3D Custom Fit Garment Design with Body Movement
Katja Wolff, Philipp Herholz, Verena Ziegler, Frauke Link, Nico Brügel, Olga Sorkine-Hornung
33D Custom Fit Garment Design with Body Movement
KATJA WOLFF and PHILIPP HERHOLZ,
ETH Zurich, Switzerland
VERENA ZIEGLER, FRAUKE LINK, and NICO BRÜGEL,
OpenDress GmbH, Germany
OLGA SORKINE-HORNUNG,
ETH Zurich, Switzerland (a) scanned poses (b) garment design (c) final rest shape (d) sewing pattern(e) simulated garment (f) physical garment
Fig. 1. Our method supports the creation of custom-fit garments that fit comfortably in a range of different poses. Our input consists of several scanned andregistered 3D poses of the intended wearer that the garment needs to accommodate (a). We provide six different tools to design and adjust the initial garmentshape for one of the poses, designated as the initial rest-shape (b). We then let the avatar move through the different poses by interpolating the pose meshesand utilizing a cloth simulation. Throughout, we adjust the rest shape of the garment to achieve the final garment shape (c), such that it accommodates all thedifferent poses (e). We use existing surface parameterization methods to create a sewing pattern of the shape (d). We sew a physical garment and verify it onthe scanned person (f). Our method is able to create garments for people who fall far outside the range of standard sizes, like this model, whose height is125 cm. Please refer to the accompanying video for a more detailed view of the dress.
The standardized sizes used in the garment industry do not cover the rangeof individual differences in body shape for most people, leading to ill-fittingclothes, high return rates and overproduction. Recent research efforts inboth industry and academia therefore focus on on-demand fabrication ofindividually fitting garments. We propose an interactive design tool forcreating custom-fit garments based on 3D body scans of the intended wearer.Our method explicitly incorporates transitions between various body posesto ensure a better fit and freedom of movement. The core of our methodfocuses on tools to create a 3D garment shape directly on an avatar withoutan underlying sewing pattern, and on the adjustment of that garment’s restshape while interpolating and moving through the different input poses. Wealternate between cloth simulation steps and rest shape adjustment stepsbased on stretch to achieve the final shape of the garment. At any step inthe real-time process, we allow for interactive changes to the garment. Oncethe garment shape is finalized for production, established techniques can
Authors’ addresses: Katja Wolff, [email protected]; Philipp Herholz, [email protected], ETH Zurich, Switzerland; Verena Ziegler, [email protected]; FraukeLink, [email protected]; Nico Brügel, [email protected], OpenDress GmbH,Germany; Olga Sorkine-Hornung, [email protected], ETH Zurich, Switzerland. be used to parameterize it into a 2D sewing pattern or transform it into aknitting pattern.CCS Concepts: •
Computing methodologies → Computer graphics ; Shape modeling ; Mesh geometry models ; Digital Garments .Additional Key Words and Phrases: computational fabrication, garmentmodeling, cloth simulation
The garment industry is a trillion-dollar, global industry that uses alarge amount of natural and human resources [Bick et al. 2018]. Ac-cording to recent research, of all annually manufactured garments,25% are never sold, and another 25% are sold but almost never worn[Morlet et al. 2017], meaning that nearly half of the produced itemsare imminently destined for landfill or incineration. One reason isthat for many people it can be challenging to find fitting clothes, asstandardized sizes often cannot account for individual differences,such as longer or shorter arms, asymmetries, body dimensions that a r X i v : . [ c s . G R ] F e b • Katja Wolff, Philipp Herholz, Verena Ziegler, Frauke Link, Nico Brügel, and Olga Sorkine-Hornung Fig. 2. We compare a standard shirt of size 43, modern fit (top row) withour custom-fit shirt (bottom row). Our shirt fits more tightly to the body(left), while allowing more comfort in a range of motions (middle and right).The standard shirt stretches and tightens uncomfortably when the arms areoutstretched to the front (middle). Lifting the arms leads to pulled downsleeves and uncomfortable stretch along the arms of the standard shirt(right). See Fig. 11 for more details of our custom-fit shirt. fall outside the commercially available range or even missing ex-tremities. Standard sizes available in stores vary over the globe; inGermany, 70% of women do not fit the commonly available stan-dard sizes [SizeGermany 2020]. Moreover, people who share thesame standard measurements might still vary drastically in theirbody shape. All these issues lead to high return rates of purchasedgarments. In recent years, on-demand, individualized fashion hasbecome a focus in research and industry as a response to these prob-lems. Still, the garment industry so far remains largely unchangedand depends on strenuous manual labor and mass-produced ready-to-wear garments. Fitting garments individually is currently a man-ual and expensive process done by professional tailors, who adjustexisting sewing patterns to the person’s size and features whileleaving the overall pattern design unchanged. This process did notsubstantially change for centuries. Even modern computer basedgarment design systems, like CLO 3D [CLO 2020], still rely on tra-ditional 2D sewing patterns as their central design space.We take a radically different design approach, which is liberatedfrom 2D sewing patterns and focuses on maximum fit and comfort under a range of individual motions for any kind of body shape. Ourmethod creates an optimized shape of the garment, and only in asecond step prepares it for production using existing techniques. Wecreate completely new sewing pattern designs via 2D parametriza-tion. Alternatively seamless knitting patterns can be generated basedon our designs, sidestepping the