Three-dimensional garment deformation prediction method and device fusing fit attributes

Abstract

The application discloses a three-dimensional garment deformation prediction method and device fusing fit attributes, which can realize dressing effects of various garment meshes in any posture, directly generate complex garment animations with fine details and realistic effects, and is superior to existing garment animation methods in generalization performance and deformation quality, and has significant application value. The method comprises the following steps: (1) creating a data set composed of garments, body shapes and animation postures for training and testing; (2) generating suitable parameters as one of network inputs by observing parameters affecting deformation quality, so as to generate real deformation for different suitability of garments; (3) performing output reconstruction by decomposing the output vector of each vertex into amplitude and direction; (4) coarse deformation prediction; and (5) detailed deformation of fine-scale wrinkles.

Description

Technical Field

[0001] This invention relates to the technical fields of computer graphics and deep learning, and in particular to a method for predicting the deformation of three-dimensional clothing by incorporating fit attributes, and a device for predicting the deformation of three-dimensional clothing by incorporating fit attributes. Background Technology

[0002] Clothing animation is a fundamental topic in computer graphics, aiming to create realistic clothing transformation effects for numerous applications such as virtual try-on, video games, and movies. With advancements in graphics, users are increasingly focused on the visual effects of clothing, including how it interacts more realistically with the human body and how wrinkles are added or removed under different movements. High-quality clothing transformations not only provide convenience for users shopping online but also offer immersive experiences for a wide range of digital applications.

[0003] The primary method for creating high-quality clothing animations is physically based simulation. This approach provides relatively accurate simulation results, but its deployment in real-time applications remains challenging due to the high cost of computer simulation. In recent years, deep learning-based clothing deformation methods have become mainstream, utilizing neural network models to map various features and directly predict clothing states. While these methods can roughly approximate clothing behavior, issues remain regarding model generalization performance and the detail quality of the results.

[0004] Most deep learning-based deformation methods employ simple multilayer perceptron (MLP) models to directly predict the nonlinear deformation of clothing. While the predicted results contain some detail of wrinkles, the fixed-size input and output vectors necessitate that the training and testing targets have the same number of vertices and topology. This constant mesh topology can only handle clothing with small-scale size variations, failing to meet the diverse deformation needs of individually customized clothing. Furthermore, current learning-based methods are generally plagued by the highly nonlinear output (e.g., vertex positions, displacements), where each element in the output vector ranges from negative infinity to positive infinity, making prediction difficult. This leads some studies to use numerous fully connected layers at the expense of model generalization performance, or to predict deformation only for specific poses to ensure quality. Summary of the Invention

[0005] To overcome the shortcomings of existing technologies, the technical problem to be solved by this invention is to provide a three-dimensional clothing deformation prediction method that integrates fit attributes. This method can realize the dressing effect of various clothing meshes in any posture and can directly generate complex clothing animations with fine details and realistic effects. It is superior to existing clothing animation methods in terms of generalization performance and deformation quality, and takes into account both deformation effect and efficiency, thus having significant application value.

[0006] The technical solution of this invention is: a three-dimensional clothing deformation prediction method that integrates fit attributes, comprising the following steps:

[0007] (1) Create a dataset consisting of clothing, body shape and animated poses for training and testing;

[0008] (2) By observing the parameters that affect the deformation quality, suitable parameters are generated as one of the network inputs to generate realistic deformation for different fits of clothing;

[0009] (3) Output reconstruction is performed by decomposing the output vector of each vertex into magnitude and direction;

[0010] (4) Coarse-grained prediction: First, a block is applied to process the input image in the first layer. Then, three blocks are applied to each branch of the vector magnitude prediction branch and the direction prediction branch. Each branch uses a different activation function to ensure the range of output values. In the last layer of the two branches, linear transformation and corresponding activation and normalization are used to achieve the final prediction: magnitude ||Δ coarse || and direction δ coarse The prediction;

[0011] (5) Detailed deformation of fine-scale folds: through the detail-aware attribute resolver W parser And Detail Clothing Generator W detail The graphic distribution related to the folds is adjusted to be adaptive, and it is approximated by detailed deformation through two branches.

