Learning-based animation of clothing for virtual try-on

Abstract

A learning-based clothing animation method and system for highly efficient virtual try-on simulations is provided. Given a garment, the system preprocess a rich database of physically-based dressed character simulations, for multiple body shapes and animations. Then, using a database, the system trains a learning-based model of cloth drape and wrinkles, as a function of body shape and dynamics. A model according to embodiments separates global garment fit, due to body shape, from local garment wrinkles, due to both pose dynamics and body shape. A recurrent neural network is provided to regress garment wrinkles, and the system achieves highly plausible nonlinear effects, in contrast to the blending artifacts suffered by previous methods.

Claims

1. A computer-implemented method for generating a digital representation of clothing on a body, the body modeled by a body template mesh having body parameters, including a shape parameter and a pose parameter, the method comprising the steps of: obtaining a clothing template mesh for the clothing, the clothing template mesh including a set of vertices and a garment parameter associated with a given size; subjecting the clothing template mesh to a non-linear regressor trained on a plurality of examples of deformations in training garment models due to a plurality of training body model parameters, the plurality of training body model parameters including a plurality of shape parameters and a plurality of pose parameters, the non-linear regressor configured to obtain deformations to the set of vertices of the clothing template mesh based, at least in part, on the body parameters of the body template mesh; and providing an output garment mesh of the clothing reflecting the deformations of the clothing in the given size by the body.

2. The computer-implemented method of claim 1, wherein the deformations correspond to a fit of the clothing on the body based on at least the shape parameter of the body template mesh.

3. The computer-implemented method of claim 2, further comprising: subjecting the clothing template mesh to a second non-linear regressor trained on a plurality of examples of wrinkle-based deformations in the training garment models due to the plurality of training body model parameters, the second linear regressor configured to obtain second deformations to the vertices of the clothing template mesh based on a body shape and the pose parameter of the body template mesh, wherein the second deformations correspond to wrinkles in the clothing based on a body pose and the shape parameter of the body template mesh.

4. The computer-implemented method of claim 1 wherein the non-linear regressor is implemented based on a single-hidden-layer multilayer perception (MLP) neural network.

5. The computer-implemented method of claim 1, wherein the deformations correspond to wrinkles in the clothing based on at least a body pose and the shape parameter of the body template mesh.

6. The computer-implemented method of claim 5, wherein the non-linear regressor is implemented based on a Recurrent Neural Network (RNN).

7. The computer-implemented method of claim 1, further comprising training the linear regressor with the plurality of examples of deformations in training garment models due to the plurality of training body model parameters, the linear regressor configured to encode a space model of variability in the clothing template mesh due to changes of at least body parameters.

8. The computer-implemented method of claim 7, wherein the variability in the clothing template mesh represent at least one of fit of the clothing on the body and wrinkles in the clothing due to the body.

9. A computer-implemented method for simulating clothing on a skeletal model, the method comprising: learning a body-shape based cloth deformation space model for a template clothing mesh from a first three-dimensional example data set of cloth mesh instances, said learnt cloth deformation space model encoding variability in said template cloth mesh based at least on changes in body shape; receiving as input a computer-based skeletal model defining a body shape and a pose and a computer model of an item of clothing, the computer model of the item of clothing defining a size of the item of clothing; computing an unposed cloth mesh representing the item of clothing fitted on the body shape defined by the computer-based skeletal model, the unposed cloth mesh preserving the size of the item of clothing; and generating a three-dimensional surface of the item of clothing in the size fitted on the body based on the unposed cloth mesh using a non-linear regressor trained on a plurality of examples including a plurality of shape parameters and a plurality of pose parameters.

10. The computer-implemented method of claim 9, wherein the computing the unposed cloth mesh comprises performing a non-linear regression of the computer model of the item of clothing based on learned instances of prior fitted unposed cloth meshes for a plurality of different training body shapes.

11. The computer-implemented method of claim 9, wherein the computing the unposed cloth mesh is based on a non-linear regressor implemented with a single-hidden-layer multilayer perception (MLP) neural network.

12. A computer-implemented method for simulating clothing on a skeletal model, the method comprising: receiving as input a computer-based skeletal model defining a body shape and a pose and one or more computer models of clothing items; computing an unposed cloth mesh for a given size of the clothing based on a non-linear regression based on learnt samples, the unposed cloth mesh fitting the one or more computer models of the clothing items to the body shape defined by the computer-based skeletal model; computing simulated garment wrinkles deformations based on the body shape and the pose defined by the computer-based skeletal model; and applying the computed garment wrinkles onto the unposed cloth mesh to generate a posed output cloth mesh in the give size of the clothing.

13. The computer-implemented method of claim 12, wherein the computing the simulated garment wrinkles deformations is based on a non-linear regressor implemented with a Recurrent Neural Network (RNN).

14. A system for generating a digital representation of clothing on a body, the body modeled by a body template mesh having body parameters, including a shape parameter and a pose parameter, the system comprising a processor and non-transitory computer readable media comprising instructions that when executed by the processor configures the processor to: obtain a clothing template mesh for the clothing, the clothing template mesh including a set of vertices and a garment parameter associated with a given size; subject the clothing template mesh to a non-linear regressor trained on a plurality of examples of deformations in training garment models due to a plurality of training body model parameters, the plurality of training body model parameters including a plurality of shape parameters and a plurality of pose parameters, the non-linear regressor configured to obtain deformations to the vertices of the clothing template mesh based, at least in part, on the body parameters of the body template mesh; and provide an output garment mesh of the clothing reflecting the deformations of the clothing in the given size by the body.

15. The system of claim 14, wherein the deformations correspond to a fit of the clothing on the body based on at least the shape parameter of the body template mesh.

16. The system of claim 14, wherein the non-linear regressor is implemented based on a single-hidden-layer multilayer perception (MLP) neural network.

17. The system of claim 14, wherein the deformations correspond to wrinkles in the clothing based on at least a body pose and the shape parameter of the body template mesh.

18. The system of claim 17, wherein the non-linear regressor is implemented based on a Recurrent Neural Network (RNN).

19. The system of claim 14, wherein the computer readable media further comprises instructions that when executed by the processor configures the processor to: subject the clothing template mesh to a second non-linear regressor trained on a plurality of examples of wrinkle-based deformations in the training garment models due to the plurality of training body model parameters, the second linear regressor configured to obtain second deformations to the vertices of the clothing template mesh based on a body shape and the pose parameter of the body template mesh, wherein the second deformations correspond to wrinkles in the clothing based on a body pose and the shape parameter of the body template mesh.