use of 2D sewing patterns com-pletely. We purposely step away from traditional sewing patterndesigns and symmetry, but nevertheless still allow the traditionalpositioning of seams if desired.We work with a variety of poses for each person: we scan eachpose using a commercial 3D scanner and use existing methods toregister a template body mesh to each scanned pose, such that weobtain pose meshes of matching connectivity. This provides us withhigher fidelity to the true shape of the user’s body than simplyusing a few standard measurements and a pre-set, symmetric avatar,and also helps eliminating measurement errors by the user. Foreach garment we use a select number of poses that this garmentneeds to accommodate. We provide a set of intuitive tools to de-sign the garment shape directly in 3D, sidestepping the necessityfor creating an underlying sewing pattern. We enable the drawingof garment boundaries directly on the avatar to create skintightclothing, but also allow for the addition of loose parts. An optionalpaintbrush tool can be used to add cloth in specified regions, andby defining a minimum distance of the garment to the body, we canexplicitly control comfort. The garment is then simulated using ex-isting cloth simulation techniques. We maintain a rest shape , whichis the garment without any forces applied and a simulation shape which represents the current simulation mesh undergoing stretch-ing, bending and shearing as it is deformed by the dynamic bodymesh. While smoothly transitioning between different poses, wecompute a stretch metric of the garment and adjust the garment restshape whenever the stretch exceeds a threshold. This process is fastand allows real-time interaction to adjust the garment at any pointin the process. The rest shape can then be used for production, e.g.,by applying existing methods to compute a distortion-minimizingsewing pattern. The resulting garments fit more tightly, while al-lowing a wider range of motions compared to garments of standardsizes (see Fig. 2).We make the following contributions: • The introduction of a toolset to easily create 3D garmentshapes on a 3D avatar; • The incorporation of varying body poses and body move-ments into the design optimization of custom-fit garments; • The formulation of an iterative garment adjustment algorithmbased on stretch.Using our software, we create a number of garments for peopleof widely varying stature and body type and demonstrate profes-sionally manufactured garments based on these designs in Figs. 1,2, 13-15, and in the accompanying video. We make our source codepublicly available to foster further research in this area.
Computational garment design has become a highly active researchfield in different scientific areas over the past years. We concentrate
D Custom Fit Garment Design with Body Movement • 3 on the most relevant works in relation to our contribution and groupthem according to their focus on garment design, fit and simulation.
Garment design.
Works on garment design focus mainly on pro-viding tools for automatically creating and manipulating the gar-ments and their underlying sewing patterns, such that designers caneasily and quickly explore design choices. Nayak and Padhye [2017]broadly survey the use of automation in garment manufacturing,including computer aided design. Early works focus on interac-tive design and modification pipelines, providing visual real timefeedback [Keckeisen et al. 2004; Volino et al. 2005]. Berthouzoz etal. [2013] scan and parse existing, traditionally published patternsand convert them into 3D garment models. Umetani et al. [2011]introduce a system for bidirectional interactive garment design thatallows to edit both the 2D pattern and the 3D garment shape, whilekeeping the correspondence between the two. To facilitate fabri-cation of computed patterns, Igarashi et al. [2008] automaticallygenerate necessary seam allowance for sewing. Several commercialCAD fashion design softwares are available nowadays, includingCLO 3D [2020] and Optitex [2020], which enable the digital designof sewing patterns and their draping, greatly accelerating the it-erative design process. However, they still follow the traditionaldesign workflow, which requires a professional garment designerwith experience in modifying 2D sewing patterns.A multitude of sketch based methods [Decaudin et al. 2006; Rob-son et al. 2011; Rose et al. 2007; Turquin et al. 2007] create 3Dgarment shapes from contours, boundary lines and seam lines thatare drawn on a digital model. Since garments are sewn from flatsheets of cloth, developable patches are automatically computedfor the sewing pattern. Incorporating advancements in machinelearning, Wang et al. [2018] use a data-driven approach to estimategarment shapes from a sketch of a desired fold pattern. Instead ofcreating a new garment, the approach by Li et al. [2018] enriches agarment with folds and pleats guided by sketches. In order to incor-porate the traditional workflow of pattern design while keeping theadvantages of working digitally, Wibowo et al. [2012] use a physicalreal-world mannequin as a guide for drawing with a specializedtool in 3D around it.Up until now, if a garment is meant to be sewn (instead of beingused solely on digital avatars), the design of the garment is based onan underlying sewing pattern created by a skilled professional whoensures fit and style. When creating a new custom-fit garment, wetake a radically different design approach, only focusing on the 3Dshape and fit of a garment first. The resulting shape can be directlyknit [Narayanan et al. 2019] or flattened with minimal distortion byexisting parametrization methods, such as [Sharp and Crane 2018].The resulting garments show a novel and intriguing style and differsignificantly from traditional sewing patterns. Similarly, Kwok et al.[2016] create styling curves directly on the avatar to create novelsewing pattern designs for sports garments, but in their work, thefit is simply assumed to exist and is not optimized.
Garment fitting.