[0012] This invention creates a dataset consisting of clothing, body shape, and animation pose for training and testing. By observing parameters that affect deformation quality, it generates suitable parameters as one of the network inputs to produce realistic deformations for different clothing fits. Output reconstruction is performed by decomposing the output vector of each vertex into amplitude and direction. Coarse deformation prediction is performed first, followed by detailed deformation of fine-scale folds. Therefore, it can realize the dressing effect of various clothing meshes in any pose and directly generate complex clothing animations with fine details and realistic effects. It outperforms existing clothing animation methods in terms of generalization performance and deformation quality, balancing deformation effect and efficiency, and has significant application value.

[0013] A three-dimensional garment deformation prediction device integrating fit attributes is also provided, the device comprising:

[0014] The dataset building module creates a dataset consisting of clothing, body shape, and animated poses for training and testing.

[0015] The fit and multi-source attribute description module generates fit parameters as one of the network inputs by observing the parameters that affect the deformation quality, so as to produce realistic deformation for different fits of clothing.

[0016] The output reconstruction module reconstructs the output by decomposing the output vector of each vertex into magnitude and direction.

[0017] The coarse deformation prediction module first applies a block to process the input image in the first layer. Then, it applies three blocks to each branch of the vector magnitude prediction branch and the direction prediction branch, with each branch using a different activation function to ensure the range of output values. In the last layer of both branches, linear transformations and corresponding activation and normalization are used to achieve the final prediction: magnitude ||Δ coarse || and direction δ coarse The prediction;

[0018] The detailed deformation module for fine-scale folds, which uses the detail-aware attribute resolver W parser And Detail Clothing Generator W detail The graphic distribution related to the folds is adjusted to be adaptive, and it is approximated by detailed deformation through two branches. Attached Figure Description

[0019] Figure 1 A flowchart of a three-dimensional garment deformation prediction method based on the present invention, incorporating fit properties, is shown.

[0020] Figure 2 A framework diagram of a three-dimensional garment deformation prediction method based on the fusion fit attribute according to the present invention is shown. Detailed Implementation

[0021] like Figure 1 , 2 As shown, this method for predicting 3D clothing deformation by incorporating fit attributes includes the following steps:

[0022] (1) Create a dataset consisting of clothing, body shape and animated poses for training and testing;

[0023] (2) By observing the parameters that affect the deformation quality, suitable parameters are generated as one of the network inputs to generate realistic deformation for different fits of clothing;

[0024] (3) Output reconstruction is performed by decomposing the output vector of each vertex into magnitude and direction;

[0025] (4) Coarse-grained prediction: First, a block is applied to process the input image in the first layer. Then, three blocks are applied to each branch of the vector magnitude prediction branch and the direction prediction branch. Each branch uses a different activation function to ensure the range of output values. In the last layer of the two branches, linear transformation and corresponding activation and normalization are used to achieve the final prediction: magnitude ||Δ coarse || and direction δ coarse The prediction;

[0026] (5) Detailed deformation of fine-scale folds: through the detail-aware attribute resolver W parser And Detail Clothing Generator W detail The graphic distribution related to the folds is adjusted to be adaptive, and it is approximated by detailed deformation through two branches.

[0027] This invention creates a dataset consisting of clothing, body shape, and animation pose for training and testing. By observing parameters that affect deformation quality, it generates suitable parameters as one of the network inputs to produce realistic deformations for different clothing fits. Output reconstruction is performed by decomposing the output vector of each vertex into amplitude and direction. Coarse deformation prediction is performed first, followed by detailed deformation of fine-scale folds. Therefore, it can realize the dressing effect of various clothing meshes in any pose and directly generate complex clothing animations with fine details and realistic effects. It outperforms existing clothing animation methods in terms of generalization performance and deformation quality, balancing deformation effect and efficiency, and has significant application value.

