Method and device for estimating physical properties parameters of fabric

Abstract

A method and apparatus for estimating physical property parameters of a fabric according to an embodiment can obtain information on a three-dimensional contour shape of a fabric placed on a three-dimensional geometric object; apply the information to a pre-trained artificial neural network to estimate material property parameters of the fabric for reproducing drape shapes of a three-dimensional garment made of the fabric; and output the material property parameters of the fabric.

Description

Technical Field

[0001] The embodiments relate to a method for generating training data of an artificial neural network for estimating fabric physical property parameters, a method and an apparatus for estimating fabric physical property parameters. Background Art

[0002] Although clothes are three-dimensional when worn on the body, in reality, clothes are closer to two-dimensional and are a combination of fabric pieces cut according to two-dimensional patterns. The fabric used as clothing material is flexible, so it may present different appearances according to the body shape or movement of the wearer. In addition, the fabric has various properties such as stiffness, stretchability, shrinkage rate, etc. Based on the physical property differences of different fabrics, even clothes with the same design will have different manifestations and feelings.

[0003] The above background art was mastered or learned by the inventor during the development of the present invention and should not be construed as generally known prior art that must be publicly available before the application of the present invention. Summary of the Invention

[0004] Technical Problem to be Solved

[0005] According to one embodiment, by applying the three-dimensional contour shape and density of the fabric to a pre-trained neural network, the fabric physical property parameters for reproducing the drape shapes of three-dimensional clothing made of the fabric can be estimated more accurately.

[0006] According to one embodiment, by using the three-dimensional contour shape of a rectangular fabric, the diversity of the drape shapes of three-dimensional clothing can be ensured.

[0007] According to one embodiment, by using the three-dimensional contour shape (contour line) of the fabric instead of the overall fabric coverage shape, the data size for training the artificial neural network can be reduced.

[0008] According to one embodiment, by using the three-dimensional contour shape of the fabric instead of the overall fabric coverage shape, the complexity of the calculations performed by the neural network for estimating the fabric physical property parameters can be reduced.

[0009] According to one embodiment, by simultaneously using the three-dimensional contour shape and the fabric density of the fabric, the training accuracy of the artificial neural network can be improved.

[0010] Technical Solution for Solving the Problem

[0011] A method for estimating fabric physical property parameters according to an embodiment may include the following steps: obtaining information including a three-dimensional contour shape of a fabric placed on a three-dimensional geometric object; applying the information to a pre-trained artificial neural network to estimate the material property parameters of the fabric for reproducing the drape shapes of a three-dimensional garment made of the fabric; and outputting the material property parameters of the fabric.

[0012] The three-dimensional contour shape of the fabric and the drape shape of the three-dimensional garment may be correlated with each other.

[0013] The step of obtaining the information may include the following steps: receiving an image capturing the three-dimensional contour shape of the fabric placed on the three-dimensional geometric object; generating a three-dimensional model including the three-dimensional contour shape of the fabric from the image; and extracting three-dimensional vertex coordinates corresponding to the three-dimensional contour shape of the fabric from the three-dimensional model.

[0014] The step of obtaining the information may include the following steps: obtaining three-dimensional vertex coordinates corresponding to the three-dimensional contour shape of the fabric placed on the three-dimensional geometric object.

[0015] The step of obtaining the three-dimensional vertex coordinates may include the following steps: sampling the three-dimensional vertices corresponding to the three-dimensional contour shape from a depth image or a three-dimensional scan image including the three-dimensional contour shape of the fabric placed on the three-dimensional geometric object; and obtaining the sampled three-dimensional vertex coordinates.

[0016] The information may further include the density of the fabric.

[0017] The fabric may include at least one of natural fiber fabrics, synthetic fiber fabrics, and blended yarn fabrics including cotton, linen, wool, polyester, nylon, elastane, etc.

[0018] At least a partial region of the fabric is placed on and supported by the three-dimensional geometric object, and the remaining region is not supported by the three-dimensional geometric object and droops. The three-dimensional contour shape of the fabric is formed by the outer contour line of the remaining region of the fabric that is not supported by the three-dimensional geometric object and droops.

[0019] The physical property parameters of the fabric may include at least one of the stretch - weft stiffness, stretch - warp stiffness, shear stiffness, banding - weft stiffness, banding - warp stiffness, and bending bias stiffness of the fabric.

[0020] The artificial neural network can be trained with training data, which is generated based on the physical property parameters of the fabric randomly sampled according to the probability distribution of the Gaussian Mixture Model (GMM).

[0021] The artificial neural network may include a first neural network for estimating the physical property parameters related to the stiffness of the fabric; and a second neural network for estimating the physical property parameters related to the bending of the fabric.

[0022] The first neural network and the second neural network may each include a Fully Connected Neural Network Model.

[0023] The step of outputting the physical property parameters of the fabric may include the following steps: applying the physical property parameters of the fabric to the 3D garment; and displaying the result covering the 3D garment.

[0024] A method for generating artificial neural network training data according to an embodiment may include the following steps: collecting a first quantity of physical property parameters of the fabric; sampling a second quantity of physical property parameters using a generation model based on the physical property parameters, where the second quantity is greater than the first quantity; simulating the covering shape of the corresponding fabric based on the sampled physical property parameters to obtain the 3D contour shape of the corresponding fabric; and generating training data corresponding to the corresponding fabric based on the sampled physical property parameters and the 3D contour shape of the corresponding fabric.

[0025] The generation model may randomly sample the physical property parameters according to the probability distribution of the Gaussian Mixture Model (GMM).

[0026] The artificial neural network can be trained using the training data corresponding to the corresponding fabric to estimate the physical property parameters of the corresponding fabric.