A number of works address custom-fitting gar-ments for different body sizes and shapes. Early research focuses onfitting the 3D shape of a pre-designed template garment to an arbi-trarily sized avatar by creating feature correspondences between theavatar and the garment [Cordier et al. 2003; Meng et al. 2012; Wang et al. 2005]. The 2D sewing patterns are only created afterwardsthrough different parametrization methods [Decaudin et al. 2006;Meng et al. 2012; Wang et al. 2005]. The seams that define thesepieces are either sketched [Decaudin et al. 2006; Wang et al. 2005]or transferred from the initial input garment [Meng et al. 2012].Introducing style and fit criteria, like proportion, scale, shape andfit, allows Brouet et al. [2012] to grade existing sewing patternsfor largely different body sizes. All these methods suffer from theso called draping effect: Since the 2D sewing patterns are directlyparameterized pieces of the 3D garment, the resulting sewn gar-ment deforms further when draped. In contrast, we work with anunderlying rest shape of the garment, incorporate movement intothe shape adjustment of the garment and allow for an optimizedplacement of seams.More recent methods adjust for the draping effect. Bartle et al.[2016] allow users to mix existing garment designs and calculatethe sewing pattern inversely from the garment. Wang [2018] solvesthe garment shape and sewing pattern design as a single nonlinearoptimization problem. This method only allows for small changesin avatar size, which are created by applying measurements todeform a base mesh; the method is demonstrated by sewing standardsized garments. In contrast, we incorporate 3D scans of extremelydifferent body shapes to create our garments. Unlike both abovemethods, our work does not rely on pre-designed sewing patternsand incorporates movement of the body. By working with a 3D restshape of the garment and optimizing the placement of seams, wealso limit the draping effect, allowing us to to incorporate largedeformations and asymmetries for custom-made apparel.The work by Montes et al. [2020] explores the automatic gen-eration of new sewing patterns. They embed the cloth as a two-dimensional elastic membrane in the surface of an elastic bodymesh and strongly adjust an initial 2D pattern to create skintightclothing. Their method allows to optimize the layout for multipleposes simultaneously. We also create novel sewing patterns for mul-tiple poses, but in contrast, we do not rely on an initial 2D pattern,and our garment does not need to be skintight and allows wrinkles.Notably few previous works demonstrate fit with actual sewngarments, and when they do, it is mostly on dolls [Bartle et al. 2016;Brouet et al. 2012; Decaudin et al. 2006]; with the exception of Wang[2018] which demonstrates the fit of the garments on real humans,as we do.
Cloth simulation.
The simulation of cloth has been of interest fordecades with different foci, approaches and models, depending onwhether an application requires speed or physical accuracy and thetype of cloth (knit or woven). Cloth elasticity can be simulated basedon fast and simple spring models [Choi and Ko 2005; Liu et al. 2013],on continuum models [Baraff and Witkin 1998; Narain et al. 2012],which can work well for woven cloth, or the individual yarns [Cirioet al. 2014; Kaldor et al. 2010], which is very useful in modelingintricate knit patterns or a snag of a single thread. Research hasalso covered aspects such as collision handling [Bridson et al. 2005;Tang et al. 2018], measurement of cloth elasticity [Miguel et al. 2012;Wang et al. 2011], inextensibility [Goldenthal et al. 2007] or plasticdeformations [Jung et al. 2016]. Though our framework is relativelyindependent of the particular choice of the simulation method, we • Katja Wolff, Philipp Herholz, Verena Ziegler, Frauke Link, Nico Brügel, and Olga Sorkine-Hornung do substantially rely on cloth simulation and base our method onthe model by Baraff and Witkin [1998]. The different componentsof the energy, such as stretch, bend or shear, can be individuallyadjusted [Bergou et al. 2006; Tamstorf and Grinspun 2013].
Human body modeling.
Our method is targeted at creating gar-ments for real people. To capture details of individual body shapes,we therefore rely on the large body of research on object scanningand body shape modeling, as well as commercially available 3Dscanners. Several methods for fitting a parameterized, articulatedhuman body model to 3D scans exist, such as the multi-person linearmodel SMPL, based on skinning and blend shapes and learned fromthousands of 3D body scans [Loper et al. 2015], the more recentSTAR model [Osman et al. 2020], as well as dynamic models likeDyna [Pons-Moll et al. 2015]. The FAUST dataset [Bogo et al. 2014]allows to evaluate and compare body models, and we use severalavatars of this dataset to demonstrate our method. Our frameworkis independent of the chosen scanning and registration technique,which can be replaced once even more accurate future methodsbecome available.
We start by describing the process to capture and register the 3Dbody poses, as well as the interpolation of scanned poses to achievesmooth motions, using existing techniques (Sec. 3.1). In order tocreate a large variety of garments directly on the 3D avatar withoutrequiring underlying sewing patterns, we introduce six interactivedesign tools (Sec. 3.2). Our system adjusts the initial garment shapeby moving through all poses, utilizing a cloth simulation (Sec. 3.3)and iteratively adjusting the garment shape (Sec. 3.4). The finalshape can then be fabricated (Sec. 3.5).
Acquisition.
In order to incorporate the movement of the humanbody in the garment optimization, we capture different static posesof the individual. The amount and nature of these poses dependson the desired garment. A comfortably fitting garment designed forsports requires a wider range of poses than a tight fitting cocktaildress. A long-sleeved shirt might need two poses to capture bentand outstretched elbows, as opposed to a short sleeved garment.The specific technique used to acquire the body poses can be chosenfreely, as long as all poses are represented as meshes with the samenumber of vertices and connectivity in full correspondence acrossthe different poses. We use a
Structure Sensor [Occipital 2020] toscan each pose. The resulting mesh can include background objects,which we remove, as well as holes and artifacts. Two cleaned upmeshes are shown in Fig. 3 (a). In order to create evenly meshedposes in full correspondence, we use
Meshcapade [Meshcapade 2020]which employs the SMPL model [Loper et al. 2015] (Fig. 3 (b)). Thesetechnologies can be easily substituted by future developments.