[0028] Preferably, in step (1), the 3D clothing design simulation software Marvelous Designer is used to design and generate clothing deformations with different mesh topologies and vertex numbers, thereby generating ground-based real data of the clothing; coarse data is obtained by applying the Laplacian smoothing operator to each generated clothing mesh; the SMPL parametric human body model is used to sample the second and seventh shape components to generate different human bodies; animation poses are selected from the CMU motion capture and AMASS datasets; the dataset is divided into training set, test set and validation set, and the data in them are not overlapping.

[0029] Preferably, in step (1), when applying the Laplace smoothing operator to obtain coarse data, the diffusion coefficient is 0.12 and the iteration is 40 times; in the training set, a total of 17 clothes and 6 bodies are used, and each pair of clothes and bodies has 2907 poses.

[0030] Preferably, in step (2), for the target human body, an SMPL model is used, and the human body mesh is M... b It means that N b Given the number of vertices, the mesh is parameterized using shape β and pose θ:

[0031]

[0032]

[0033] Among them, W smpl (·) represents the skinning function, derived from the skeleton J(β) pose and skinning weights. For stationary grid To deform, It is by mixing shapes into shape B s (β) and pose blending shape B p (θ) is applied to the average template mesh. To calculate;

[0034] To obtain SMPL human body and clothing The correspondence is defined by defining an indicator matrix. To represent the correlation between clothing vertices and body vertices, we obtain the indicator matrix I by finding the nearest vertex from the clothing to the body. For each pair of clothing and body vertices in a static pose... The corresponding distance vector between the clothing vertex and the body vertex is in ||·|| represents the Euclidean norm operation along the last dimension of the vertex matrix;

[0035] The distance vectors of all clothing pairs are concatenated in parallel and represented by a matrix, thus concisely representing the distance information in each clothing pair. Since the vector length N... g Different clothing pairs cannot be directly connected, based on the minimum number of vertices in the dataset. Uniformly select a fixed length from the distance vector d of each clothing-body pair. Elements to form a new distance vector d * ,in against The choice is through rejection Specific radius This is achieved using vertices that are closer together, where A g This is the area of ​​the clothing mesh; if not enough vertices are returned, the radius is gradually decreased by incrementing the integer C until the number of vertices equals the area of ​​the mesh. Connect the distance vectors of fixed length to form a distance matrix. Store the distance information between all garment pairs.

[0036] Preferably, in step (2), factor analysis (FA) is used to calculate the fitting parameters, and the variance of each vertex is independently modeled; considering the convergence speed, SVD-based probability optimization is used; the FA in the matrix term is defined as:

[0037] D-μ≈LA

[0038] in It is the average vector, and it should be broadcast to... Subtraction with the same size and Represents the loading matrix and factors, where A is a vector. The model is composed of three parameters (α, β, θ) and provides an effective F-dimensional representation for each pair of garments. The parameter α is the fit attribute. During runtime, given a test garment pair in a static pose, the trained FA model is used to perform a matrix multiplication to directly obtain the fit relationship α. The three parameters (α, β, θ) are collectively referred to as multi-source attributes.

[0039] Preferably, in step (3),

[0040] Δ i =||Δ i ||⊙δ i

[0041] The original output is The magnitude of the decomposed vector is Direction is The operator ⊙ is the Hadamard product. In both generators, both networks are designed with two branches to predict the decomposed items separately.

[0042] Based on the numerical characteristics, different activation functions were used in the two branches: in ||Δ i || Use ReLU in the branch to output positive values; in δ i Tanh is used in the branch to map the result value between -1 and 1.

[0043] Preferably, in step (4), the target body M in a fixed posture is... b The clothing worn achieves a seemingly reasonable coarse deformation M. coarse :

[0044] M coarse =IM b +Δ coarse ,

[0045] In the formula Let be the indicator matrix relating the clothing to the vertices of the human body. For the remaining residual... By learning a model W coarse To automatically infer the offset; through the index matrix I, N b′ Each body vertex and N g Each clothing vertex is associated with N. b′ =N g Based on these volume vertices, construct a graph. This graph stores N b′ Vertex features v of the volume vertex b′ and its edge ε b′ , where (i, j) ∈ ε b′ ε represents the edge connection between node i and node j. b′ Indicates the connection between the vertices of the clothing garment;

[0046] Vertex normal Attached to each graph node, the body mesh geometry is encoded, and corresponding skin features are used to reflect the body skin features under different poses. To each node feature, where w s，i G is the skinning weight of vertex i affected by joint s. s (θ) is the rotation matrix of joint s in pose θ, where θ * Indicates a static posture. It is a static posture position; the relative skin feature is a variant of the body vertex position. When the body moves to any position, the body vertex position changes, while the relative skin feature remains unchanged.