20. The system of claim 14, wherein the computer readable media further comprises instructions that when executed by the processor configures the processor to train the linear regressor with the plurality of examples of deformations in training garment models due to the plurality of training body model parameters, the linear regressor configured to encode a space model of variability in the clothing template mesh due to changes of at least body shape parameters.

21. The system of claim 20, wherein the variability in the clothing template mesh represent at least one of fit of the clothing on the body and wrinkles in the clothing due to the body.

22. The system of claim 14, wherein the processor is a distributed processor including a plurality of processing units communicatively coupled via a computer network.

23. A computer implemented method for obtaining a digital representation of clothing on a body, the body modeled by a body template mesh having body shape parameters, including a shape and a pose, the method comprising: providing an input comprising a clothing template mesh for the clothing, the clothing template mesh including a set of vertices and a garment parameter associated with a given size; and receiving by a virtual try-on application an output garment mesh of the clothing reflecting deformations of the clothing in the given size by the body, wherein the output garment mesh of the clothing reflecting the deformations is generated by subjecting the clothing template mesh to a non-linear regressor trained on a plurality of examples of deformations in training garment models due to a plurality of training body model parameters, the plurality of training body model parameters including a plurality of shape parameters and a plurality of pose parameters, the non-linear regressor configured to obtain the deformations to the set of vertices of the clothing template mesh based, at least in part, on the shape parameters of the body template mesh.

24. The method of claim 23, wherein the input comprises a size of the clothing.

25. The method of claim 23, wherein the input comprises a design of the garment.

26. The method of claim 23, wherein the deformations correspond to a fit of the clothing on the body based on the body parameters of the body template mesh.

27. The method of claim 23, wherein the deformations correspond to wrinkles in the clothing based on the pose and the shape of the body parameters of the body template mesh.

28. The method of claim 23, wherein the non-linear regressor is implemented based on a single-hidden-layer multilayer perception (MLP) neural network.

29. The method of claim 23, wherein the non-linear regressor is implemented based on a Recurrent Neural Network (RNN).

Description

BRIEF DESCRIPTION OF THE DRAWINGS

(1) FIG. 1A is a block diagram illustrating a data-driven cloth modeling system according to embodiments of the disclosure.

(2) FIG. 1B provides illustrations of body shapes and poses with a garment simulated according to embodiments of the disclosure.

(3) FIG. 2 is flow chart of a method for modeling clothing according to embodiments of the disclosure.

(4) FIG. 3 provides illustrations of results simulations for comparison according to aspects of embodiments of the disclosure.

(5) FIG. 4 is flow chart of a method for modeling clothing according to embodiments of the disclosure.

(6) FIG. 5 is flow chart of a method for modeling clothing according to embodiments of the disclosure.

(7) FIG. 6 provides illustrations of results of postprocessing functions according to aspects of embodiments of the disclosure.

(8) FIG. 7 provides illustrations of simulated garment fit deformations and corresponding body shapes according to embodiments of the disclosure.

(9) FIG. 8 provides illustrations of simulated garment wrinkle deformations and corresponding body poses according to embodiments of the disclosure.

(10) FIG. 9 provides illustrations of simulations of garment deformations for comparison according to embodiments of the disclosure.

(11) FIG. 10 provides illustrations of simulations of garment fit deformations for comparison according to embodiments of the disclosure.

(12) FIG. 11 provides illustrations of simulations of garment wrinkle deformations for comparison according to embodiments of the disclosure.

(13) FIG. 12 provides illustrations of frames of simulations of garment wrinkle deformations for comparison according to embodiments of the disclosure.

(14) FIG. 13 provides illustrations of frames of simulations of garment wrinkle deformations according to embodiments of the disclosure.

(15) The figures depict various example embodiments of the present disclosure for purposes of illustration only. One of ordinary skill in the art will readily recognize form the following discussion that other example embodiments based on alternative structures and methods may be implemented without departing from the principles of this disclosure and which are encompassed within the scope of this disclosure.

DETAILED DESCRIPTION

(16) The above and other needs are met by the disclosed methods, a non-transitory computer-readable storage medium storing executable code, and systems for 3D modeling of clothing and cloth items in computer applications such as virtual try-on but may used in other applications, including, for example, garment design and virtual modeling, motion capture applications, biomechanics and ergonomics design and simulation, education, business, virtual and augmented reality shopping, and entertainment applications, including animation and computer graphics for digital movies, interactive gaming and videos, human, animal, or character simulations, virtual and augmented reality applications, robotics, and the like. The Figures and the following description describe certain embodiments by way of illustration only. One of ordinary skill in the art will readily recognize from the following description that alternative embodiments of the structures and methods illustrated herein may be employed without departing from the principles described herein. Reference will now be made in detail to several embodiments, examples of which are illustrated in the accompanying figures.

(17) Referring now to FIG. 1A, a diagram of a learning-based data-driven system to animate the clothing of a virtual character according to embodiments is provided. According to one embodiment, a learning-based system 100 is provided to augment a skinning-based character animation with realistic nonlinear garment deformations. A runtime pipeline 120, takes as input a current body shape β 105a, obtained, for example, using motion capture or by editing a rigged character, avatar, or other representation of a body. The runtime pipeline 120 also takes as input frames of skeletal animation of the body shape β 105a at time-varying pose θ.sub.t 105b. For each frame of the skeletal animation, the system 100 produces a fitted garment and body mesh M.sub.t 109, including garment wrinkle effects due to body shape and pose in the given frame. The runtime pipeline 120 includes three main blocks: a garment fit regression module 121, a garment wrinkle regression module 122, and a post-processing module 123. According to embodiments, the garment fit regression module 121 applies global body-shape-dependent deformations to the input garment 106. In embodiments, the garment wrinkle regression module 122 predicts dynamic wrinkle deformations as a function of body shape β 105a and time-varying pose θ.sub.t 105b. Similarly, in embodiments, the post-processing module 123 includes a skinning module that combines the unposed fitted garment model with corresponding wrinkle deformations with a body pose θ.sub.t 105b for the current frame t to produce the deformed surface mesh (i.e., a fitted garment and body mesh) M.sub.t 109.