[0027] An apparatus for estimating physical property parameters of a fabric according to an embodiment may include: a communication interface that acquires information including a three-dimensional contour shape of a fabric placed on a three-dimensional geometric object; a processor that applies the information to a pre-trained artificial neural network to estimate the physical property parameters of the fabric for reproducing a covering shape of a three-dimensional garment made of the fabric; and an output device that outputs the physical property parameters of the fabric.

[0028] The three-dimensional contour shape of the fabric and the covering shape of the three-dimensional garment may be correlated with each other.

[0029] The communication interface may receive an image capturing the three-dimensional contour shape of the fabric placed on the three-dimensional geometric object, and the processor may generate a three-dimensional model including the three-dimensional contour shape of the fabric from the image and extract three-dimensional vertex coordinates corresponding to the three-dimensional contour shape of the fabric from the three-dimensional model.

[0030] Advantages of the Invention

[0031] According to one aspect, by applying the three-dimensional contour shape and density of a fabric to a pre-trained neural network, the physical property parameters of the fabric for reproducing the covering shapes of three-dimensional garments made of the fabric can be estimated more accurately.

[0032] According to one aspect, by using the three-dimensional contour shape of a rectangular fabric, the diversity of the covering shapes of three-dimensional garments can be ensured.

[0033] According to one aspect, by using the three-dimensional contour shape (contour line) of a fabric instead of the overall covering shape of the fabric, the data size for training an artificial neural network can be reduced.

[0034] According to one aspect, by using the three-dimensional contour shape of a fabric instead of the overall covering shape of the fabric, the complexity of the calculations performed by the neural network for estimating the physical property parameters of the fabric can be reduced.

[0035] According to one aspect, by using both the three-dimensional contour shape and the fabric density of a fabric, the training accuracy of the artificial neural network can be improved. BRIEF DESCRIPTION OF THE DRAWINGS

[0036] Figure 1 is a flowchart showing a method for estimating physical property parameters according to an embodiment.

[0037] Figure 2 is a drawing illustrating a method for acquiring three-dimensional contour shape information of a fabric according to an embodiment.

[0038] Figure 3It is a drawing for explaining the training principle of an artificial neural network according to an embodiment.

[0039] Figure 4 It is a drawing for explaining the operation of an artificial neural network according to an embodiment.

[0040] Figure 5 It is a drawing for explaining the correlation between physical property parameters estimated according to an embodiment and the three-dimensional contour shape of a fabric.

[0041] Figure 6 It is a drawing showing the structure of an artificial neural network according to an embodiment.

[0042] Figure 7 It is a flowchart showing a method for generating training data of an artificial neural network according to an embodiment.

[0043] Figure 8 It is a block diagram of an apparatus for estimating physical property parameters according to an embodiment.

[0044] Explanation of reference numerals

[0045] 800: Estimation apparatus

[0046] 805: Communication bus

[0047] 810: Communication interface

[0048] 830: Processor

[0049] 850: Memory

[0050] 870: Output device Detailed description of the embodiments

[0051] Embodiments will be described in detail below with reference to the drawings. However, various changes can be made to the embodiments, and the scope of the rights of this application is not limited to the above embodiments. It should be understood that all changes to the embodiments, their equivalents, and their alternatives are included in the scope of the rights.

[0052] The terms used in the embodiments are only for explaining specific embodiments and are not used to limit the scope of protection. Without special explanation in the content, singular expressions include plural meanings. In this specification, terms such as "including" or "having" are used to express the presence of features, numbers, steps, operations, components, parts, or combinations thereof described in the specification, and do not exclude the presence of one or more other features, numbers, steps, operations, components, parts, or combinations or additional functions.

[0053] Unless otherwise defined, all terms used herein, including technical or scientific terms, shall have the ordinary meaning as understood by those of ordinary skill in the art. Terms commonly used, such as those defined in a dictionary, shall be understood to have the meaning in the context of the relevant technical content. Without specific definition in this specification, they shall not be construed as having an idealized or overly formal meaning.

[0054] Also, in the process of description with reference to the drawings, regardless of the reference numerals, the same components are denoted by the same reference numerals, and repeated descriptions are omitted. In the process of describing the embodiments, when it is determined that the detailed description of the relevant well-known technologies will unnecessarily obscure the embodiments, the detailed description thereof is omitted.

[0055] In addition, when describing the components of the embodiments, terms such as first, second, A, B, (a), (b), etc. may be used. Such terms are only used to distinguish the components from other components, however, such terms do not limit the nature, order or sequence of the components. It should be understood that when describing that a certain component "connects", "combines" or "contacts" other components, each component may directly connect or contact another component, but there may also be other components "connecting", "combining" or "contacting" between the components.

[0056] When a certain component has the same function as the components in other embodiments, the same name is used for description in other embodiments. Without mention of counterexamples, the description of one embodiment can be applied to other embodiments, and detailed descriptions of repeated content are omitted.

[0057] Figure 1 is a flowchart showing a method for estimating physical property parameters according to an embodiment. Refer to Figure 1 , according to an embodiment, an apparatus for estimating physical property parameters (hereinafter referred to as "estimating apparatus") can estimate the physical property parameters of a fabric through steps 110 to 130.

[0058] In step 110, the estimation device obtains information including the three-dimensional contour shape of a fabric placed on a three-dimensional geometric object. Non-limiting examples of the three-dimensional geometric object include a cylinder, a cube, a sphere, a microbody, or a mannequin. The fabric may include natural fiber fabrics, synthetic fiber fabrics, and at least one of blended yarn fabrics such as cotton, linen, wool, polyester, nylon, and elastane. Non-limiting examples of the fabric shape are rectangular or circular. At this time, at least a part of the fabric area is placed on and supported by the three-dimensional geometric object, and the remaining area hangs down without being supported by the three-dimensional geometric object. In one embodiment, the three-dimensional contour shape of the fabric may be the outer contour line of the remaining area of the fabric that hangs down without being supported by the three-dimensional geometric object. In an exemplary embodiment, the three-dimensional contour shape of the fabric may be the three-dimensional contour shape of the fabric in a fixed static state.