[0047] Each node is assigned a fit attribute to represent the relationship between the body and the clothing. The first component α1, v of the most discriminative fit attribute α is used. b′ Features of each node in It consists of three attributes, represented as:

[0048] Design a model W using a graph with defined features as input. coarse This is used to obtain the latent representation of the graph data and map it to the final prediction Δ. coarse ,

[0049] The entire process through the coarse generator can be represented as follows:

[0050]

[0051] During the training process, the displacement Δ coarse and ground truth Minimize the MSE loss between them.

[0052] Preferably, in step (5), the detail-aware attribute parser takes multi-source attributes (α, β, θ) as input and uses W... parser Detail-aware encoding is performed on it, adaptively adjusting the distribution of graph features based on a given input instance; detail-aware encoding is used when the dimension is equal to d. [1] The vector is multiplied along the feature dimension by the transformed graph:

[0053]

[0054] in This refers to the diagram and detail-aware encoding W parser The graph of features injected after the first layer of (α, β, θ). The original features in the data have been adaptively modified by high-dimensional attribute encoding. The new features are expressed in a more detailed manner;

[0055] Will Enter into W detail Subsequent layers, except for the first layer The rest is represented as Detail Generator W detail There are two branches, which approximate the detailed output elements after decomposition: magnitude ||Δ detail || and direction δ detail .

[0056] Preferably, in step (5), for each branch, in addition to graph attention blocks, graph pooling and depooling operations are also applied to avoid overfitting and improve the model's generalization ability; after the first layer The detail generator after the details is represented as:

[0057]

[0058] The predicted ||Δ detail || and δ detail Multiplication yields the corrected displacement Δ detail This is then added to the rough deformation to obtain the final garment detail deformation:

[0059] M detail =M coarse +Δ detail

[0060] During the training process, W was simultaneously... parser and W detail To optimize, MSE loss is used as the loss function to minimize the predicted Δ. detail With the real ground The differences between them.

[0061] Those skilled in the art will understand that all or part of the steps in the methods of the above embodiments can be implemented by a program instructing related hardware. The program can be stored in a computer-readable storage medium. When executed, the program includes the steps of the methods of the above embodiments. The storage medium can be ROM / RAM, magnetic disk, optical disk, memory card, etc. Therefore, corresponding to the method of the present invention, the present invention also includes a three-dimensional clothing deformation prediction device that integrates fit attributes. This device is typically represented in the form of functional modules corresponding to the steps of the method. The device includes:

[0062] The dataset building module creates a dataset consisting of clothing, body shape, and animated poses for training and testing.

[0063] The fit and multi-source attribute description module generates fit parameters as one of the network inputs by observing the parameters that affect the deformation quality, so as to produce realistic deformation for different fits of clothing.

[0064] The output reconstruction module reconstructs the output by decomposing the output vector of each vertex into magnitude and direction.

[0065] The coarse deformation prediction module first applies a block to process the input image in the first layer. Then, it applies three blocks to each branch of the vector magnitude prediction branch and the direction prediction branch, with each branch using a different activation function to ensure the range of output values. In the last layer of both branches, linear transformations and corresponding activation and normalization are used to achieve the final prediction: magnitude ||Δ coarse || and direction δ coarse The prediction;

[0066] The detailed deformation module for fine-scale folds, which uses the detail-aware attribute resolver W parser And Detail Clothing Generator W detail The graphic distribution related to the folds is adjusted to be adaptive, and it is approximated by detailed deformation through two branches.