(18) Further, according to embodiments, a preprocessing stage 110 generates physics-based simulations of multiple animated bodies wearing the same garment that is used in the runtime regression modules. The preprocessing stage 110 includes a cloth simulation module 111. In one embodiment, cloth simulation module 111 is a physics-based cloth simulation engine, for example implemented using ARCSim [NSO12, NPO13]. The cloth simulation module 111 takes as input garment mesh template 101 and shape parameters 104 and pose parameters 103 from a template body mesh T.sub.b The cloth simulation module 111 provides ground-truth garment fit and garment wrinkle data Δ.sub.GT 102 as output. The cloth simulation module 111 allows for user-defined parameters to provide a simulated clothing item. For example, different knit types, different fabric compositions (e.g., cotton, polyester, wool, linen, silk, etc.) with single fabrics or mixes of fabrics in different ratios, and the like. In one embodiment, a T-shirt was simulated using an interlock knit with 60% cotton and 40% polyester.

(19) Referring now to FIG. 1B, an illustration of the input and outputs of a learning-based data-driven system to animate the clothing of a virtual character according to embodiments is provided. Given a garment (e.g., garment 101 in FIG. 1A or a template cloth mesh 106, which may be a low-pass filtered version of the garment), a learnt deformation model enables virtual try-on by bodies 109a-109d with different shapes and poses. The models according to embodiments of the dislcosure produce cloth animations 110 with realistic dynamic drape and wrinkles.

(20) According to embodiments, a deformed human body mesh M.sub.b may be determined by shape parameters β (e.g., the principal components of a database of body scans) and pose parameters θ (e.g., joint angles). Further, according to embodiments, a deformed garment mesh worn by the human body mesh M.sub.c 109 is provided. In these embodiments, a cloth mesh Sc(β, θ) 108 (in FIG. 1A) simulating the deformation and contact mechanics of the garment on a body mesh with shape β and pose θ is approximated using a data-driven model instead of a physics-based simulation.

(21) Based on the observation that most garments closely follow the deformations of the body, a clothing model according to embodiments is provided based in part on the Pose Space Deformation (PSD) literature [LCF00] and subsequent human body models [ASK*05, FCS15, LMR*15]. The body mesh is assumed to be deformed according to a rigged parametric human body model:
M.sub.b(β,θ)=W(T.sub.b(β,θ),β,θ,.sub.b),  (1)
where W (.Math.) is a skinning function, which deforms an unposed body mesh T.sub.b(β, θ)∈.sup.3×V.sup.b with V.sub.b vertices based on: first, the shape parameters β∈.sup.|β|, which define joint locations of an underlying skeleton; and second, the pose parameters θ∈.sup.|θ|, which are the joint angles to articulate the mesh according to a skinning weight matrix .sub.b. The unposed body mesh may be obtained additionally by deforming a template body mesh T.sub.b to account for body shape and pose-based surface corrections (See, e.g., [LMR*15]).

(22) Referring now to FIG. 2, according to embodiments, cloth deformations may be modeled following a similar overall process. For a given garment, a template cloth mesh 106 T.sub.c∈.sup.3×V.sup.c with V.sub.c vertices is provided 200 as input. The template cloth mesh 106 includes garment parameters for a given garment of a given size and design. The template cloth mesh 106 is subjected 201 to a non-linear garment fit regressor R.sub.G( ) that takes as input body shape parameters β 105a that are applied to the input cloth mesh 106. According to this embodiment, an unposed cloth mesh T.sub.c(β, θ) is output 202. Thus, a fitted garment model, deformed to match the shape parameters of a given body shape is thereby provided. Using this approach, overcomes the prior garment modeling methods that automatically resize the modeled garment to fit the different body shapes. For example, FIG. 3 illustrates the results of this approach. Referring to FIG. 3, illustrations 301a and 301b show the result of prior art method DRAPE [GRH*12] when the same garment 302 is worn by two avatars 303a and 303b with significantly different body shapes. DRAPE approximates the deformation of the garment by scaling it such that garment 302a is smaller in size than garment 302b adjusted in size to fit the target shape. This produces plausible but unrealistic results. In contrast, as illustrated by illustrations 302a and 302b, applying a method according to embodiments of this disclosure deforms the garment 305 in a realistic manner When the garment 305a is modeled on a smaller avatar body shape 306a, it provides a loose fit. While when the same garment model 305b is modeled over the larger avatar body shape 306b, the garment is stretched and wrinkled while maintaining the same size, showing a more realistic virtual try-on result.

(23) According to these embodiments, the fitted garment mesh T.sub.c(β, θ) may be used in post-processing to provide an un-posed (or default posed) skinned avatar result. Now referring to FIG. 4, according to other embodiments, a method for simulating full cloth deformation, for example for a virtual try-on application, is described. For a given garment, a template cloth mesh 106 T.sub.c∈.sup.3×V.sup.c with V.sub.c vertices is provided 400 as input. In addition, a template body mesh T.sub.b to account for body shape and pose-based surface corrections. The template cloth mesh 106 includes garment parameters for a given garment of a given size and design and the template body mesh also includes body shape 105a and pose 105b parameters as described above. The template cloth mesh 106 and body mesh are subjected 401 to a non-linear garment wrinkle regressor R.sub.L( ) that takes as input body shape β 105a and pose θ.sub.t 105b parameters that are applied to the input cloth mesh 106. According to this embodiment, a cloth mesh Sc(β, θ) 108 simulating the deformation and contact mechanics of the garment on a body mesh with shape β and pose θ is output. According to these embodiments, the deformed cloth mesh Sc(β, θ) 108 may be used in post-processing to provide a posed skinned avatar for each frame of a series of frames, for example in a virtual try-on application.

(24) Now referring to FIG. 5, according to embodiments, a method for modeling cloth deformations, for example for virtual try-on applications, is provided. According to these embodiments, a two-level strategy is used to learn the complex nonlinear deformations of clothing. A template cloth mesh 106 T.sub.c∈.sup.3×V.sup.c and a template body mesh T.sub.b are provided as input at 500. According to these embodiments, for a given garment, the template cloth mesh 106 T.sub.c∈.sup.3×V.sup.c with V.sub.c vertices is subjected at 501 to a non-linear garment fit regressor R.sub.G( ) that takes as input body shape parameters β 105a that are applied to the input cloth mesh 106. The resulting unposed cloth mesh T.sub.c(β, θ) is then deformed at 502 using the skinning function W (.Math.) to produce the full cloth deformation. To compute body shape-dependent garment fit and shape-and-pose-dependent garment wrinkles as corrective displacements to the template cloth mesh, to produce the unposed cloth mesh:
T.sub.c(β,θ)=T.sub.c+R.sub.G(β)+R.sub.L(β,θ),  (2)
where garment fit regressor R.sub.G( ) 501 and garment wrinkle regressor R.sub.L( ) 502 represent two nonlinear regressors, which take as input body shape parameters and shape and pose parameters, respectively.