[0059] In step 110, the estimation device may extract three-dimensional vertex coordinates corresponding to the three-dimensional contour shape of the fabric from an image capturing the three-dimensional contour shape of the fabric placed on the three-dimensional geometric object. Since the possibility of changes in photo or video data due to external factors (e.g., including all external factors other than the nature of the fabric itself) is low, the three-dimensional vertex coordinates extracted by the estimation device using the image have the same effect as the actual three-dimensional vertex coordinates. Alternatively, the estimation device may directly obtain the three-dimensional vertex coordinates corresponding to the three-dimensional contour shape of the fabric placed on the three-dimensional geometric object. The method by which the estimation device obtains information including the three-dimensional contour shape of the fabric will be described in more detail with reference to Figure 2 More specifically.

[0060] According to an embodiment, in step 110, in addition to obtaining the three-dimensional contour shape of the fabric, the estimation device may also obtain the fabric density. At this time, the fabric density may be the value obtained by dividing the mass of the corresponding fabric by the total fabric area. When obtaining the fabric density in addition to the three-dimensional contour shape of the fabric, the estimation device may further consider characteristics that cannot be confirmed by the naked eye when estimating the physical properties of the fabric in step 120.

[0061] In step 120, the estimation device applies the information obtained in step 110 to a pre-trained artificial neural network, thereby estimating the material property parameters of the fabric for reproducing the drape shapes of the three-dimensional clothing made of the fabric. In this specification, "draping" can be understood as a process of dressing a three-dimensional avatar or the like in the three-dimensional clothing using a computer program, where the three-dimensional clothing is made of a fabric reflecting the estimated fabric material properties. In one embodiment, the three-dimensional contour shape of the fabric and the drape shape of the three-dimensional clothing may be correlated with each other.

[0062] An artificial neural network according to an embodiment may be a neural network trained in terms of the material property parameters of the fabric, where the material property parameters of the fabric are used to reproduce the static drape shape of the fabric.

[0063] An artificial neural network according to an embodiment may utilize an estimation model such as Equation 1 for definition. The estimation model may be defined by a linear regression model.

[0064] Equation 1

[0065]

[0066] where may represent a vector composed of 6 material property parameters to be estimated by the estimation model . is the element-wise logarithm of each element, is the feature vector. is a set of sampling points of the three-dimensional contour shape of the fabric, a set of sampling points of the fabric density.

[0067] In one embodiment, an estimation model trained with training data composed of a large dataset may be used to estimate the material property parameters of the fabric. At this time, as an example, a large dataset may be generated using a generation model of 400 mechanical properties of actual fabric materials. As an example, the generation model may be a Gaussian Mixture Model (GMM).

[0068] For example, the artificial neural network may be trained with training data generated from the material property parameters of the fabric randomly sampled based on the probability distribution of the Gaussian Mixture Model (GMM).

[0069] The training principle of the artificial neural network according to an embodiment and the operation of the artificial neural network will be described with reference to Figures 3 to 4 more specifically. And the structure of the artificial neural network will be described with reference to Figure 6 more specifically.

[0070] In step 120, non-limiting examples of the estimated fabric physical property parameters may include the stretch - weft stiffness, stretch - wrap stiffness, shear stiffness, banding - weft stiffness, banding - wrap stiffness, and bending bias stiffness of the fabric, etc. Among them, "weft" represents the horizontal line of the fabric, also known as "weft thread". And "warp" represents the vertical line of the fabric, also known as "warp thread".

[0071] In step 130, the estimation device outputs the physical property parameters of the fabric estimated in step 120. The estimation device can explicitly output or implicitly output the physical property parameters of the fabric estimated in step 120. In an exemplary embodiment, "explicitly outputting the physical property parameters of the fabric" may include directly displaying the physical property parameters of the fabric on a display panel and / or outputting them to paper for display. Or as an example, "implicitly outputting the physical property parameters of the fabric" may include displaying the simulation result of a three - dimensional garment, or the simulation result of covering the corresponding three - dimensional garment on a three - dimensional avatar, etc., where the three - dimensional garment is made of the fabric using the fabric physical property parameters.

[0072] Figure 2 It is a drawing illustrating a method for obtaining three - dimensional contour shape information of a fabric according to an embodiment. Figure 2 Taking a cylinder 211 as a three - dimensional geometric object, the state 210 of spreading a rectangular fabric 213 on the cylinder 211 is shown. The fabric 213 can correspond to a specimen of a certain size of the fabric.

[0073] At this time, as an example, the diameter of the cylinder 211 can be 10 cm and the height can be 20 cm. And the fabric 213 can be a square with a length and width of 30 cm. The fabric 213 is placed on the cylinder 211, and the center of the fabric 213 and the center on the cylinder 211 can be at the same point.

[0074] When Figure 2 the fabric 213 is placed as shown by 210 in Figure 2 , the portion of the fabric 213 not supported by the cylinder 211 will droop as shown by 230 in

[0075] In one embodiment, as shown by 230 in Figure 2 , the "cylinder test" is a process of causing the fabric area not supported by the cylinder to droop to generate wrinkles.

