

[0037] The internal force constraint loss includes strain energy and bending energy, i.e. The strain energy is obtained through the deformation gradient. Let the initial position of the clothing be X, and its position after deformation be x. Then the deformation gradient... The strain energy was then calculated. Where λ and μ are Lame constants, and G is the strain tensor. I is the identity matrix, tr(·) denotes the trace of the matrix, and finally the strain energy is calculated. V represents the volume (area * thickness) of each facet; bending energy is defined by the difference between the local smoothness of the clothing and the consistency of the original clothing position, expressed as... ,in θt represents the dihedral angle of the deformed garment, θr represents the corresponding dihedral angle of the template garment in its static state, l is the length of the common side, and a is the sum of the areas of the two triangles. The balance coefficient α∈[0,1] is proportional to the distance from the current side to the original position of the garment;

[0038] The collision loss is defined as the nearest directed distance d(x) from the clothing vertex to the human safety boundary ε. Vertices smaller than the safety boundary are given collision energy, the magnitude of which is quantified by the distance value, and expressed as: ,in The collision coefficient, .

[0039] In some embodiments, step 5-2 includes the following steps:

[0040] Step 5-2-1: Calculate the covariance between the vertex positions of the clothing and the vertex positions after deformation. The covariance describes the distribution characteristics of the vertices in 3D space. The covariance matrix is ​​used for this purpose. Where x represents the position of the vertex of the clothing. The mean of the positions is N, where N is the total number of vertices.

[0041] Step 5-2-2: Assuming the clothing has limited space, then ,in The singular values ​​of the initial position covariance matrix are given by `det(·)`, which represents the determinant calculation. A balance factor `k` is set to an initial value of 1; `k=1` indicates non-stretchable, and `k>1` indicates stretchable.

[0042] Step 5-2-3: Identify the areas where clothing vertices may penetrate the human body geometry, and adjust the balance factor k for these areas. ,in The directed depth of penetration is represented by k. By dynamically adjusting k, clothing can be stretched to avoid collisions, enhancing the network's geometric perception ability in this area. The geometric collision loss function is expressed as... A geometric collision loss term is added to constrain the covariance of the clothing vertex position and the position after deformation, thereby avoiding the clipping phenomenon that occurs during the inference stage.

[0043] The proposed method utilizes a fused graph neural network, multi-layer message passing, and unsupervised training to predict realistic clothing dynamics for any garment style and body type in real time. Compared to previous unsupervised fabric simulation schemes, which first use a network to predict the deformation of clothing in the original space and then attach it to the human body via a linear skin, resulting in clothing clinging tightly to the body during simulation—a highly unsatisfactory effect for loose garments like dresses—this graph neural network-based and multi-layer message passing approach avoids the distortions associated with such methods. It also generalizes to unseen clothing and supports dynamic adjustments to material parameters and clothing topology. Attached Figure Description

[0044] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments, and the advantages of the present invention in the above and / or other aspects will become clearer.

[0045] Figure 1 This is a flowchart of a self-supervised clothing simulation method based on multi-layer message passing and geometric collision of graph neural networks provided by the present invention.

[0046] Figure 2 The flowchart for constructing the geometric collision constraint loss term provided by this invention is shown.

[0047] Figure 3 This invention provides a flowchart for predicting the vertex state of clothing during the inference phase of the network.

[0048] Figure 4 The diagram shows a comparison of the proposed solution and the skin-based self-supervised neural network method. Detailed Implementation

[0049] The purpose of this invention is to provide a self-supervised clothing simulation method based on graph neural networks, multi-layer message passing, and geometric collision. By utilizing graph neural networks, hierarchical message passing, and a physically defined loss function, the resulting clothing simulation effect is smoother and better meets visual requirements. This solves the problems of high computational complexity and difficulty in achieving real-time performance in traditional methods, high equipment requirements and labor costs for data-driven methods, and poor dynamic effects in skinning-based solutions.