(25) According to these embodiments, the final cloth skinning step 503 can be formally expressed as:
M.sub.c(β,θ)=W(T.sub.c(β,θ),β,θ,.sub.c)  (3)

(26) According to embodiments, a skinning weight matrix .sub.c is defined by projecting each vertex of the template cloth mesh onto the closest triangle of the template body mesh, and interpolating the body skinning weights .sub.b.

(27) Referring back to FIG. 1, for illustration, a pipeline according to these embodiments shows avatar 101 with the template body mesh T.sub.b wearing the template cloth mesh T.sub.c, and then the template cloth mesh 106 in isolation, with the addition of garment fit 107, with the addition of garment wrinkles 108, and the final deformation after the skinning step 109. According to another aspect of embodiments, one or more of the regressors, R.sub.G( ) and R.sub.L( ) as used in different embodiments, are trained with collision-free data. According to these embodiments, the data-driven model learns naturally to approximate contact interactions, but it does not guarantee collision-free cloth outputs. In particular, when the garments are tight, interpenetrations with the body can become apparent. Accordingly, in embodiments, the output step 202, 402, or 503, after the skinning process may also include additional postprocessing functions, including a step to correct cloth vertices that collide with the body, by pushing them outside their closest body primitive. An example of collision postprocessing is shown in FIG. 6.

(28) Garment Fit Regressor

(29) The learning-based cloth deformation models according to embodiments can represent corrective displacements on the unposed cloth state, as discussed above. Such displacements are produced by two distinct sources. On one hand, the shape of the body produces an overall deformation in the form of stretch or relaxation, caused by tight or oversized garments, respectively. According to embodiemnts, this deformation source is captured as a static global fit, determined by body shape alone. On the other hand, body dynamics produce additional global deformation and small-scale wrinkles. According to embodiments, this deformation source is captured as time-dependent displacements, determined by both body shape and motion. Different embodiments may provide for modeling that capture one or both sources of deformation depending on the intended application. In some embodiments, a higher accuracy is provided by training garment fit and garment wrinkles separately due to their static vs. dynamic nature.

(30) According to embodiments, static garment fit is defined as a vector of per-vertex displacements Δ.sub.G∈.sup.3×V.sup.c. These displacements represent the deviation between the cloth template mesh T.sub.c and a smoothed version of the simulated cloth worn by the unposed body. The ground-truth garment fit displacements can be formally defined as
Δ.sub.G.sup.GT=ρ(S.sub.c(β,θ))−T.sub.c,  (4)
where S.sub.c(β, θ)) represents a simulation of the garment on a body with shape β and pose θ=0, and ρ represents a smoothing operator.

(31) In these embodiments, to compute garment fit displacements 201/501 in the data-driven models, a nonlinear garment fit regressor R.sub.G:.sup.|β|.sub..fwdarw..sup.3×V.sup.c may be used, which takes as input the shape of the body β. In embodiments, the regressor Δ.sub.G=R.sub.G(β) is implemented using a single-hidden-layer multilayer perception (MLP) neural network. However, other non-linear regressors may be used in different embodiments. In these embodiments, the MLP network may be trained by minimizing the mean squared error between predicted displacements Δ.sub.G and ground-truth displacements Δ.sub.G.sup.GT. FIG. 7 illustrates this process with a visualization of garment fit regression. The original template mesh 701a has is globally deformed 701b and 702c to match the corresponding body shape 702b and 702c but lacks pose-dependent wrinkles.

(32) Garment Wrinkle Regressor

(33) According to embodiments, dynamic cloth deformations (e.g., wrinkles) are characterized as a vector of per-vertex displacements Δ.sub.L∈.sup.3×V.sup.c. These displacements represent the deviation between the simulated cloth worn by the moving body, S.sub.c(β, θ), and the template cloth mesh T.sub.c corrected with the global garment fit Δ.sub.G. In embodiments, this deviation in the body's rest pose, may be expressed by applying the inverse skinning transformation W.sup.−1(.Math.) to the simulated cloth. The ground-truth garment wrinkle displacements can be formally defined as
Δ.sub.L.sup.GT=W.sup.−1(S.sub.c(β,θ),(β,θ,.sub.c)−T.sub.c−Δ.sub.G.  (5)

(34) In these embodiments, to compute garment wrinkle displacements 401/502 in the data-driven models, a nonlinear garment wrinkle regressor R.sub.L:.sup.|β|+|θ|.fwdarw..sup.3×V.sup.c may be used, which takes as input the shape β and pose θ of the body. In contrast to the static garment fit, garment wrinkles exhibit dynamic, history-dependent deformations. In embodiments, these dynamic effects are accounted for by introducing recursion within the regressor. In some embodiments, the regressor θ.sub.L=R.sub.L(β, θ) may be implemented using a Recurrent Neural Network (RNN) based on Gated Recurrent Units (GRU) [CVMG*14], which has proven successful in modeling dynamic systems such as human pose prediction [MBR17]. GRU networks do not suffer from the well-known vanishing and exploding gradients common in vanilla RNNs [PMB13]. However, different recursive non-linear regressors may be used in different embodiments. According to embodiments, the GRU network may be trained by minimizing the mean squared error between predicted displacements ΔL and ground-truth displacements Δ.sub.L.sup.GT. FIG. 8 illustrates this process with a a visualization of garment wrinkle regression. In FIG. 8, a template garment 801a fit to the body shape 802a, as for example may be obtained with a garment fit regressor according to embodiments, is further deformed and enriched with pose-dependent dynamic wrinkles, 801b and 801c corresponding to the changes in pose of the avatar model 802b and 802c respectively.

(35) Training Data and Regressor Settings

(36) According to another aspect of embodiemnts of the disclosure, the generation of synthetic training sequences and the extraction of ground-truth data to train the regressor networks is described. In addition, network settings and hyperparameters that may be used in different embodiments are also provided for illustrative purposes.