[0076] In one embodiment, the cylinder test can be simulated to extract feature vectors from the fabric, and an artificial neural network is used to predict physical property parameters, and the visual similarity between the prediction result and the result using the actual fabric is evaluated. During the cylinder test, the three-dimensional contour shape of the fabric will vary according to the method of placing the corresponding fabric sample. At this time, it is very important to make the four corners of the fabric fall at the same height simultaneously. This process can be repeated multiple times for a fabric, and the three-dimensional contour shape of the fabric that appears repeatedly can be selected from them.

[0077] In one embodiment, using a square fabric to perform the cylinder test can make the depression and twisting direction of the fabric clearer, and more covering areas with various potential covering shapes can be derived. For a square fabric, the area of the corner region is larger than that of the remaining region, so the weight of the corner region is also heavier than that of the remaining region. As a result, the change in the covering shape due to the weight distribution of the rectangular fabric can be observed more clearly.

[0078] An estimation device according to an embodiment can obtain information through the cylinder test, where the information includes the three-dimensional contour shape of the fabric corresponding to a partial feature vector for estimating the physical property parameters of the fabric. The three-dimensional contour shape of the fabric can be a three-dimensional closed curve.

[0079] When Figure 2 as shown by 250 in

[0080] In one embodiment, when estimating the physical property parameters of a fabric, using the three-dimensional contour line of the fabric to replace the overall covering shape of the fabric can significantly reduce the complexity of the artificial neural network training process. This is generally based on the assumption that there is a one-to-one correspondence between the drape shapes of the fabric shown in the cylinder test and the three-dimensional contour shape of the fabric.

[0081] In one embodiment, as shown in 510 below, Figure 5 it is possible to confirm the possibility of obtaining the physical property parameters trained from the three-dimensional contour line of the fabric by using the visual correlation matrix between the three-dimensional contour line of the fabric and the physical property parameters. For this purpose, the physical property parameters and simulation results of 400 actual fabric samples are required.

[0082] According to an exemplary embodiment, the estimation device can receive an image capturing the three-dimensional contour shape of the fabric placed on a three-dimensional geometric object and generate a three-dimensional model including the three-dimensional contour shape of the fabric from the image. Among them, the three-dimensional model can correspond to the mesh model of the fabric. The estimation device can extract three-dimensional vertex coordinates corresponding to the three-dimensional contour shape of the fabric from the three-dimensional model. For example, the estimation device extracts vertices from the boundary of the mesh model of the fabric and determines the extracted vertex coordinates as the three-dimensional vertex coordinates corresponding to the three-dimensional contour shape of the fabric.

[0083] According to an embodiment, the estimation device can restore the three-dimensional contour shape of the covered fabric by the following method. For example, the estimation device can use an artificial neural network capable of restoring the three-dimensional contour shape from the two-dimensional plan view of the fabric. The artificial neural network can be trained based on the pair data of [two-dimensional plan view - three-dimensional contour shape]. Alternatively, the estimation device can reconstruct a three-dimensional model based on two-dimensional images taken from multiple angles, generate a three-dimensional mesh of the fabric, and then restore the three-dimensional contour shape from the three-dimensional mesh. Alternatively, the estimation device can use a three-dimensional scanner to scan the covered fabric, generate a three-dimensional mesh, and then restore the three-dimensional contour data from the three-dimensional mesh.

[0084] As an example, a mesh model of a fabric can be modeled by a mesh including a plurality of polygons (e.g., triangles). The three vertices of the polygon (triangle) can be point masses with mass, and the sides of the polygon can be represented as elastic springs connecting the point masses. Thus, as an example, the fabric can be modeled by a Mass-Spring Model. Among them, the springs can have different resist values for stretch, shear, bending, etc. according to the physical property parameters of the fabric (fabric). Each vertex can move according to external forces (such as gravity) and internal forces of stretch, shear, and bending. For example, by calculating the external force and internal force to find the force applied to each vertex, the displacement and movement speed of each vertex can be obtained. And the movement of the virtual clothing can be simulated by the movement of the vertices of the polygon at each time step.

[0085] In one embodiment, draping a garment made of a fabric modeled by a mesh over a three-dimensional avatar can embody a natural three-dimensional virtual garment based on physical laws.

[0086] In an exemplary embodiment, the image of the three-dimensional contour shape of the fabric is an image obtained by scanning the three-dimensional contour shape of the fabric placed on a three-dimensional geometric object using a mobile device equipped with a depth sensor or a depth camera. The estimation device generates a three-dimensional mesh model of the fabric from the scanned image and samples the three-dimensional vertices on the mesh surface, thereby extracting the three-dimensional contour shape of the fabric. The estimation device can use a Bezier curve, etc. to interpolate the sampled vertices. According to the embodiment, in order to extract the three-dimensional contour shape of the fabric, the estimation device can also first scan the three-dimensional mesh model of the corresponding fabric sample, and then sample the control points of the Bezier spline.

[0087] Alternatively, the estimation device can also directly obtain the three-dimensional vertex coordinates corresponding to the three-dimensional contour shape of the fabric placed on the three-dimensional geometric object. For example, the estimation device can sample the three-dimensional vertices corresponding to the three-dimensional contour shape from a depth image or a three-dimensional scan image including the three-dimensional contour shape of the fabric placed on the three-dimensional geometric object. The estimation device can obtain the sampled three-dimensional vertex coordinates.

[0088] For example, during the simulation of a fabric, the fabric weight has a great influence on the final covering result. Therefore, in one embodiment, the fabric density (weight) is additionally considered to estimate the physical property parameters. In many cases, the fabric density may not be visually recognizable. For example, the area of the fabric that sags without being supported by a three-dimensional geometric object may have a greater density or a smaller stiffness than the area supported by the three-dimensional geometric object. In one embodiment, to consider the invisible covering aspect, the fabric density is used as part of the feature vector. The fabric density can be measured by dividing the corresponding fabric mass by the corresponding total fabric area. In one embodiment, in addition to the three-dimensional contour shape of the fabric, the fabric density is also used to improve the training accuracy.