Interpolation.
During the later garment simulation, we want tosmoothly transition between the captured poses in order to opti-mize the garment for a range of motions. Given the captured posemeshes with corresponding vertices and faces, we can readily in-terpolate these poses. Even though Meshcapade provides skeletons (a) (b) (c)
Fig. 3. Using a 3D scanner, we obtain a mesh for each pose (a) and weregister all the meshes based on the SMPL model [Loper et al. 2015] (b). Tocreate smooth motions between poses, we compute interpolated in-betweenmeshes (c). with each registered pose, we do not use those for interpolation,for two reasons: (i) We want our algorithm to be independent ofthe registration technique and work for methods that do not supplya skeleton, and (ii) we wish to be able to incorporate changes inthe body shape that are not necessarily captured by skeletal meshdeformations (skinning), such as movement of fat and tissues, orvolumetric changes for the same pose (e.g., growing belly in preg-nancy). Therefore, we directly interpolate the pose meshes using anonlinear morphing technique akin to deformation transfer [Sum-ner and Popović 2004] and Poisson shape interpolation [Xu et al.2006]. In order to get a unique solution, we need to constrain theposition of a single vertex. We choose to constrain the vertex thatmoves least between both poses to a linearly interpolated position.We evenly sample between both poses with a predefined numberof interpolation steps (see Fig. 3 (c)), though the number can bechosen higher or lower, depending on how much the scanned posesdiffer. Starting with any of the captured poses, we use it to design thegarment directly in 3D. We propose a set of simple put powerfultools that allow to explore the complex space of garment designs:(1)
Boundary tool to draw garment boundaries on the avatar.(2)
Extension tool for loose fitting parts.(3)
Paint tool to add cloth in specific areas.(4)
Comfort tool to set a minimum distance to the body.(5)
Pinning tool for garment vertices.(6)
Seam tool to predefine garment seams.
Boundary tool.
We allow the user to draw closed loop boundariesdirectly on the avatar by consecutively clicking points on the mesh,which we connect by shortest edge paths. After the loop is closed,we perform a smoothing operation, which creates a polyline thatis defined on the mesh through barycentric coordinates. Since thedifferent poses have a corresponding mesh structure, the definedboundaries are valid for all poses (Fig. 4 (a)-(c)). We can create agarment from these boundaries by specifying a boundary-enclosedregion on the avatar. To define the initial garment rest shape weduplicate the enclosed submesh and remesh it to a desired resolution,
D Custom Fit Garment Design with Body Movement • 5 (a) (b) (c) (d) (e) (f) (g) (h)
Fig. 4. Our toolset allows creating garment boundaries on the avatar mesh by connecting clicked vertices on a shortest path (a). These boundaries are thensmoothed (b) and expressed in barycentric coordinates on the triangle mesh, which allows us to transfer them to different poses (c). Clicking an area on theavatar mesh creates a garment mesh enclosed by the created boundaries (d). Garments can be extended with loose-fitting parts at a chosen boundary (e). Evenduring simulation (f), changes to the garment are possible. We allow painting areas where cloth should be added (g,h). All interactions are marked by a cursor.
Fig. 5. Different offsets for comfort control. The top row shows the simulatedgarment shape after no offset or an offset of 1 or 3 cm was applied. Thebottom row shows the adjusted rest shape. generating triangles with similar area (Fig. 4 (d)). This tool is usedto create a new garment shape, whereas all following tools are usedto edit existing garment shapes.
Extension tool.
In order to allow for the designof dresses, skirts, tops with wide sleeves and sim-ilar features that are not skintight, we allow toextend a garment at a chosen boundary. An axisis created from this boundary by calculating itscenter of mass c and its vector area a . The vectorarea is a vector that is well defined by the vertexpositions of the boundary and points in the direction that maxi-mizes the enclosed area when projected onto a plane, and thereforecan also be calculated for non-planar boundary loops. The chosenboundary is duplicated, translated and scaled along this axis, suchthat it lies on an additionally specified point, which can be definedwith a mouse click by the user. The duplicated boundary is thenconnected to the existing garment mesh. A remeshing operationensures even meshing. See Fig. 4 (e). (a) (b) (c) Fig. 6. When boundaries are not pinned to the avatar while changing poses,the large stretch forces can result in widened hems, such as a wide neckline(b). Pinned boundaries ensure that the garment stays in place and the stretchforces result in enlarged cloth area instead (here in the shoulder regions) toallow movement of the specific body parts (c).
Paint tool.
This tool can be used to specify areas that need to beenlarged. The intensity of the color per triangle defines a scalingfactor. The garment mesh is then adjusted accordingly by applyingthe methods described in Sec. 3.4. See Fig. 4 (g)-(h).
Comfort tool.
We allow the user to set a minimum offset distancebetween the garment and the body. When the garment is simulated(Sec. 3.3), a simple collision detection with the body pushes thegarment vertices away from the body by the set distance and createsa small stretch everywhere. The rest shape adjustment step (Sec. 3.4)then automatically adjusts the garment to counteract this stretch.A small offset is useful to allow for thick textiles or to compensatesmall errors introduced through the 3D scanning process. Largeoffsets can also be used as a design choice. See Fig. 5.