[0067] To verify the effectiveness of the method, this invention used seven types of clothing, three body types, and 405 poses in the test set. Furthermore, to effectively track training progress, a total of three outfits, two bodies, and 103 poses were used for verification.

[0068] For training W coarse , the diagram The features are input into a graph attention block with a hidden feature size of 256, where the multi-head count is 4, and the feature sizes for the self-enhancing flow and the convergent flow are 128 and 32, respectively. The features are applied through Tanh activation and then input into ||Δ coarse ||Prediction branch and δcoarse In the prediction branch, both branches contain three graph attention blocks with hidden feature sizes of [512, 512, 256]. After graph convolution, three fully connected layers are used to adjust ||Δ coarse The hidden feature sizes in the prediction branch are [256, 128, 1] and δ. coarse The features with hidden size [256, 128, 3] in the prediction branch are transformed.

[0069] To ensure a reasonable output range, ReLU and Tanh activation functions are used after each layer of the two branches, respectively. Normalization is also used for δ. coarse Features in the branches. For training W... parser The multi-source attributes (α, β, θ) are processed through three fully connected layers ([256, 512, 1024]) and the ReLU activation function. Convert it into detail-aware encoding. For training W detail The graph feature G is input into a graph attention block with a hidden feature size of 1024. After injecting detail-aware encoding into the graph feature, the feature dimension remains unchanged, and the features are input into ||Δ detail ||Prediction branch and δ detail In the prediction branch, the structure of the four graph attention blocks ([256, 256, 128, 96]) and the pooling (N... g The operations of roughly reducing it to half and unpooling (restoring) are the same in each branch. Finally, in ||Δ detail ||Prediction branch and δ detail The prediction branch uses fully connected layers with hidden feature sizes of [128, 64, 1] and [128, 64, 3] and their corresponding activations, respectively.

[0070] This invention leverages generalization results from thin, normal, and overweight body types not visible in the training set. Based on predicted coarse deformations, it can generate fine-scale wrinkles that are not significantly different from ground-based data. Furthermore, it can successfully predict personalized and detailed clothing deformations for different body shapes, including rich and believable wrinkles in the left waist region. During training, influencing attributes are transformed into detail-aware encodings, allowing for the learning of clothing deformations in a detail-aware manner, thus effectively and accurately predicting new body types.

[0071] This invention statistically analyzed the error distribution of these three test subjects wearing the same training clothing in the same training posture. The number of vertices with near-zero clothing deformation error was highest for slender individuals, with an average error of approximately 1.33 mm per vertex, as shown in Table 1. Because the clothing folds for slender individuals are simpler than those for overweight individuals, their clothing deformation prediction error is relatively smaller. In contrast, clothing for overweight individuals has more complex folds, making them relatively difficult to predict. This demonstrates the impact of clothing fit on the level of detail.

[0072] Table 1. Average error of single-vertex deformation for different body shapes (mm)

[0073] Body type Slim Normal Fat Rough deformation 2.82 3.01 3.27 Detail deformation 1.33 1.52 1.74

[0074] In summary, through W detail and W parser The deformation refinement reduces the deformation error by about half compared to coarse deformation.

[0075] This invention activates the clothed body and successfully generates details through new postures such as raising hands, walking, and swaying. The wrinkles in areas such as the armpits, waist, and shoulders are rich and closely resemble the real ground effect, with a prediction error of 1.67mm per vertex.

[0076] In addition to constructing the graph during the training process, this invention designs a W... parser This generates detail-aware codes and injects them into a graph neural network, enabling the model to learn personalized deformations caused by different poses.

[0077] The model is generalized to new garments (i.e., long-sleeved T-shirts and vests), with the test garments having different garment meshes and vertex numbers than the training garments. Based on graph learning and the proposed detail-aware strategy, the model can approximate deformations with rich detail regardless of the garment design. Due to the influence of the hem, long-sleeved T-shirts fit the body more closely than vests, resulting in denser wrinkles in the deformations around the abdomen and waist.

[0078] Table 2 shows the results for clothing, body shape, and posture.