[0089] Figure 3 is a drawing for explaining the training principle of an artificial neural network according to an embodiment. Referring to Figure 3 , a photograph 310 of a fabric placed on a three-dimensional geometric object (such as a cylinder), a machine learning model 330, and physical property parameters 350 of the fabric output from the machine learning model 330 are shown. The machine learning model may correspond to an artificial neural network according to an embodiment.

[0090] For example, the drape property of a fabric can have a decisive influence on the appearance of the fabric. As a visually recognizable property, the drape property of the fabric can be derived through machine learning from a photo or video database. However, when using photos or videos as training data, due to uncontrollable external factors such as the shooting angle, lens properties, lighting, and fabric color, the required data diversity is usually too high. To process diverse data, more training data and a more complex training model such as a convolutional layer are required.

[0091] Therefore, in one embodiment, the machine learning model 330 is trained to estimate the physical property parameters 350 of the fabric from the photograph 310 of the fabric placed on a three-dimensional geometric object (such as a cylinder), thereby reproducing the static covering shape of the target fabric.

[0092] Figure 4 is a drawing for explaining the operation of an artificial neural network according to an embodiment. Referring to Figure 4 , the artificial neural network according to an embodiment performs a first step of restoring a three-dimensional model 420 from a photograph 410 of a fabric placed on a three-dimensional geometric object (such as a cylinder), and then performs a second step of extracting a three-dimensional boundary curve 430 of the fabric from the restored three-dimensional model and estimating the physical property parameters 440 of the fabric, thereby outputting the physical property parameters 450.

[0093] According to an embodiment, the operations of the first step and the second step are both performed by an artificial neural network included in the estimation device, or the operation of the first step is performed by the estimation device and the operation of the second step is performed by an artificial neural network included in the estimation device.

[0094] For example, the photo 410 may be an image of the three-dimensional contour shape of the fabric placed on a three-dimensional geometric object.

[0095] After receiving the photo 410 of the fabric, the artificial neural network according to an embodiment may restore 420 or reconstruct a three-dimensional model including the three-dimensional contour shape of the fabric from the photo 410. Among them, the three-dimensional model is equivalent to the mesh model of the fabric mentioned above. Since the fabric is in a static state and its original shape and size are known, the artificial neural network can more easily restore the three-dimensional model.

[0096] The artificial neural network may extract the boundary curve 430 of the fabric equivalent to the three-dimensional contour shape of the fabric from the restored three-dimensional model 420. At this time, the artificial neural network may extract the three-dimensional vertex coordinates corresponding to the boundary curve 430 of the fabric. Before providing the three-dimensional coordinates of the boundary curve 430 to the neural network, a normalization operation needs to be performed first. For example, the three-dimensional coordinates may be normalized with respect to the minimum and maximum values of each coordinate. Assuming that the cylinder is in the middle, for the minimum and maximum values of the x and z coordinates, they may be half of the horizontal / vertical dimensions of the sample fabric, and for the minimum value of the y coordinate, it is 0 and the maximum value is the height of the cylinder. The normalized value can be calculated as (coordinate - minimum value) / (maximum value - minimum value). The artificial neural network estimates 440 the physical property parameters of the fabric from the three-dimensional vertex coordinates corresponding to the three-dimensional boundary curve 430 of the fabric and outputs 450 them.

[0097] Among them, the input of the machine learning model 330 is the same as Figure 4 the input of the first step of Figure 4 and the output of the machine learning model 330 is the same as Figure 4 the output of the second step of Figure 3 The physical property parameters shown in 350 of Figure 4 are the physical property parameters commonly used for fabric simulation. In the embodiment, the physical property parameters shown in 450 of Figure 3 may be used, that is, a partial set equivalent to 350 of Figure 4 After the artificial neural network completes learning, by inputting the set of three-dimensional positions equivalent to the three-dimensional contour shape of a specific fabric into the artificial neural network, the physical property parameters defined in 450 of

[0098] The method for collecting learning data required for learning is as follows. Initially, hundreds of actual fabric physical property parameter data are given (provided by a preset method). In order to generate new random physical property parameters that have a statistically similar distribution to the given set of physical property parameters, a GMM model is learned from the initial data. A set of physical property parameters of the required quantity is generated from the GMM model, and virtual simulation is performed using the newly generated physical property parameters, thereby generating learning data for the artificial neural network model, that is, a three-dimensional contour model covering the corresponding fabric.

[0099] Figure 5 It is a drawing showing the correlation between the estimated physical property parameters and the three-dimensional contour shape of the fabric according to an embodiment. Figure 5 It shows a correlation matrix 510 representing the correlation between the physical property parameters of the fabric and the three-dimensional contour shape of the fabric according to an embodiment, and an autocorrelation matrix 530 between the physical property parameters of the fabric according to an embodiment.

[0100] In Figure 5 , the horizontal axis can display the x, y, and z coordinates of the first 30 sampling points of the three-dimensional contour shape of the fabric. In Figure 5 , each color in the color map shown in 510 can indicate that the correlation between the physical property parameters of the fabric and the three-dimensional contour shape of the fabric is not low.