Pinning tool.
We allow the user to pin selected vertices of thegarment mesh to the body during the cloth simulation, specificallywhole garment boundaries. The pinning is implemented as an ad-ditional constraint (Sec. 3.3). This tool is especially useful when • Katja Wolff, Philipp Herholz, Verena Ziegler, Frauke Link, Nico Brügel, and Olga Sorkine-Hornung the avatar moves through different poses. As illustrated in Fig. 6, agarment’s neckline might stretch when the avatar lowers its arms.Pinning the neckline prevents this enlargement and keeps the neck-line of the rest shape in place.
Seam tool.
After the final garment shape is computed by movingthrough several poses and ranges of motion, we allow to optionallypredefine seams by creating boundaries, similar to the
Boundary tool ,but directly on the garment rest shape. This is not necessary, butmight be desired to give the garment a certain look, e.g., by definingtraditional shoulder seams for men’s shirts (Fig. 10, 11, 13) or bycreating seams between regions with different textiles (Fig. 14).To design a garment, we can choose an arbitrary pose as a startingpoint, as long as this pose has no self-overlaps. Poses with arms heldat the sides often have self-overlaps below the shoulders, whichleads to self-intersections in the garment created from this pose.Such self-overlaps often do not untangle during the cloth simulation,even if the arm is lifted or the cloth stretched. The choice of poseused for designing the initial shape of the garment influences thefinal shape, and one can achieve different designs, as we discussfurther in Sec. 4. Additionally to the tools above, our systems allowsthe user to load any existing garment mesh to work with.
After designing the initial shape of a garment, we simulate its drap-ing on the avatar with respect to its 3D rest shape. In this section,this garment rest shape is equal to the initial shape design and doesnot change. As soon as the avatar moves through different posesthough, we allow the rest shape to deform as well to adjust to largestretch forces, as described in Sec. 3.4.We use a standard cloth model that treats bending, stretching andshearing models separately [Baraff and Witkin 1998; Bridson et al.2005]. Similarly to previous work [Umetani et al. 2011], we use theisometric bending model [Bergou et al. 2006], which is efficient interms of computation time, since it has a constant positive semi-definite energy Hessian. This model penalizes out of plane anglesbetween neighbouring triangles. Even though we are working witha 3D rest shape where this is not the case, most materials usedto fabricate the final garment are flat, motivating our choice ofbending energy. Alternatively, a more computationally involvedbending energy can be used, such as penalizing the deviation ofangles from the rest shape, based on the work by Tamstorf et al.[2013].As has been described in [Umetani et al. 2011], compression in thecloth can induce instabilities in the usual St. Venant-Kirchhoff (StVK)constant strain triangle (CST) due to an indefinite force Jacobian[Volino et al. 2009]. Therefore we use a stabilized StVK CST byeliminating negative eigenvalues in the Jacobian when the elementis under compression, similar to [Teran et al. 2005].Whenever the user specifies boundaries that are to be pinnedonto the body with the
Pinning tool , we add additional constraintsto the cloth simulation. We find that incorporating the constraintsvia mass modification, as described in [Baraff and Witkin 1998],introduces large stretch forces in the triangles adjacent to the fixedboundaries, which interferes with the later rest shape adjustment (a) (b) (c) (d) (e)
Fig. 7. Starting from a shirt with a rest shape that was designed for raisedarms (c) and therefore does not display any significant stretch in that pose(a), stretches once the arms are lowered (b). Red areas signify high stretch,as seen on top of the shoulders, whereas green areas signify compression,as seen in the armpits. Once we deform the rest shape according to section3.4 (e), the stretch is reduced significantly (d). In this example, we pinnedthe neckline to the avatar with the
Pinning tool that is based on stretch. Instead, we introduce penalty forces, keep-ing the garment boundary vertices close to the defined boundaries.The corresponding condition 𝐶 is defined as: 𝐶 ( X ) = ∥ x 𝑖 − b 𝑖 ∥ , (1)where X ∈ R 𝑛 × contains the coordinates of the simulated garmentmesh, x 𝑖 are the vertex positions of the pinned boundaries and b 𝑖 are the corresponding boundary vertex positions on the avatar.Right after garment creation, we create these pairs of vertices byassociating the garment boundary vertices x 𝑖 with their closestpoint on the avatar mesh b 𝑖 , which we represent in barycentriccoordinates. As the avatar moves, this target position changes andis also offset in the normal direction from the avatar mesh by the Comfort control distance.We use a simple collision detection of the garment with the avatarby calculating a signed distance for each garment vertex to the avatarmesh. We resolve collisions by moving the garment vertices alongthe normal, which is calculated at the closest point on the avatarmesh. We incorporate the
Comfort tool by moving the vertex bythe specified offset away from the avatar mesh. Since the collisiondetection operates solely on the vertices, a dense enough meshresolution of the garment is essential to avoid scenarios where armsslip through the garment mesh.Many avatar poses can exhibit self-intersections, for examplebelow the armpits when the arms are held close to the body. Weobserve that these self-intersections can induce instability in thecloth simulation, especially when there is an offset defined betweenthe garment and the avatar. This is due to collisions not being ade-quately resolved. We deal with these problem regions by assumingthat ideally, the cloth would get stuck between the two opposingsurfaces in these areas. Therefore, we detect all vertices affectedby self-intersection regions during the collision detection step. Wethen mark these as fixed during the simulation and add additionalconstraint forces equivalent to Eq. (1).