(37) Dressed Character Animation Dataset

(38) To produce ground-truth data for the training of Garment Fit and Garment Wrinkle Regressors according to various embodiments, a novel dataset of dressed character animations with diverse motions and body shapes may be provided. According to one embodiment, a dataset was created using only one garment, but the same techniques can be applied to other garments or their combinations in other embodiments encompassed within the scope of this disclosure.

(39) According to various embodiments described herein, the use of a parametric human model is relied upon for the modeling of garments, for example in virtual try-on applications. In an exemplary embodiment, the SMPL [LMR*15] paramtric human model is used. In this embodiment, 17 training body shapes were selected as follows. For each of the 4 principal components of the shape parameters β, 4 samples were generated, leaving the rest of the parameters in β as 0. To these 16 body shapes, the nominal shape with β=0 is added.

(40) According to an exemplary embodiment, as animations, character motions from the CMU dataset [CMU], applied to the SMPL body model [VRM*17] were used. Specifically, 56 sequences containing 7 character motions were used, with 117 frames in total (at 30 fps, downsampled from the original CMU dataset of 120 fps). Each of the 56 sequences for each of the 17 body shapes were simulated wearing the same garment mesh (i.e., the exemplary T-shirt shown throughout the figures in this disclosure, which consists of 8, 710 triangles). It should be noted that any garment mesh may be used in different embodiments and even combinations of garments may be used within the scope of the methods and systems of this disclosure. The combination of garments may be modeled as a single mesh or each garment mesh may be processed separately and adjusted during post-processing as for example described with reference to FIG. 6.

(41) All simulations according to an exemplary embodiment were produced using the ARCSim physics-based cloth simulation engine [NSO12, NPO13], with remeshing turned off to preserve the topology of the garment mesh. Note however, that any other sutiable physics-based cloth simulation engine may used in different embodiemnts. According to this exemplary embodiment, ARCSim required setting several material parameters. In this example, since a T-shirt was simulated, an interlock knit with 60% cotton and 40% polyester was selected from a set of measured materials [WRO11]. In this exemplary embodiment, simulations were executed using a fixed time step of 3.33 ms, with the character animations running at 30 fps and interpolated to each time step. The output database with the simulation results from 1 out of every 10 time steps was stored to match the frame rate of the character animations. This produced a total of 120, 989 output frames of cloth deformation.

(42) In this exemplary embodiment, ARCSim required a valid collision-free initial state. To this end, the garment mesh was manually pre-position once on the template body mesh T.sub.b. However, in other embodiments, for example if collision-free is not a requirement, this step may be omitted. The simulation according to this embodiment was run to let the cloth relax first, and thus define the initial state for all subsequent simulations. In addition, a smoothing operator ρ(.Math.) may be applied to this initial state to obtain the template cloth mesh T.sub.c. Any other suitable approaches to obtain the template cloth mesh may be used in different embodiments.

(43) In embodiemnts, the generation of ground-truth garment fit data may require the simulation of the garment worn by unposed bodies of various shapes. In one exemplary embodiment, the shape parameters were incrementally interpolated from the template body mesh to the target shape, while simulating the garment from its collision-free initial state. Once the body reaches its target shape, the cloth is allowed to rest, and the ground-truth garment fit displacements Δ.sub.G.sup.GT may be computed according to Equation 4.

(44) Similarly, in one exemplary embodiment, to simulate a garment on animations with arbitrary pose and shape, both shape and pose parameters are incrementally interpolated from the template body mesh to the shape and initial pose of the animation. Then, the cloth is allowed to rest before starting the actual animation. In this embodiment, the simulations produce cloth meshes S.sub.c(β, θ), and from these the ground-truth garment wrinkle displacements Δ.sub.L.sup.GT may be computed according to Equation 5.

(45) Network Implementation and Training

(46) According to embodiments, neural networks for garment fit and garment wrinkle regressors may be implemented using Tensorflow [AAB*15]. In other embodiments, other suitable neural network implementations may be used. In one embodiment, an MLP network for garment fit regression contains a single hidden layer with 20 hidden neurons. In this embodiment, this configuration was sufficient to predict the global fit of the garment however multiple layers with a different number of neurons may be used in different embodiments. Similarly, in this exemplary embodiment, a GRU network for garment wrinkle regression was used with a single hidden layer but used instead 1500 hidden neurons. In this embodiment, this configuration was found to provide a good fit of the test data but, as one of skill in the art will recognize, other neural network configurations, with more layers or different numbers of neurons may be used. According to one exemplary embodiment, in both networks, dropout regularization was optionally applied to avoid overfitting the training data, which may not be implemented in different embodiments. In this exemplary embodiment, 20% of the hidden neurons were randomly disabled on each optimization step and the training data was shuffled at the beginning of each training epoch. Different techniques may be implemented in other embodiments within the scope of this disclosure as those of skill in the art will recognize.

(47) In this exemplary embodiment, the training process was implemented with a learning rate of 0.001, using the Adam optimizer, with 2000 epochs. For the garment fit MLP network, the ground-truth data from all 17 body shapes was used for training. For the garment wrinkle GRU network, the ground-truth data from 52 animation sequences was used for training, leaving 4 sequences for testing purposes. When training the GRU network, a batch size of 128 was used in this embodiment. Furthermore, to speed-up the training process of the GRU network in this embodiment, the training error was computed using Truncated Backpropagation Through Time (TBPTT), with a limit of 90 time steps.

(48) Runtime Performance

(49) According to one exemplary embodiment, a modeling system was implemented using an Intel Core i7-6700 CPU, with an Nvidia Titan X GPU with 32 GB of RAM. Table 1 shows average per-frame execution times of this exemplary embodiment including garment fit regression, garment wrinkle regression, and skinning, with and without collision postprocessing. For reference, simulation timings of a CPU-based implementation of full physics-based simulation using ARCSim is provided.

(50) TABLE-US-00001 TABLE 1 Per-frame execution times in milliseconds of exemplary embodiments, with and without collision postprocessing. Full physics-based simulation times are also provided for reference. Exemplary Exemplary Embodiment Embodiment ARCSim (w/o post- (w/ post- [NSO12] process) process) Mean 5635.4 ms 1.51 ms 4.01 ms Standard 2488.5 ms 0.28 ms 0.27 ms Deviation

(51) The low computational cost of the exemplary embodiment shows that the approach disclosed herein is suitable for interactive applications, such as virtual try-on applications. The memory footprint for the exemplary embodiment is as follows: 1.1 MB for the Garment Fit Regressor MLP, and 108.1 MB for the Garment Wrinkle Regressor GRU, both without any compression.