[0101] Figure 5 The x-axis in the graph is the one-dimensional expansion of the x, y, and z coordinates. For example, the x, y, and z of the first three-dimensional position are the 0th, 1st, and 2nd data, the x, y, and z of the second three-dimensional position are the 3rd, 4th, and 5th data, and the x, y, and z of the kth three-dimensional position are the 3*(k - 1)+0, 3*(k - 1)+1, and 3*(k - 1)+2nd data. As Figure 5 shown, the 0th to 23rd data are the x, y, and z values of 8 sampling data. The correlation values between the horizontal axis data and the corresponding vertical axis data (tensile - weft, bending - warp, shear, bending - weft, bending - warp, bending - twill) are marked in color on the vertical axis. The closer the correlation value is to 1 or -1 (the darker the color), the greater the correlation between the two data.

[0102] Figure 6 It is a drawing showing the structure of the artificial neural network according to an embodiment. Figure 6 It shows the structure of an artificial neural network 600 including a first neural network 610 and a second neural network 630 according to an embodiment.

[0103] For example, the artificial neural network 600 may include a first neural network 610 that estimates physical property parameters related to fabric stiffness; and a second neural network 630 that estimates physical property parameters related to fabric bending. As an example, the physical property parameters related to fabric stiffness may include weft stretch stiffness, warp stretch stiffness, shear stiffness, etc. And as an example, the physical property parameters related to fabric bending may include weft bending stiffness, warp bending stiffness, and bending bias stiffness, etc.

[0104] At this time, the first neural network 610 and the second neural network 630 may be composed of independent fully connected neural network (FCNN) models. The foregoing formula 1 can define the fully connected neural network model. The input of both neural networks is the three-dimensional position of the contour sampling points. The stretching / bending learning two learning models are respectively trained with the same input, and the stretching / bending parameters are respectively output from the two models.

[0105] Referring to Figure 5 530, among the physical property parameters related to fabric stiffness ("first set") output by the first neural network 610 and the physical property parameters related to fabric bending ("second set") output by the second neural network 630, the physical property parameters belonging to the same set have a high autocorrelation, and conversely, the correlation between the two sets is relatively weak. The same set means that there is a great correlation between the stretching parameters and other stretching parameters (weft stretch / warp stretch / shear), and between the bending parameters and other bending parameters (weft bending / warp bending / twill bending). The number of hidden layers and nodes, activation types, and other parameters that respectively constitute the first neural network 610 and the second neural network 630 can be optimized through experiments.

[0106] In an exemplary embodiment, the neural network model 600 may be trained for 100 epochs with a 32 batch size.

[0107] Figure 7 is a flowchart showing a method for generating training data of an artificial neural network according to an embodiment. Referring to Figure 7, according to an embodiment, a training data generation device (hereinafter referred to as the "generation device") generates artificial neural network training data for estimating physical property parameters of a fabric through steps 710 to 740.

[0108] In step 710, the generation device collects a first quantity of physical property parameters of the fabric. Among them, as an example, the first quantity of physical property parameters may include 400 mechanical properties of the actual fabric materials. At this time, in addition to the physical property parameters of the fabric, the generation device may also collect the fabric density. A non-limiting example of the first quantity is 400.

[0109] In step 720, the generation device samples a second quantity of physical property parameters by using a generation model based on the physical property parameters collected in step 710, where the second quantity is greater than the first quantity. The generation model may randomly sample the physical property parameters by using a probability distribution such as a Gaussian Mixture Model (GMM). As an example, the second quantity is a quantity larger than 400.

[0110] According to an embodiment, in order to generate a sufficiently large set required for training, the generation device may randomly collect physical property parameters and density. Then, in order to collect information including a three-dimensional contour shape for each parameter set, the generation device simulates a cylinder test through a simulation system. Sampling without prior parameter space information has the risk of generating bias or invalid data sets, and invalid data will generate physically unrealizable or unsuitable parameters for clothes. In one embodiment, to prevent this risk, the generation model used may be generated based on the mechanical properties measured from actual fabrics.

[0111] The generation device may convert each set of mechanical properties into 6 physical property parameters and density. The following Table 1 shows the statistical values of 6 parameters and density measured from 400 actual fabrics.

[0112] Table 1

[0113] In Table 1, the mean, min, max, and standard deviation (std.dev.) may include a wide range of parameter spaces for training machine learning. Among them, 1,000,000 is the maximum limit of stiffness defined by a simulation system according to an embodiment.

[0114] As an example, the generating device can also use 400 actual fabric data sets to generate a Gaussian mixture model (GMM) with 7 variables (for example, 6 physical property parameters and density). In an exemplary embodiment, a Gaussian mixture model with an optimal number of clusters of 5 is used as the generating model for sampling the training data. This has two very important advantages. First, the generating device can easily avoid data deviation by collecting the same number of samples for each cluster. Second, the generating device can reduce the possibility of collecting parameters that are not suitable for clothing, that is, the set of invalid parameters.

[0115] In an exemplary embodiment, using the aforementioned Figure 2 The cylinder test described above was used to simulate and collect feature vectors for 100,000 samples (20,000 samples were collected for each cluster). At this time, the maximum particle distance of the virtual fabric can be 5 mm, and the time step is 0.33 seconds. The generating device stops the simulation when all the particles are in a stable state and discards the samples that cannot converge.

[0116] In step 730, the generating device simulates the covering shape of the corresponding fabric based on the physical property parameters collected in step 720, so as to obtain the three-dimensional contour shape of the corresponding fabric. At this time, as an example, the covering shape of the corresponding fabric is the covering shape when the corresponding fabric is placed on a three-dimensional geometric object.

[0117] In step 740, the generating device generates training data corresponding to the corresponding fabric based on the physical property parameters collected in step 720 and the three-dimensional contour shape of the corresponding fabric obtained in step 730.

[0118] According to an embodiment, the generating device uses the training data for the corresponding fabric generated in step 740 to train an artificial neural network so that it can estimate the physical property parameters of the corresponding fabric.