D Custom Fit Garment Design with Body Movement • 7
Our system features automatic adjustment operations that adapt thegarment to fit a specific dynamic avatar, which is moving througha set of poses (see Fig. 7). This way we can ensure a good fit formany everyday situations that would otherwise introduce signifi-cant stretch. This can be seen, for example, in a long-sleeved shirtin Fig. 10, designed for a T-pose (arms are outstretched to the side).When the avatar lowers the arms, there is not enough cloth atopthe shoulders, so that stretch is high in that region, resulting in abad fit and a risk of eventually tearing the garment. We detect suchsituations by monitoring per-triangle stretch while the avatar ismoving through a set of scanned input poses, as described in Sec. 3.1.The user can optionally introduce manual edits to the garment thatare not directly related to fit, like puffy sleeves.
Per triangle stretch.
Garment mesh adaptions are defined locally,since reducing stretch can be expressed per mesh element. We followthe basic methodology of gradient domain processing [Botsch andSorkine 2008], similar to as-rigid-as-possible mesh editing [Sorkineand Alexa 2007] to deform the current rest shape ^X ∈ R 𝑛 × with 𝑛 vertices in order to reflect the desired changes expressed w.r.t.the current simulation mesh X ∈ R 𝑛 × . To this end we alternatethe modification of individual triangles with stitching them into aconsistent mesh by solving a Poisson system. The coordinates ofeach rest shape and simulation triangle are given respectively by ^x 𝑡 = (cid:0) ^x 𝑡 ^x 𝑡 ^x 𝑡 (cid:1) ∈ R × , x 𝑡 = (cid:0) x 𝑡 x 𝑡 x 𝑡 (cid:1) ∈ R × . Each triangle x 𝑡 can be rigidly transformed such that it is con-tained in the 2D plane and its first vertex coincides with the origin.The two vectors spanning this triangle form the × matrix 𝑃 x 𝑡 (see Fig. 8). This matrix is invertible whenever the triangle x 𝑡 isnon-degenerate. In order to measure stretch, we first construct thedeformation gradient F ∈ R × , which can be conveniently ex-pressed using the × matrix representation: F = 𝑃 x 𝑡 ( 𝑃 ^x 𝑡 ) − . (2)The deformation gradient exists and is invertible if the simulationtriangle x 𝑡 and the rest shape triangle ^x 𝑡 are both non-degenerate. Fig. 8. The deformation gradient relates the triangles of the rest shape andthe simulated garment.
Moreover, given an invertible deformation gradient, we can mapbetween the projected 2D rest shape and simulation triangle shape.We use this property to adapt the rest shape such that the simulationtriangle exhibits less stretch. The eigenvalues of the right Cauchy-Green tensor F T F are the squared principal stretches 𝑠 and 𝑠 .Consequently the singular values of the deformation gradient F aregiven by the principal stretches: F = U (cid:18) 𝑠 𝑠 (cid:19) V T . (3) Reference triangle adaption.
Our goal is to find a rest shape suchthat the stretch is within a certain range [ , + 𝛿 ] . A stretch valueabove means the cloth exhibits stretch, and a value below in-dicates compression. Our simulation always starts with a fittinggarment for one specific pose. Therefore we never want to decreasetriangle area, lest we destroy the fit. Hence, a threshold 𝛿 is onlyneeded to limit stretch, but not compression. To find new referencetriangles 𝑃 ^x ′ 𝑡 that meet our criteria locally, we keep the triangle ofthe simulated garment fixed and modify the principal stretches 𝑠 ′ 𝑖 = (cid:40) 𝑠 𝑖 / 𝛽 𝑡 𝑠 𝑖 / 𝛽 𝑡 < + 𝛿 + 𝛿 otherwise. (4)By limiting the principal stretches and setting them to smaller value( 𝑠 ′ 𝑖 < 𝑠 𝑖 ), we force the rest shape triangle to become larger accord-ingly. Additionally, we can manually enforce a larger rest shapetriangle by dividing by a user defined local scaling factor 𝛽 𝑡 > .Using the Paint tool , the user can add fabric by directly markingregions of the simulated mesh using a sketch-based interface. Bymodifying the brush intensity the amount of fabric that should beintroduced in a specific area can be controlled. Our system translatesthis input into the local scaling factor 𝛽 𝑡 for each mesh triangle anduses it to adapt the rest shape. The initial value for triangles thatare not manually enlarged is 𝛽 𝑡 = .We modify the singular value decomposition of F to find thedeformation gradient that is close to the original one while beingconsistent with the modified principal stretches: F ′ = U (cid:18) 𝑠 ′ 𝑠 ′ (cid:19) V T . (5)We can also directly construct the inverse of this matrix (cid:0) F ′ (cid:1) − = V (cid:18) / 𝑠 ′
00 1 / 𝑠 ′ (cid:19) U T . (6)Using this new deformation gradient, we can conveniently find twovectors spanning the corresponding projected rest shape triangle: (cid:0) y 𝑡 y 𝑡 (cid:1) = (cid:0) F ′ (cid:1) − 𝑃 x 𝑡 . (7) Rest shape update.