[0079] Table 2. Average error of deformation for each vertex (mm)

[0080] Test case Rough prediction Detail prediction Skirt + slim body + new pose 3.46 2.11 Jacket + regular + new pose 3.65 2.39 Coat + fat body + new pose 3.63 2.27

[0081] The test garments included dresses, short-sleeved jackets, and 3 / 4-sleeved coats with different mesh topologies. Humans of new body types wearing these garments were animated in new poses. Although all three variables were novel and not present in the training set, the predictions still matched the underlying facts and successfully generated wrinkles at most fine scales. Overall, the average point error of the predictions across all test data was approximately 2.24 mm. The results demonstrate that the method has strong generalization ability, can handle novel variational terms simultaneously, and is easily integrated into practical applications.

[0082] This invention verifies the effectiveness of output decomposition, detail-aware attribute parser, step-by-step deformation estimation, and graph pooling operations by conducting ablation studies.

[0083] To evaluate the proposed output decomposition, the output displacement of each vertex is first preserved as the original three-dimensional vector Δ. detail (w / o decomposition) allows the network to have an unbounded prediction range. Furthermore, this invention sets a finite range of values ​​for the 3D output displacement (w / output constraint), where the finite range is obtained by scaling each original displacement value to (-1, 1) using a scaling factor of 5 determined by the dataset. The detail-aware attribute parser was evaluated by testing the case of removing the attribute parser (w / o attribute parser), where attributes are directly assigned to each graph node, and the case of replacing element-wise smart multiplication with concatenation in Equation 7 (concat attribute parser).

[0084] To evaluate the effectiveness of the two-step deformation estimation and graph pooling operations, this invention uses a single model instead of two-step approximation (without two steps) and removes the pooling operators (without pooling). With the layers removed, the number of parameters in the remaining layers is increased to ensure approximate consistency of capacity, and the best-tested initialization scheme is selected for all evaluations: Glorot initialization for graph convolutional layers and Kaiming initialization for other layers.

[0085] This invention plots the average error of each vertex during the verification process.

[0086] The method without output decomposition produces the largest error because the output is difficult to approximate using only three values ​​from negative infinity to positive infinity. As the epoch increases, it still suffers from highly complex outputs and an unreducible error between 2.5 and 3 mm. Secondly, setting the output within a finite range results in a smaller prediction error than in the case of unbounded output, but convergence is not ideal. Even with a network structure of roughly the same capacity as the original network after removing the attribute parser, the deformation error remains significant. Applying concatenation instead of element-wise multiplication in Equation 7 leads to even greater errors because attribute information cannot be accurately injected into the graph features.

[0087] Without a two-step strategy and pooling operations, quantitative results will be affected to varying degrees. Reducing the output to a finite range can improve the deformation effect, but it will lose much detailed wrinkle detail. For results without an attribute parser, the waist and collar deformations show obvious and wide wrinkles, reflecting a certain degree of wrinkling trend. When using concatenation to combine features processed by the attribute parser with graphic features, the qualitative results have a larger error.

[0088] Without the pooling operation, the location and trend of some wrinkles differ from reality, indicating that graph pooling contributes to the model's generalization performance, i.e., it enables effective inference using learned information. In contrast, the complete method of this invention successfully generates vivid wrinkles and restores near-realistic detail.

[0089] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any way. Any simple modifications, equivalent changes, and alterations made to the above embodiments based on the technical essence of the present invention shall still fall within the protection scope of the present invention.