[0119] Figure 8 is a block diagram of an estimating device for physical property parameters according to an embodiment. Referring to Figure 8 , according to an embodiment, the estimating device 800 may include a communication interface 810, a processor 830, a memory 850, and an output device 870. The communication interface 810, the processor 830, the memory 850, and the output device 870 communicate with each other through a communication bus 805.

[0120] The communication interface 810 obtains information including the three-dimensional contour shape of the fabric placed on the three-dimensional geometric object. According to an embodiment, the communication interface 810 may also obtain the weight of the fabric or the fabric density.

[0121] The processor 830 applies the information to a pre-trained artificial neural network, thereby estimating the fabric physical property parameters for reproducing the covering shape of the three-dimensional clothing made of the fabric. Among them, the three-dimensional contour shape of the fabric and the covering shape of the three-dimensional clothing may be correlated with each other.

[0122] Alternatively, the communication interface 810 may receive an image capturing the three-dimensional contour shape of the fabric placed on the three-dimensional geometric object. At this time, the processor 830 generates a three-dimensional model including the three-dimensional contour shape of the fabric from the image. The processor 830 may extract the three-dimensional vertex coordinates corresponding to the three-dimensional contour shape of the fabric from the three-dimensional model.

[0123] The memory 850 may store the three-dimensional contour shape of the fabric obtained by using the communication interface 810, and / or the fabric density. And, the memory 850 may store the physical property parameters of the fabric estimated by the processor 830. The memory 850 may store various information generated during the processing of the above-mentioned processor 830. In addition, the memory 850 may store various data and programs, etc. The memory 850 may include a volatile memory or a non-volatile memory. The memory 850 has a large-capacity storage medium such as a hard disk, thereby storing various data.

[0124] The output device 870 may output the physical property parameters of the fabric estimated by the processor 830. The output device 870 may output the physical property parameters themselves, or may also output an image of covering a virtual clothing made of the fabric with the physical property parameters on a three-dimensional avatar. As an example, the output device 870 may be a display device, or a printing device that prints paper patterns on paper or fabric.

[0125] Alternatively, the processor 830 may execute based on Figures 1 to 7At least one method of description or an algorithm corresponding to at least one method. The processor 830 may be a data processing device embodied by hardware, where the hardware includes circuits having physical structures for performing desired operations. For example, the desired operations may include code or instructions included in a program. As an example, the processor 830 may be configured as a central processing unit (CPU), a graphics processing unit (GPU), or a neural network processing unit (NPU). For example, the estimation device 800 implemented by hardware may include a microprocessor, a central processing unit, a processor core, a multi-core processor, a multiprocessor, an application-specific integrated circuit (ASIC), or a field programmable gate array (FPGA).

[0126] The processor 830 may execute a program and control the estimation device 800. The program code executed by the processor 830 may be stored in the memory 850.

[0127] The method according to an embodiment is embodied in the form of program commands that can be executed by various computer means and is recorded in a computer-readable and writable medium. The computer-readable and writable medium can include program commands, data files, data structures, etc. either individually or in combination. The program instructions recorded in the medium can be instructions specially designed and constructed to implement the embodiment, or instructions that can be used by those of ordinary skill in the computer software field based on known knowledge. The computer-readable and writable recording medium can include magnetic media such as hard disks, floppy disks, and magnetic tapes; optical media such as CD-ROMs, DVDs, etc.; magneto-optical media such as floptical disks, and hardware devices specially configured to store and execute program commands such as read-only memories (ROMs), random access memories (RAMs), flash memories, etc. Examples of program instructions include not only machine language codes generated by compilers but also high-level language codes that can be executed by a computer by using interpreters, etc. To perform the operations of the embodiment, the hardware device can be configured in such a way that the operations are implemented by one or more software modules, and vice versa.

[0128] Software can include a computer program, code, instruction, or a combination of one or more of them, which can cause a processing device to operate in the desired manner, or, individually or collectively, command the processing device. To be interpreted by the processing device or to provide commands or data to the processing device, software and / or data can be permanently or temporarily embodied in any type of device, component, physical device, virtual equipment, computer storage medium or device, or transmitted signal wave. Software is distributed on computer systems connected by a network and can be stored or executed in a distributed manner. Software and data can be stored in one or more computer-readable and writable storage media.

[0129] In summary, the embodiments have been described with reference to a limited number of drawings, but it is obvious that various modifications can be made to the forms and details of these examples by those of ordinary skill in the art without departing from the spirit and scope of the claims and their equivalents. The embodiments described herein are for illustrative purposes only and not for purposes of limitation. The description of the features or aspects of each embodiment is equally applicable to similar features or aspects in other embodiments. Appropriate results can also be obtained when in a different order, and / or the components of the system, architecture, device, or circuit are combined in a different way, and / or replaced or supplemented by other components or their equivalents.

[0130] Accordingly, the scope of the present disclosure is not limited by the specific embodiments, but is defined by the claims and their equivalents, and all variations within the scope of the claims and their equivalents should be construed as being included in the present disclosure.

Claims

1. A method for estimating physical property parameters of a fabric, characterized in that, Comprising the following steps: Obtaining information on the three-dimensional contour shape of a fabric placed on a three-dimensional geometric object; Applying the information to a pre-trained artificial neural network to estimate the physical property parameters of the fabric for reproducing the covering shape of a three-dimensional garment made of the fabric; the artificial neural network is trained using the following training data based on the relationship between the physical property parameters of the fabric and the three-dimensional contour shape corresponding to the physical property parameters of the fabric, wherein the training data includes the physical property parameters of a second sample fabric generated by upsampling the physical property parameters of a first quantity of a first sample fabric using a generative model, and the second quantity is greater than the first quantity, and the three-dimensional contour shape is generated by physical simulation for each of the second quantity of physical property parameters; And Outputting the physical property parameters of the fabric.