The previous step yields individual, 2D refer-ence triangles that are compatible with the requirements on stretch.In order to obtain a modified rest shape ^X ′ , these triangles have tobe stitched to form a consistent triangle mesh given by the set of tri-angles T and coordinates ^x ′ 𝑡 . To this end, we employ Poisson basedstitching using the cotan Laplacian, akin to as-rigid-as-possible • Katja Wolff, Philipp Herholz, Verena Ziegler, Frauke Link, Nico Brügel, and Olga Sorkine-Hornung Fig. 9. We stitch the 2D reference triangles y 𝑡 (left) to form a consistenttriangle mesh with 3D triangles ^x ′ 𝑡 (right). shape deformation [Sorkine and Alexa 2007]. This amounts to mini-mizing the objective function 𝐸 ( R , ^X ′ ) = ∑︁ 𝑡 ∈T ∑︁ 𝑘 = , , ( 𝑖,𝑗 ) is 𝑘 th edge of 𝑡 cot 𝛼 𝑡,𝑘 (cid:13)(cid:13)(cid:13) ( ^x ′ 𝑡 ) 𝑗 − ( ^x ′ 𝑡 ) 𝑖 − R 𝑡 y 𝑘𝑡 (cid:13)(cid:13)(cid:13) , (8)with respect to the new rest shape positions ^X ′ and the per trianglerotations R . Here 𝑖 and 𝑗 refer to the vertices of the 𝑘 − th edge oftriangle 𝑡 and y 𝑘𝑡 denotes the corresponding edge of the referencetriangle y 𝑡 . The matrix R 𝑡 ∈ R × represents a rotation from the2D plane into 3D and 𝛼 𝑡,𝑘 is the angle opposite of edge 𝑘 in triangle 𝑡 of the previous rest shape (see Fig. 9). The objective is optimizedby alternating minimization with respect to R and ^X ′ .Our overall pipeline consist of repeatedly adjusting the avatarmesh, followed by 𝑛 steps of cloth simulation based on a fixed restshape and a final step of adjusting the rest shape while the simulatedcloth is fixed. Computing several steps during the cloth simulationallows the garment to settle on the slightly adjusted avatar posebefore we calculate the stretch to adjust the rest shape. At the sametime, a small 𝑛 improves computation time. We observe that we canachieve good results with 𝑛 = . After the final rest shape of the garment is computed, we use it asa basis for further production. In principle, users can choose anyfabrication method that suits their needs, such as creating seamlessknitting patterns using the method of Wu et al. [2019]. In this paper,we opted to create all example garments by generating sewingpatterns through variational surface cutting [Sharp and Crane 2018].This method parameterizes surfaces over flat domains by directlyoptimizing the distortion induced by cutting and flattening. Assuggested in this work, we employ the Hencky energy, which isuseful for physical fabrication. We weigh this energy against thecut lengths to balance the number of sewing pattern pieces versusthe introduced distortion. We initialize the pattern through normalclustering, choose a length normalization weight of and Henckydistortion weight of and take steps. As was shown in [Sharpand Crane 2018], even short cuts can yield sewing pattern piecesthat nicely approximate the rest shape. We found that the flatteningintroduces only a very small amount of distortion across all ourexamples. In Fig. 12 we show an optimized and simulated dress (b)along with the same dress (f) generated from stitching the individualsewing pattern pieces (e) back together. Our implementation is based on several existing libraries: libigl [Jacobson et al. 2018],
PMP [Sieger and Botsch 2020],
OpenMP [Dagum and Menon 1998] and
Cholmod [Chen et al. 2008]. ForFigures 4, 5 and 6 we use avatars from the FAUST dataset [Bogoet al. 2014]. All other avatars have been scanned by us. We use acomputer with a 12-core 2.7 GHz CPU and 64 GB memory.The dress shown in Fig. 14 takes approximately 2.8 minutes to ad-just to the two additional poses. Computation times vary mainly dueto the varying number of poses and garment resolution. In Table 1we report average frame rates for all modeling sessions presentedin this paper. Our algorithm is implemented by augmenting a tradi-tional Newton-based simulation framework, which is responsiblefor the bulk of computation time. Running the simulation withoutadapting the garment results in the performance reported under“fps (sim)”; the full algorithm runs at a frame rate of “fps (full)”.Consequently, any method accelerating the chosen type of clothsimulation technique immediately benefits our algorithm. The framerate of our interactive application largely depends on the garmentmesh resolution reported as the number of vertices.For all examples shown in this paper we use the same clothparameters, chosen to simulate textiles that practically do not stretchin order to highlight the capabilities of our method, as they requirethe largest adjustments to the rest shapes. However, our methodworks with any kind of textile, if the cloth simulation parameters arechosen accordingly. We use the stiffness constants 𝑘 stretch = , 𝑘 shear = and 𝑘 bend = − and the damping stiffness constants 𝑘 d,stretch = , 𝑘 d,shear = and 𝑘 d,bend = − for the stretch,shear and bend constraints, respectively (for details we refer to[Baraff and Witkin 1998]). We advance the simulation with a timestep of ℎ = . and adjust the rest shape every 8 time steps witha stretch threshold of 𝛿 = . . We move through interpolationsteps between two poses. We empirically found these parameters toconsistently yield good results, however, any other set of parametersmight be chosen in order to achieve a different trade off betweenaccuracy and performance.We created several garments and computed the sewing patternsas detailed in Sec. 3.5. All are sewn by professional tailors. We listthe design parameters for the individual garments in Table 2. Men’s shirt.