[0050] like Figure 1-3 As shown, this solution includes the following steps:

[0051] Step 1: Select the axis-angle rotation data of key human skeletal points and the human body shape parameters as input to the model;

[0052] Specifically, based on the type of clothing to be simulated, key human skeletal points that affect the deformation of the clothing are selected, and ineffective skeletal points are excluded. Then, motion data of the selected key human skeletal points are collected, and the local rotation information θ of each skeletal point is recorded in an axis-angle rotation manner. Next, the shape parameters that characterize the height, weight, and build of the human body are determined, and the human body shape parameter β is set to cover different human body shape features. Finally, the obtained axis-angle rotation data and human body shape parameters are integrated, and different types of action sequences are selected to form the input data set for the subsequent model.

[0053] Step 2: Construct a graph model of the clothing mesh, and enhance it by adding connection edges between the clothing and the nearest human node;

[0054] Step 2-1: Normalize the clothing mesh and the corresponding human body mesh to be simulated using AABB bounding boxes respectively;

[0055] Specifically, the axial bounding box (AABB) of the mesh vertices is calculated, and its minimum and maximum vertex coordinates are determined. Then, the center of the AABB is moved to the origin of the coordinate system through a translation transformation, and then scaled proportionally based on the maximum span (length, width, height) of each axis, so that the entire bounding box is normalized to a standard unit space (such as [-1,1]³).

[0056] Step 2-2: Construct a graph model using the vertices and edges of the clothing mesh, where each vertex and edge corresponds to a feature vector of the graph network. and Here, i and j represent vertex indices. The vertex feature vector contains the type (clothing / human body), current state variables (velocity, normal vector), and physical properties (mass). The edge feature vector records the relative positional relationship between its two endpoints in the current state.

[0057] Steps 2-3: Find the nearest vertex on the human body model for each clothing vertex. If the distance is less than the threshold ε, connect the two vertices and add a new edge. This indicates that the apex of the human body will affect the movement of the clothing.

[0058] Step 3: Convert the graph model obtained in Step 2 into a hierarchical graph structure to enable rapid information propagation, and process it with a hierarchical message-passing graph neural network to update the features of nodes and edges.

[0059] Step 3-1: Coarsen the input graph using a recursive process. First, divide the nodes of the input graph into progressively coarser sets, so that the distance between nodes (i.e. the number of edges in the shortest path between two nodes) increases accordingly. Then, generate a new set of coarsened edges for each partition, which is divided into three layers: L0, L1, and L2. L0 represents the finest layer, and L2 represents the coarsest layer.

[0060] Specifically, the process is as follows: First, starting from the central node of the graph structure, find the adjacent nodes of all odd-numbered layers and connect the adjacent nodes to create new edges; then, process the boundary nodes, repair nodes that are only adjacent on one side, and ensure that reasonable connections can be formed in the boundary areas; finally, randomly supplement the even-numbered layers of the central node to prevent the coarsened graph from becoming too sparse and to ensure the connectivity of the graph.

[0061] Step 3-2: The nested hierarchical graph structure used supports simultaneous information processing at multiple levels, with each level L of the graph equipped with an independent set of edge feature vectors. And node features It is then shared across all levels.

[0062] Step 3-3: At the beginning of the message passing step, first use node features to update the edge features of all levels. Then, the node features are updated using edge features. , where f represents the neural network mapping.

[0063] Specifically, in each message passing step, the human body edge set will be dynamically updated. Only edges connected to the currently processed clothing node are retained. The multi-level message passing scheme can support any number of levels of synchronous processing. Experiments show that using a three-level structure (fine layer - intermediate layer - coarsest layer - intermediate layer - fine layer), synchronously processing two adjacent layers at a time achieves the best balance between inference speed and result quality.

[0064] Step 4: Based on the clothing node features obtained after information transmission in Step 3, input them into the decoder multilayer perceptron to predict the acceleration value of each clothing vertex, and then calculate the final clothing position according to the time step (set to 0.0333s in this method) and the basic laws of kinematics.

[0065] Step 4-1: Pass the node features after message transmission to the decoder multilayer perceptron to predict the acceleration of each clothing vertex. Based on the set time step, use the laws of kinematics and the obtained acceleration to calculate the displacement of the clothing vertex.