2. The method for estimating the physical property parameters of a fabric according to claim 1, wherein The three-dimensional contour shape of the fabric and the covering shape of the three-dimensional garment are correlated with each other.

3. The method for estimating the physical property parameters of a fabric according to claim 1, wherein The step of obtaining the information includes the following steps: Receiving an image capturing the three-dimensional contour shape of the fabric placed on the three-dimensional geometric object; Generating a three-dimensional model including the three-dimensional contour shape of the fabric from the image; and Extracting three-dimensional vertex coordinates corresponding to the three-dimensional contour shape of the fabric from the three-dimensional model.

4. The method for estimating the physical property parameters of a fabric according to claim 1, wherein The step of obtaining the information includes the following steps: Obtaining three-dimensional vertex coordinates corresponding to the three-dimensional contour shape of the fabric placed on the three-dimensional geometric object.

5. The method for estimating the physical property parameters of a fabric according to claim 4, wherein The step of obtaining the three-dimensional vertex coordinates includes the following steps: Sampling the three-dimensional vertices corresponding to the three-dimensional contour shape from a depth image or a three-dimensional scan image including the three-dimensional contour shape of the fabric placed on the three-dimensional geometric object; and Obtaining the three-dimensional vertex coordinates of the sampling.

6. The method for estimating the physical property parameters of a fabric according to claim 1, wherein The information further includes the density of the fabric.

7. The method for estimating the physical property parameters of a fabric according to claim 1, wherein The fabric includes at least one of a natural fiber fabric, a synthetic fiber fabric, and a blended spun fabric including cotton, linen, wool, polyester fiber, nylon, and spandex.

8. The method for estimating the physical property parameters of a fabric according to claim 1, wherein At least a part of the region of the fabric is placed on the three-dimensional geometric object and supported by the three-dimensional geometric object, and the remaining part of the region is not supported by the three-dimensional geometric object and droops, The three-dimensional contour shape of the fabric is formed by the outer contour line of the remaining region of the fabric that is not supported by the three-dimensional geometric object and droops.

9. The method for estimating the physical property parameters of a fabric according to claim 1, wherein The physical property parameters of the fabric include at least one of the weft tensile stiffness, warp tensile stiffness, shear stiffness, weft bending stiffness, warp bending stiffness, and bending twill stiffness of the fabric.

10. The method for estimating the physical property parameters of a fabric according to claim 1, wherein the artificial neural network is trained with training data, and the training data is generated based on the physical property parameters of the fabric randomly sampled according to the probability distribution of the Gaussian mixture model.

11. The method for estimating the physical property parameters of a fabric according to claim 1, wherein the artificial neural network includes a first neural network for estimating the physical property parameters related to the stiffness of the fabric; and a second neural network for estimating the physical property parameters related to the bending of the fabric.

12. The method for estimating the physical property parameters of a fabric according to claim 11, wherein the first neural network and the second neural network each include a fully connected neural network model.

13. The method for estimating the physical property parameters of a fabric according to claim 1, wherein the step of outputting the physical property parameters of the fabric includes the following steps: applying the physical property parameters of the fabric to the three-dimensional garment; and displaying the result covering the three-dimensional garment.

14. A method for generating artificial neural network training data, wherein it includes the following steps: collecting the physical property parameters of a first number of first different sample fabrics; using a generation model to upsample the physical property parameters of the first number to generate the physical property parameters of a second number of second different sample fabrics, the second number being greater than the first number; performing a simulation to generate the three-dimensional contour shape of the second different sample fabrics corresponding to the physical property parameters of the second number; and generating training data, the training data including the generated three-dimensional contour shape and the physical property parameters of the second number, and the three-dimensional contour shape is generated by physical simulation for each of the physical property parameters of the second number.

15. The method for generating training data according to claim 14, wherein the generation model randomly upsamples the physical property parameters according to the probability distribution of the Gaussian mixture model.

16. The method for generating training data according to claim 14, wherein using the training data corresponding to the respective fabrics to train the artificial neural network to estimate the physical property parameters of the respective fabrics.

17. A computer program stored on a computer-readable storage medium, wherein it is combined with hardware to execute the method according to claim 1.

18. An apparatus for estimating the physical property parameters of a fabric, wherein it includes: a communication interface that acquires information including the three-dimensional contour shape of the fabric placed on a three-dimensional geometric object; A processor that applies the information to a pre-trained artificial neural network to estimate the fabric physical property parameters for reproducing the covering shape of a three-dimensional garment made of the fabric; the artificial neural network is trained using the following training data based on the relationship between the physical property parameters of the fabric and the three-dimensional contour shapes corresponding to the physical property parameters of the fabric, wherein the training data includes the physical property parameters of a second sample fabric generated by upsampling the physical property parameters of a first quantity of a first sample fabric using a generative model, and the second quantity is greater than the first quantity, and the three-dimensional contour shapes are generated by physical simulation for each of the second quantity of physical property parameters; and An output device that outputs the fabric physical property parameters.

19. The estimating device for fabric physical property parameters according to claim 18, wherein The three-dimensional contour shape of the fabric and the covering shape of the three-dimensional garment are correlated with each other.

20. The estimating device for fabric physical property parameters according to claim 18, wherein The communication interface receives an image capturing the three-dimensional contour shape of the fabric placed on a three-dimensional geometric object, The processor generates a three-dimensional model including the three-dimensional contour shape of the fabric from the image and extracts three-dimensional vertex coordinates corresponding to the three-dimensional contour shape of the fabric from the three-dimensional model.

Images