(52) Quantitative Evaluation

(53) Linear vs. nonlinear regression. FIG. 9 provides an illustrative comparison of the fitting quality of the nonlinear regression method according to the exemplary embodiment described above 903 vs. linear regression (implemented using a single-layer MLP neural network), on a training sequence. While our method retains the rich and history-dependent wrinkles, linear regression suffers smoothing and blending artifacts.

(54) Generalization to new body shapes and poses.

(55) The inventors have we quantitatively evaluated the generalization of the proposed approach to new shapes (i.e., not in the training set) and motion sequences with constant body shape but varying pose. The per-vertex mean error achieved on a static pose and a dynamic sequence were examined, as the body shape was changed over time. To provide a quantitative comparison to existing methods, the error suffered by cloth retargeting [LCT18, PMPHB17] was also examined. As discussed above, such retargeting methods scale the garment in a way analogous to the body to retain the garment's style. However, even if retargeting produces appealing results, it does not suit the purpose of virtual try-on, and produces larger error with respect to a physics-based simulation of the garment. The inventors observed that the error with retargeting increases as the shape deviates from the nominal shape, while it remains stable in the results obtained with an exemplary embodiment of the disclosed system.

(56) Similarly, the inventors analyzed the per-vertex mean error in the results achieved with an exemplary embodiment with 2 test motion sequences with constant body shape but varying pose. Cloth animation results according to this exemplary embodiment were obtained using the CMU sequences 01_01 and 55_27 [CMU], which were excluded from the training set, and exhibit complex motions including jumping, dancing and highly dynamic arm motions. The error suffered based on two baseline methods for cloth animation were also examined for comparison: (1) Linear Blend Skinning (LBS), which consists of applying the kinematic transformations of the underlying skeleton directly to the garment template mesh; and (2) a Linear Regressor (LR) that predicts cloth deformation directly as a function of pose, implemented using a single-layer MLP neural network. The results demonstrate that a two step approach according to one embodiment of the disclosure, with separate nonlinear regression of garment fit and garment wrinkles, outperforms the linear approach.

(57) Qualitative Evaluation: Generalization to new shapes and poses.

(58) FIG. 10 provides an illustration of the clothing deformations produced by the exemplary embodiment on a static pose while changing the body shape 1004, the mean body shape 1004a wearing an appropriately sized garment. As the ground-truth physics-based simulation row 1001 shows, as the body shape decreases towards 1004c, the garment becomes oversized for the decreasing body shape. As the body shape increases towards 1004e, the garment becomes undersized for the increasing body shape. The simulation results with retargeting techniques [LCT18, PMPHB17] shown in row 1002, scale the garment, and suffer noticeable artifacts away from the base shape. The results achieved with an exemplary embodiment according to this disclosure (shown in row 1003) successfully reproduces ground-truth deformations, including the overall drape (i.e., how the T-shirt slides up the belly due to the stretch caused by the increasingly fat character) and mid-scale wrinkles, even when changing the body shape beyond training values.

(59) Similarly, FIG. 11 illustrates the performance of one exemplary embodiment (row 1102) with a comparison to ClothCap [PMPHB17] (row 1101), a performance capture method that reconstructs clothing and shape independently, from 4D scans. A potential application of ClothCap is to retarget or transfer a captured garment to different shapes. However, as the illustrations in row 1101 show, retargeting lacks realism because cloth deformations are simply copied across different shapes. In contrast, the results from an exemplary embodiment in row 1102 show a realistic pose- and shape-dependent deformations.

(60) Similarly, FIG. 12 illustrates the performance of a system according to an exemplary embodiment showing results comparing ground-truth physics-based simulations (row 1201) with data-driven cloth deformations results on a test sequence based on the exemplary embodiment (row 1202). The overall fit and mid-scale wrinkles are successfully predicted using the data-driven model according to this exemplary embodiment, with a performance gain of three orders of magnitude. Similarly, in FIG. 13, more frames 1302a-f of a test sequence produced with the data-driven model according to this exemplary embodiment show realistic wrinkles in the belly area that appear when the avatar crouches.

(61) Conclusions

(62) We have presented a novel data-driven method for animation of clothing that enables efficient virtual try-on applications at over 250 fps. Given a garment template worn by a human model, our two-level regression scheme independently models two distinct sources of deformation: garment fit, due to body shape; and garment wrinkles, due to shape and pose. We have shown that this strategy, in combination with the ability of the regressors to represent nonlinearities and dynamics, allows our method to overcome the limitations of previous data-driven approaches.

(63) We believe our approach makes an important step towards bridging the gap between the accuracy and flexibility of physics-based simulation methods and the computational efficiency of data-driven methods. Nevertheless, there are a number of limitations that remain open for future work.

(64) Our method involves independent training per garment. Given the low computational cost of the regressor, it would be possible to animate multiple garments at the same time, for example with a dataset for each garment. Mix-and-match virtual try-on may be performed by training each possible combination of test garments.

(65) The foregoing description of the embodiments has been presented for the purpose of illustration; it is not intended to be exhaustive or to limit the patent rights to the precise forms disclosed. Persons skilled in the relevant art can appreciate that many modifications and variations are possible in light of the above disclosure.

(66) Some portions of this description describe the embodiments in terms of algorithms and symbolic representations of operations on information. These algorithmic descriptions and representations are commonly used by those skilled in the computer modeling arts to convey the substance of their work effectively to others skilled in the art. These operations, while described functionally, computationally, or logically, are understood to be implemented by computer programs or equivalent electrical circuits, microcode, or the like. Furthermore, it has also proven convenient at times, to refer to these arrangements of operations as modules, without loss of generality. The described operations and their associated modules may be embodied in software, firmware, hardware, or any combinations thereof.

(67) Any of the steps, operations, or processes described herein may be performed or implemented with one or more hardware or software modules, alone or in combination with other devices. In one embodiment, a software module is implemented with a computer program product comprising a computer-readable medium containing computer program code, which can be executed by a computer processor for performing any or all of the steps, operations, or processes described.