[0066] Step 4-2: Based on the actual spatial position of the vertices of the clothing after deformation, update the normal information of the clothing, map the clothing model that has completed deformation calculation and normal update back to its original size, and spatially align it with the corresponding human body model.

[0067] Step 5: Construct a loss function based on physical and geometric collisions, and construct a geometric collision constraint loss term to solve the clipping problem that occurs in previous solutions.

[0068] Step 5-1: Construct a loss function L based on physical constraints, whereby the physical constraints include dynamic loss. Gravity constraint loss Internal force constraint loss and collision damage ;

[0069] Step 5-2: Construct the geometry-aware loss term To address the clipping problem that may occur during the reasoning phase;

[0070] Specifically, the process involves: first, calculating the covariance between the vertex positions of the clothing and the vertex positions after deformation; then, using the covariance to describe the distribution characteristics of the vertices in 3D space, resulting in the covariance matrix. Where x represents the position of the vertex of the clothing. Let N represent the average value of the positions, and N be the total number of vertices; then, assuming that the local space of the clothing is not extensible, then... ,in The singular values ​​of the initial position covariance matrix are represented by `det(·)`, which denotes the determinant calculation. A balance factor `k` is set to an initial value of 1, where `k=1` indicates non-stretchable and `k>1` indicates stretchable. Finally, the regions where clothing vertices may penetrate the human geometry are identified, and the balance factor `k` for these regions is adjusted. ,in The directed depth of penetration is represented by k. By dynamically adjusting k, clothing can be stretched to avoid collisions, enhancing the network's geometric perception ability in this area. The geometric collision loss function is expressed as... .

[0071] Step 5-3: Add the physical constraint loss term from Step 5-1 to the geometric collision loss term from Step 5-2 to form the complete loss function. .

[0072] Specifically: the dynamic definition of loss is inertial energy. Where x is the motion state of the clothing vertex in the current frame, This represents the motion state of the previous frame. M is the mass matrix of the clothing, and its purpose is to prevent the velocity of clothing particles from changing continuously over time. Gradient backpropagation of this loss term is prohibited during training. The external force constraint loss is defined as gravitational potential energy. Where m is the mass of each garment vertex, and g is the gravitational acceleration, simulating the effect of gravity on the movement of the garment; the internal constraint loss includes strain energy and bending energy, i.e. The strain energy is obtained through the deformation gradient. Let the initial position of the clothing be X, and its position after deformation be x. Then the deformation gradient... The strain energy was then calculated. Where λ and μ are Lame constants, and G is the strain tensor. I is the identity matrix, tr(·) denotes the trace of the matrix, and finally the strain energy is calculated. V represents the volume (area * thickness) of each facet; bending energy is defined by the difference between the local smoothness of the clothing and the consistency of the original clothing position, expressed as... ,in Let θt be the bending coefficient, θr be the dihedral angle of the deformed garment, l be the length of the common side, and a be the sum of the areas of the two triangles. The balance coefficient α∈[0,1] is proportional to the distance from the current edge to the original position of the garment. Collision loss is defined as the nearest directed distance d(x) from the garment vertex to the human safety boundary ε. Collision energy is added to vertices smaller than the safety boundary, and the energy magnitude is quantified by the distance value, expressed as: ,in The collision coefficient, In this method, ε is set to 0.003.

[0073] Step 6: For the defined loss function, use the Adam optimizer to iteratively optimize the neural network parameters. During the training process, expand the feature vector of the graph neural network with dynamic random material parameters until the model reaches convergence.

[0074] Specifically, to calculate the mass matrix M of the garment, measured parameters of pure cotton fabric were used: thickness 0.47 mm, density 426 kg / m³. Dynamic random material parameters include the lame constants λ and μ; the initial values ​​for this method are set to λ = 20.9 and μ = 11.1; the curvature coefficient... The initial value for this method is set to 3.96e-5. Regarding network hyperparameter settings, a batch size of 16 is used for training: initially, 10 epochs of training are performed with a learning rate of 0.001, followed by adjusting the learning rate to 0.0001 and continuing training until the model converges. This method offers fast training speed, is suitable for modeling various types of clothing, and effectively avoids error states.