Claims

1. A three-dimensional garment deformation prediction method fusing fit attributes, characterized in that: The method comprises the following steps: (1) creating a dataset consisting of clothes, body shapes and animation poses for training and testing; (2) generating fitting parameters as one of the network inputs by observing parameters affecting deformation quality to produce realistic deformation for different fitting degrees of clothes; (3) performing output reconstruction by decomposing the output vector of each vertex into amplitude and direction; (4) Coarse deformation prediction: first apply one block to process the input image at the first layer, then apply three blocks on each branch of the vector magnitude prediction branch and the direction prediction branch, each branch adopts different activation functions to ensure the value range of the output value; at the last layer of the two branches, use linear transformation and the corresponding activation and normalization, so as to realize the final prediction: the prediction of the magnitude ||Δ coarse and the direction δ coarse ; (5) Detailed deformation of fine-scale wrinkles: through the fine detail perceptual property resolver W parser and the detailed garment generator W detail , adjust the graph-adaptive distribution related to wrinkles and pass it through two branches for detailed deformation approximation; In the step (5), the detail-aware attribute parser takes the multi-source attribute (a, b, q) as input, and passes it through W parser The detail-aware encoding is a d-dimensional vector, which is multiplied with the transformed graph along the feature dimension: [1] The detail-aware encoding is a d-dimensional vector, which is multiplied with the transformed graph along the feature dimension: wherein refers to the graph and the details-aware encoding W parser the first layer of (a, b, q) injects features into the graph, the original features in have been adaptively modified by the high-dimensional attribute encoding, the new features in are expressed in a more detailed manner; will be input to the subsequent layers, except the first layer detail The rest of the representation is The detail generator W detail has two branches, approximating the decomposed detail output elements: magnitude ||Δ detail || and direction δ detail .​ 2. The three-dimensional garment deformation prediction method of fusion fit attributes according to claim 1, characterized in that: In step (1), 3D clothing design simulation software MarvelousDesigner is used to design and generate clothes deformation with different mesh topologies and vertex numbers, thereby generating ground truth data of clothes; Rough data is obtained by applying a Laplacian smoothing operator to each generated clothing mesh; The SMPL parameterized human body model is used to sample the second and seventh shape components to generate different human bodies; animation poses are selected from the CMU motion capture and AMASS datasets; The dataset is divided into a training set, a test set and a validation set, and the data in them are not overlapped.

3. The three-dimensional garment deformation prediction method of fusion fit attributes according to claim 2, characterized in that: In step (1), when applying the Laplacian smoothing operator to obtain rough data, the diffusion coefficient is 0.12 and the iteration is 40 times; in the training set, a total of 17 clothes and 6 bodies are used, and there are 2907 poses for each pair of clothes and body.

4. The three-dimensional garment deformation prediction method of fusion fit attributes according to claim 3, characterized in that: In the step (2), for the target human body, the SMPL model is adopted, and the human body mesh is parameterized by M b N b represents the number of vertices, and the mesh is parameterized by shape β and pose θ: where W smpl (·) is the skinning function that applies the skeleton J(β) pose and skinning weights to the rest pose mesh to deform it, is computed by applying the shape blend shape B s (β) and pose blend shape B p (θ) to the mean template mesh ; To get the correspondence between SMPL human body and the garment , the garment vertices are associated with the body vertices by defining an indicator matrix , which is obtained by finding the closest body vertex from the garment for each pair of garment and body correspondence in the rest poses The distance vector between the corresponding garment vertex and body vertex is where ||·|| denotes the Euclidean norm operation along the last dimension of the vertex matrix; The distance vectors of all garment-body pairs are concatenated and represented by a matrix, which succinctly represents the distance information in each pair of garment and body, since the vector length N g In different garment-body pairs, different cannot be directly connected, based on the minimum number of vertices in the dataset From the distance vector d of each garment-body pair, a fixed length of Elements are selected to form a new distance vector d * Where The selection of Is achieved by rejecting Vertices closer than a radius Where A g Is the area of the garment mesh; if not enough vertices are returned, the radius is gradually decreased by increasing the integer C until the number of vertices is equal to The fixed length of distance vectors are concatenated to form a distance matrix Stores the distance information between all garment-body pairs.

5. The three-dimensional garment deformation prediction method of fusion fit attributes according to claim 4, characterized in that: In step (2), factor analysis FA is used to calculate fitting parameters to independently model the variance of each vertex; considering the convergence speed, SVD-based likelihood optimization is used; the FA in the matrix term is defined as: D-μ≈LA where is the mean vector, which should be broadcast to the same size as for subtraction, and denote the loading matrix and factors, A is composed of vectors , which provides an effective F-dimensional representation for each pair of body, the parameter a is the fit attribute, in the running, given the test body pair of static poses, using the trained FA model to perform a matrix multiplication, directly get the fit relationship a; the three parameters (a, b, q) are collectively referred to as multi-source attributes.