(68) Embodiments may also relate to an apparatus for performing the operations herein. This apparatus may be specially constructed for the required purposes, and/or it may comprise a general-purpose computing device selectively activated or reconfigured by a computer program stored in the computer. Such a computer program may be stored in a non-transitory, tangible computer readable storage medium, or any type of media suitable for storing electronic instructions, which may be coupled to a computer system bus. Furthermore, any computing systems referred to in the specification may include a single processor or may be architectures employing multiple processor designs for increased computing capability. Finally, the language used in the specification has been principally selected for readability and instructional purposes, and it may not have been selected to delineate or circumscribe the inventive subject matter. It is therefore intended that the scope of the patent rights be limited not by this detailed description, but rather by any claims that issue on an application based hereon. Accordingly, the disclosure of the embodiments is intended to be illustrative, but not limiting, of the scope of the patent rights, which is set forth in the following.

REFERENCES

(69) [AAB*15] ABADI M., AGARWAL A., BARHAM P., BREVDO E., CHEN Z., CITRO C., CORRADO G. S., DAVIS A., DEAN J., DEVIN M., GHE-MAWAT S., GOODFELLOW I., HARP A., IRVING G., ISARD M., JIA Y., JOZEFOWICZ R., KAISER L., KUDLUR M., LEVENBERG J., MANE D., MONGA R., MOORE S., MURRAY D., OLAH C., SCHUSTER M., SHLENS J., STEINER B., SUTSKEVER I., TALWAR K., TUCKER P., VANHOUCKE V., VASUDEVAN V., VIEGAS F., VINYALS O., WARDEN P., WATTENBERG M., WICKE M., YU Y., ZHENG X.: TensorFlow: Large-scale machine learning on heterogeneous systems, 2015. Software available from tensorflow.org.

(70) [ASK*05] ANGUELOV D., SRINIVASAN P., KOLLER D., THRUN S., RODGERS J., DAVIS J.: SCAPE: Shape Completion and Animation of People. ACM Transactions on Graphics (TOG) 24, 3 (2005), 408-416.

(71) [BMO*14] BENDER J., MULLER M., OTADUY M. A., TESCHNER M., MACKLIN M.: A Survey on Position-Based Simulation Methods in Computer Graphics. Computer Graphics Forum 33, 6 (2014).

(72) [BMWG07] BERGOU M., MATHUR S., WARDETZKY M., GRINSPUN E.: TRACKS: Toward Directable Thin Shells. ACM Transactions on Graphics (SIGGRAPH) 26, 3 (July 2007), 50:1-50:10.

(73) [BPS*08] BRADLEY D., POPA T., SHEFFER A., HEIDRICH W., BOUBEKEUR T.: Markerless garment capture. ACM Trans. Graphics (Proc. SIGGRAPH) 27,3 (2008), 99.

(74) [BSBC12] BROUET R., SHEFFER A., BOISSIEUX L., CANI M.-P.: De sign preserving garment transfer. ACM Trans. Graph. 31,4 (2012), 36:1-36:11.

(75) [CLMMO14] CIRIO G., LOPEZ-MORENO J., MIRAUT D., OTADUY M. A.: Yarn-level simulation of woven cloth. ACM Transactions on Graphics (Proc. of ACM SIGGRAPH Asia) 33,6 (2014).

(76) [CMU] CMU Graphics Lab Motion Capture Database. http://mocap.cs.cmu.edu/.

(77) [CO18] CASAS D., OTADUY M. A.: Learning nonlinear soft-tissue dynamics for interactive avatars. Proceedings of the ACM on Computer Graphics and Interactive Techniques 1,1 (May 2018).

(78) [CVMG*14] CHO K., VAN MERRIENBOER B., GULCEHRE C., BAH-DANAU D., BOUGARES F., SCHWENK H., BENGIO Y.: Learning phrase representations using RNN encoder-decoder for statistical machine translation.

(79) [dASTH10] DE AGUIAR E., SIGAL L., TREUILLE A., HODGINS J. K.: Stable spaces for real-time clothing. ACM Transactions on Graphics 29,4 (July 2010), 106:1-106:9.

(80) [FCS15] FENG A., CASAS D., SHAPIRO A.: Avatar reshaping and automatic rigging using a deformable model. In Proceedings of the 8th ACM SIGGRAPH Conference on Motion in Games (2015), pp. 57-64.

(81) [FTP16] FRATARCANGELI M., TIBALDO V., PELLACINI F.: Vivace: A practical gauss-seidel method for stable soft body dynamics ACM Trans. Graph. 35,6 (November 2016), 214:1-214:9.

(82) [GRH*12] GUAN P., REISS L., HIRSHBERG D. A., WEISS A., BLACK M. J.: DRAPE: DRessing Any PErson. ACM Transactions on Graphics 31, July (2012), Volume 31 Issue 4.

(83) [HFE13] HILSMANN A., FECHTELER P., EISERT P.: Pose space image based rendering. In Computer Graphics Forum (2013), vol. 32, pp. 265-274.

(84) [HSR13] HAUSWIESNER S., STRAKA M., REITMAYR G.: Virtual try-on through image-based rendering. IEEE Transactions on Visualization and Computer Graphics (TVCG) 19, 9 (2013), 1552-1565.

(85) [HTC*14] HAHN F., THOMASZEWSKI B., COROS S., SUMNER R. W., COLE F., MEYER M., DEROSE T., GROSS M.: Subspace Clothing Simulation Using Adaptive Bases. ACM Transactions on Graphics 33, 4 (July 2014), 105:1-105:9.

(86) [HWW*18] HAN X., WU Z., WU Z., YU R., DAVIS L. S.: VITON: An image-based virtual try-on network. In Proceedings of IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (2018).

(87) [JF03] JAMES D. L., FATAHALIAN K.: Precomputing Interactive Dynamic Deformable Scenes. ACM Transactions on Graphics (Proc. SIGGRAPH) 22, 3 (July 2003), 879-887.

(88) [JTST10] JAIN A., THORMAHLEN T., SEIDEL H.-P., THEOBALT C.: MovieReshape: Tracking and Reshaping of Humans in Videos. ACM Transactions on Graphics (Proc. SIGGRAPH Asia 2010) 29, 5 (2010).

(89) [KGBS11] KAVAN L., GERSZEWSKI D., BARGTEIL A. W., SLOAN P.-P.: Physics-inspired upsampling for cloth simulation in games. ACM Transactions on Graphics 30, 4 (2011)

(90) [KJM08] KALDOR J. M., JAMES D. L., MARSCHNER S.: Simulating knitted cloth at the yarn level. ACM Trans. Graph. 27, 3 (2008), 65:1-65:9.