[0075] like Figure 4 As shown, the comparison of the effects of this scheme (left) and the skin-based (right) self-supervised neural network method shows that the simulation effect of this scheme on loose fabric is significantly better than that of the skin-based scheme.

[0076] This invention provides a concept and method for a self-supervised clothing simulation method based on multi-layer message passing and geometric collision in graph neural networks. Many methods and approaches exist for implementing this technical solution; the above description is merely a preferred embodiment of the invention. It should be noted that those skilled in the art can make various improvements and modifications without departing from the principles of this invention, and these improvements and modifications should also be considered within the scope of protection of this invention. All components not explicitly stated in this embodiment can be implemented using existing technologies.

Claims

1. A self-supervised clothing simulation method based on multi-layer message passing and geometric collision in graph neural networks, characterized in that, include: Step 1: Select the axis-angle rotation data of key human skeletal points and the human body shape parameters as input to the model; Step 2: Construct a graph model of the clothing mesh, and enhance it by adding connection edges between the clothing and the nearest human node; Step 3: Convert the graph model obtained in Step 2 into a hierarchical graph structure to enable rapid information propagation, and process it using a hierarchical message-passing graph neural network to update the features of nodes and edges; Step 4: Based on the clothing node features obtained after information transmission in Step 3, input them into the decoder multilayer perceptron to predict the acceleration value of each clothing vertex, and calculate the final clothing position using the basic laws of kinematics. Step 5: Construct a loss function based on physical and geometric perception, and construct a geometric collision constraint loss term to solve the clipping problem that occurs in previous solutions; Step 6: For the loss function defined in Step 5, iteratively optimize the neural network parameters using the Adam optimizer. During training, expand the feature vector of the graph neural network using dynamic random material parameters until the model reaches convergence.

2. The self-supervised clothing simulation method based on multi-layer message passing and geometric collision of graph neural networks according to claim 1, characterized in that, Step 1 includes the following steps: Step 1-1: Based on the type of clothing to be simulated, filter out the key human skeletal points that affect the deformation of the clothing and exclude skeletal points that have no effect; Step 1-2: Collect motion data of key human skeletal points after screening, and record the local rotation information θ of each skeletal point in each frame using axis-angle rotation. Steps 1-3: Determine the body shape parameters that characterize the height, weight, and build of the human body, and set the body shape parameter β to cover different body shape characteristics; Steps 1-4: Integrate the axis-angle rotation data obtained in Step 1-2 with the human body shape parameters determined in Step 1-3 to form the input data set for the subsequent model.

3. The self-supervised clothing simulation method based on multi-layer message passing and geometric collision of graph neural networks according to claim 1, characterized in that... Step 2 includes the following steps: Step 2-1: Normalize the clothing mesh to be simulated and the corresponding human body mesh using AABB bounding boxes respectively; Step 2-2: Construct a graph model using the vertices and edges of the clothing mesh, where each vertex and edge corresponds to a feature vector of the graph network. and , where i and j represent vertex indices; the vertex feature vector contains type, current state variables, and physical properties; the edge feature vector records the relative positional relationship between its two endpoints in the current state; Steps 2-3: Find the nearest vertex on the human body model for each clothing vertex. If the distance is less than the threshold ε, connect the two vertices and add a new edge. This indicates that the apex of the human body will affect the movement of the clothing.

4. The self-supervised clothing simulation method based on multi-layer message passing and geometric collision of graph neural networks according to claim 1, characterized in that, Step 3 includes the following steps: Step 3-1: Coarsen the input graph using a recursive process. Divide the nodes of the input graph into progressively coarser sets, so that the distance between nodes increases accordingly. The distance between nodes is the number of edges in the shortest path between two nodes. Generate a new set of coarsened edges for each partition, which is divided into three layers: L0, L1, and L2. L0 represents the finest layer, and L2 represents the coarsest layer. Step 3-2: Process information synchronously across multiple levels using a nested hierarchical graph structure, equipping each level L of the graph with an independent set of edge feature vectors. And node features It is then shared across all levels; Step 3-3: At the beginning of the message passing step, update the edge features of all levels using node features. Update node features using edge features , where f represents the neural network mapping.