6. The three-dimensional garment deformation prediction method of fusion fit attributes according to claim 5, characterized in that: In step (3), Δ i =||Δ i ||⊙δ i where the original output is The decomposed vector magnitude is The direction is The operator is Hadamard product, in two generators, two networks are designed two branches, in order to predict the decomposed item respectively; According to the numerical characteristics, different activation functions are employed in the two branches: in the ||Δ i branch, ReLU is used to output positive values; in the δ i branch, Tanh is used to map the resulting values between -1 and 1.

7. The three-dimensional garment deformation prediction method of fusion fit attributes according to claim 6, characterized in that: In said step (4), the target body M in the fixed pose b Realized on the clothes worn, the garment coarse deformation M coarse : M coarse = IM b + Δ coarse , In the formula is an indication matrix of the garment associated with the vertices of the human body, for the remaining residual part by learning a model W coarse to automatically infer the offset; N b′ body vertices are associated with N g garment vertices by means of an index matrix I, where N b′ = N g , based on these body vertices, a graph G is constructed The graph stores the vertex features v b′ of the N b′ body vertices and its edges ε b′ , where (i, j) ∈ ε b′ represents an edge connection between node i and node j, ε b′ represents the connection of garment vertices; Vertex normal Attached to each graph node, the body mesh geometry is encoded, and corresponding skin features are used to reflect the body skin features under different poses. To each node feature, where w s，i G is the skinning weight of vertex i affected by joint s. s (θ) is the rotation matrix of joint s in attitude θ, where θ * Indicates a static posture. It is a static posture position; the relative skin feature is a variant of the body vertex position. When the body moves to any position, the body vertex position changes, while the relative skin feature remains unchanged. Each node is attached with a fit property to represent the relationship between body and garment, the first component of the most discriminative fit property α Each node feature in Composed of three properties, denoted as: A model W is designed to take a graph with defined features as input coarse , to obtain a latent representation of the graph data and map it to the final prediction Δ coarse The whole process through the coarse generator is represented as: During training, the MSE loss between the displacement Δ coarse and the ground truth is minimized.

8. The three-dimensional garment deformation prediction method of fusion fit attributes according to claim 7, characterized in that: In step (5), for each branch, in addition to the graph attention block, graph pooling and unpooling operations are also applied to avoid overfitting problems and improve the model generalization ability. After the first layer The details-after-detail generator is represented as: The predicted ||Δ detail || and δ detail are multiplied to get the modified displacement Δ detail and added to the rough deformation to get the final garment detail deformation: M detail = M coarse + Δ detail During the training process, W parser and W detail are optimized simultaneously with the MSE loss as the loss function to minimize the difference between the predicted Δ detail and the ground truth Δ .

9. A three-dimensional garment deformation prediction apparatus fusing fit attributes, characterized by: The device is used to perform the three-dimensional clothing deformation prediction method of fusion fitting property according to claim 1, and the device comprises: A dataset construction module creates a dataset consisting of clothes, body shapes and animation poses for training and testing; A fitting degree and multi-source attribute description module generates fitting parameters as one of the network inputs by observing parameters affecting deformation quality to produce realistic deformation for different fitting degrees of clothes; An output reconstruction module performs output reconstruction by decomposing the output vector of each vertex into amplitude and direction; a coarse deformation prediction module which first applies one block at the first layer to process the input image, then applies three blocks on each branch of the vector magnitude prediction branch and the direction prediction branch, each branch employs different activation functions to guarantee the range of the output values; at the last layer of the two branches, linear transformation and corresponding activation and normalization are used to achieve the final prediction: the prediction of the magnitude ||Δ coarse and the direction δ coarse ; a detailed deformation module for fine-scale wrinkles, which resolves the fine detail perceptual attributes with a detail-aware attribute resolver W parser and a detailed garment generator W detail adjusts the graph-adaptive distribution associated with the wrinkles and approximates its detailed deformation with two branches.

Images