(91) [KKN*13] KIM D., KOH W., NARAIN R., FATAHALIAN K., TREUILLE A., O'BRIEN J. F.: Near-exhaustive precomputation of secondary cloth effects. ACM Transactions on Graphics 32, 4 (2013)

(92) [KV08] KIM T.-Y., VENDROVSKY E.: Drivenshape: a data-driven approach for shape deformation. In Proceedings of the 2008 ACM SIG-GRAPH/Eurographics Symposium on Computer Animation (2008), Eurographics Association, pp. 49-55.

(93) [LCF00] LEWIS J. P., CORDNER M., FONG N.: Pose space deformation: a unified approach to shape interpolation and skeleton-driven deformation. In Annual Conference on Computer Graphics and Interactive Techniques (2000), pp. 165-172.

(94) [LCT18] LAHNER Z., CREMERS D., TUNG T.: Deepwrinkles: Accurate and realistic clothing modeling. In European Conference on Computer Vision (ECCV) (2018).

(95) [LMR*15] LOPER M., MAHMOOD N., ROMERO J., PONS-MOLL G., BLACK M. J.: SMPL: A skinned multi-person linear model. ACM Transactions on Graphics (Proc. SIGGRAPH Asia) 34,6 (October 2015), 248:1-248:16.

(96) [MBR17] MARTINEZ J., BLACK M. J., ROMERO J.: On human motion prediction using recurrent neural networks. In 2017 IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (2017), IEEE, pp. 4674-4683.

(97) [NH13] NEOPHYTOU A., HILTON A.: Shape and pose space deformation for subject specific animation. In 3DV (2013), pp. 334-341.

(98) [NPO13] NARAIN R., PFAFF T., O'BRIEN J. F.: Folding and crumpling adaptive sheets. ACM Transactions on Graphics 32,4 (July 2013), 51:1-51:8.

(99) [NSO12] NARAIN R., SAMII A., O'BRIEN J. F.: Adaptive anisotropic remeshing for cloth simulation. ACM Transactions on Graphics 31,6 (November 2012), 152:1-152:10.

(100) [OBLN17] OVERBY M., BROWN G. E., LI J., NARAIN R.: ADMM Projective Dynamics: Fast Simulation of Hyperelastic Models with Dynamic Constraints. IEEE Transactions on Visualization and Computer Graphics 23,10 (October 2017), 2222-2234.

(101) [OLL18] OH Y. J., LEE T. M., LEE I.-K.: Hierarchical cloth simulation using deep neural networks. In Proceedings of Computer Graphics International 2018 (2018), CGI 2018, pp. 139-146.

(102) [PMB13] PASCANU R., MIKOLOV T., BENGIO Y.: On the difficulty of training recurrent neural networks. In International Conference on Machine Learning (ICML) (2013), pp. 1310-1318.

(103) [PMPHB17] PONS-MOLL G., PUJADES S., HU S., BLACK M.: Cloth-Cap: Seamless 4D Clothing Capture and Retargeting. ACM Transactions on Graphics, (Proc. SIGGRAPH) 36,4 (2017).

(104) [RPC*10] ROHMER D., POPA T., CANI M.-P., HAHMANN S., SHEFFER A.: Animation wrinkling: augmenting coarse cloth simulations with realistic-looking wrinkles. In ACM Transactions on Graphics (TOG) (2010), vol. 29, ACM, p. 157.

(105) [SM06] SCHOLZ V., MAGNOR M.: Texture replacement of garments in monocular video sequences. In Eurographics Conference on Rendering Techniques (2006), pp. 305-312.

(106) [SRIC01] SLOAN P.-P. J., ROSE III C. F., COHEN M. F.: Shape by example. In Proceedings of Symposium on Interactive 3D graphics (2001), ACM, pp. 135-143.

(107) [SSIF09] SELLE A., SU J., IRVING G., FEDKIW R.: Robust high-resolution cloth using parallelism, history-based collisions, and accurate friction. IEEE Transactions on Visualization and Computer Graphics 15, 2 (March 2009), 339-350.

(108) [TWL*18] TANG M., WANG T., LIU Z., TONG R., MANOCHA D.: I-Cloth: Incremental collision handling for GPU-based interactive cloth simulation. ACM Transaction on Graphics (Proceedings of SIGGRAPH Asia) 37, 6 (November 2018), 204:1-10.

(109) [VRM*17] VAROL G., ROMERO J., MARTIN X., MAHMOOD N., BLACK M. J., LAPTEV I., SCHMID C.: Learning from synthetic humans. In Conference on Computer Vision and Pattern Recognition (2017).

(110) [WHRO10] WANG H., HECHT F., RAMAMOORTHI R., O'BRIEN J.: Example-based wrinkle synthesis for clothing animation. ACM Transactions on Graphics 29, 4 (2010)

(111) [WPLS18] WEIDNER N. J., PIDDINGTON K., LEVIN D. I. W., SUEDA S.: Eulerian-on-lagrangian cloth simulation. ACM Transactions on Graphines 37, 4 (2018), 50:1-50:11.

(112) [WRO11] WANG H., RAMAMOORTHI R., O'BRIEN J. F.: Data-driven elastic models for cloth: Modeling and measurement. ACM Transactions on Graphics (Proc. SIGGRAPH) 30, 4 (July 2011), 71:1-11.

(113) [XUC*14] XU W., UMENTANI N., CHAO Q., MAO J., JIN X., TONG X.: Sensitivity-optimized Rigging for Example-based Real-time Clothing Synthesis. ACM Transactions on Graphics 33, 4 (July 2014), 107:1-107:11.

(114) [YFHWW18] YANG J., FRANCO J.-S., HETROY-WHEELER F., WUHRER S.: Analyzing Clothing Layer Deformation Statistics of 3D Human Motions. In European Conference on Computer Vision (ECCV) (September 2018).

(115) [ZBO13] ZURDO J. S., BRITO J. P., OTADUY M. A.: Animating wrinkles by example on non-skinned cloth. IEEE Transactions on Visualization and Computer Graphics 19, 1 (2013), 149-158.

(116) [ZPBPM17] ZHANG C., PUJADES S., BLACK M. J., PONS-MOLL G.: Detailed, Accurate, Human Shape Estimation From Clothed 3D Scan Sequences. In The IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (July 2017).

(117) [ZSZ*12] ZHOU Z., SHU B., ZHUO S., DENG X., TAN P., LIN S.: Image-based clothes animation for virtual fitting. In SIGGRAPH Asia 2012 Technical Briefs (2012), ACM, p. 33.