5. The self-supervised clothing simulation method based on multi-layer message passing and geometric collision of graph neural networks according to claim 1, characterized in that, Step 4 includes the following steps: Step 4-1: Pass the node features after message transmission to the decoder multilayer perceptron to predict the acceleration of each clothing vertex. Based on the set time step, calculate the displacement of the clothing vertex using kinematic laws and the obtained acceleration. Add the displacement of the clothing vertex to the position of the clothing vertex in the previous frame to obtain the position of the deformed clothing vertex. Step 4-2: Based on the actual spatial position of the vertices of the clothing after deformation, update the normal information of the clothing, map the clothing model that has completed deformation calculation and normal update back to its original size, and spatially align it with the corresponding human body model.

6. The self-supervised clothing simulation method based on multi-layer message passing and geometric collision of graph neural networks according to claim 1, characterized in that, Step 5 includes the following steps: Step 5-1: Construct a loss function L based on physical constraints, whereby the physical constraints include dynamic loss. Gravity confinement loss Internal force constraint loss and collision damage ; Step 5-2: Construct the geometric collision loss term To address the clipping problem that may occur during the reasoning phase; Step 5-3: Add the physical constraint loss term from Step 5-1 to the geometric collision loss term from Step 5-2 to form the complete loss function. .

7. The self-supervised clothing simulation method based on multi-layer message passing and geometric collision of graph neural networks according to claim 6, characterized in that, Step 5-1 defines the physical constraints as follows: The dynamic loss is inertial energy. Where x is the motion state of the clothing vertex in the current frame, It represents the motion state of the previous frame, and M is the mass matrix of the clothing. External constraint loss is defined as gravitational potential energy. , where m is the mass of each garment vertex and g is the gravitational acceleration; Internal constraint losses include strain energy and bending energy, i.e. The strain energy is obtained through the deformation gradient. Let the initial position of the clothing be X, and its position after deformation be x. Then the deformation gradient... To obtain strain energy Where λ and μ are Lame constants, and G is the strain tensor. I is the identity matrix, tr(·) denotes the trace of the matrix, and finally the strain energy is calculated. V represents the volume of each facet, i.e., area * thickness; bending energy is defined by the difference between the local smoothness of the clothing and the consistency of the original clothing position, expressed as... ,in θt represents the dihedral angle of the deformed garment, θr represents the corresponding dihedral angle of the template garment in its static state, l is the length of the common side, and a is the sum of the areas of the two triangles; the balance coefficient α∈[0,1] is proportional to the distance from the current side to the original position of the garment. The collision loss is defined as the nearest directed distance d(x) from the clothing vertex to the human safety boundary ε. Vertices smaller than the safety boundary are given collision energy, the magnitude of which is quantified by the distance value, and expressed as: ,in The collision coefficient, .

8. The self-supervised clothing simulation method based on multi-layer message passing and geometric collision of graph neural networks according to claim 5, characterized in that, Step 5-2 includes the following steps: Step 5-2-1: Calculate the covariance between the vertex positions of the clothing and the vertex positions after deformation. The covariance describes the distribution characteristics of the vertices in 3D space. The covariance matrix is ​​used for this purpose. Where x represents the position of the vertex of the clothing. The mean of the positions is N, where N is the total number of vertices. Step 5-2-2: Assuming the clothing has limited space, then ,in The singular values ​​of the initial position covariance matrix are given by det(·), which represents the determinant calculation. The balance factor k is set to an initial value of 1, where k=1 indicates non-stretchable and k>1 indicates stretchable. Step 5-2-3: Identify the areas where clothing vertices may penetrate the human body geometry, and adjust the balance factor k for these areas. ,in The directed depth of penetration is represented by k. By dynamically adjusting k, clothing can be stretched to avoid collisions, enhancing the network's geometric perception ability in this area. The geometric collision loss function is expressed as... .

Images