We created two men’s shirts for the same person.The first one, shown in Fig. 10, is created from a single T-pose, while
Table 1. Performance measurements for all editing sessions. We report thenumber of vertices of the garment and average frames per second (fps), bothfor the full algorithm (fps (full)) and the simulation part only (fps (sim)).
Figure
D Custom Fit Garment Design with Body Movement • 9 (a) (b) (c) (d) (e)
Fig. 10. Starting from a single T-pose (a), we create the shape of a shirt (b) and its sewing pattern (c). This shirt fits perfectly in the T-pose, but createsuncomfortable stretch in other poses (d). Especially when the arms are held at the sides, visible stress is produced on the buttons of the shirt. This is alsonoticeable for arms outstretched to the front and up (e). (a) (b) (c) (d)(e) (f)
Fig. 11. Starting from the same shirt design for the T-pose as in Fig. 10, we now move through five different poses (a) to update the rest shape of the garment(b) and create a sewing pattern for fabrication (c). We show the final digital garment on all five poses (e), as well as the sewn shirt (f). The shirt accommodatesall five poses, and no uncomfortable stretch occurs (d). the second one is based on five different poses shown in Fig. 11,starting from the T-pose. For both shirts we use the
Boundary tool tocreate the shape of the garment and add an offset of 1.5 cm with the
Comfort tool . While moving through the different poses to adjustthe rest shape, we pin the collar and cuffs with the
Pinning tool toensure that the fabric of the sleeves is stretched sufficiently insteadof being pulled back when moving the arms. Finally, we pre-definethe shoulder seams and button border typical for shirts with the
Seam tool . The sewing pattern is then created from this pre-cut mesh(here we use a Hencky distortion weight of to create fewer seams).A professional tailor added the collar, cuffs and button border. Acomparison of both shirts (Fig. 10 (d) and Fig. 11 (e)) shows thatthe shirt created solely from the T-pose fits in this specific poseand stretches uncomfortably in other poses, close to tearing. The second shirt allows a wider range of motions. Fig. 2 shows additionalphotographs of the second shirt and compares to an off-the-shelf,standard shirt bought in a store for the same person. Both shirtstake approximately 7 hours to fabricate, including the cutting of thecloth, sewing and adding the cuffs, collar and buttons. Accordingto the tailors we worked with, the production time for garmentsdepends mostly on the number of pieces and the length of the seams.Therefore they estimate that the production time for our shirt and astandard one would not differ significantly. Maternity wear.
Our method can also be used to make garmentsfor special body shapes. We create a dress that fits during pregnancyand afterwards (Fig. 12), starting from the slim shape. After therest shape adjustment from the slim shape to the pregnant one, we (a) (b) (c) (d) (e) (f) (g)
Fig. 12. Our method can be used to create garments that fit during and after pregnancy. Instead of different poses, we start from different scanned bodyshapes (a). We design the initial dress shape on the slim pose (c), compute its adjustment to both poses to get the final shape (d) and create a sewing pattern(e). Comparison of the draped final rest shape on both poses (b) with the sewn and draped pattern (f) reveals only negligible difference. The sewn garment (g)can be worn during and after pregnancy. additionally adjust the rest shape slightly by increasing the clotharea using the
Paint tool below the belly for a smother transition. InFig. 12 (b) and (e) we compare the garment shape before generatingthe sewing pattern and after cutting, flattening and re-simulatingin 3D. The 3D garment shape does not change visibly.
Jumpsuit.
By creating a jumpsuit, we show that our method canhandle complex cases of larger garments that cover the whole body(Fig. 15). We again use the
Boundary tool to define the upper part ofthe jumpsuit, and the
Extension tool to create the legs. Starting fromthe A-pose with half-raised arms at the sides, we move through 4poses, while pinning the collar and cuffs with the
Pinning tool andadding an offset of 1 cm with the
Comfort tool . Using the
Seam tool ,we create a straight seam in the back for a zipper. Fabricating thedenim jumpsuit takes 14 hours, where half of the time is used tocreate the elaborate seaming.
Inclusive garment design.
The strength of our method is show-cased by creating garments for people who fall far outside the stan-dard sizes available in stores, see Fig. 1. In this example, imperfec-tions in the registered avatars for the different poses do appear,especially around the hands and feet, likely due to the SMPL modelnot being trained on such body shapes. Still, our method is robustto small scan errors, and we can still design a custom-fit dress.We show two additional simple results, created from just twoposes: a shirt (Fig. 13) and a white dress (Fig. 14), to highlight thevariety of possible garments. The shirt in Fig. 13 is designed inthe same way as the previous shirt from Fig. 11. The white dressis created similarly to the pregnancy dress from Fig. 12, but weuse the
Seam tool to cut a strip of cloth that is replaced by a redband. By calculating the underlying garment shape, but allowingthe professional tailor to add small details like buttons, trims andcollars, a diverse set of garments can be created.
Discussion.
The choice of the initial pose plays a vital role in thefinal garment shape, as the garment exhibits the least amount offolds in this pose. The further the avatar moves away from that pose, (a) (b) (c) (d)
Fig. 13. A shirt created from two poses (a). We show the final rest shape (b),the sewing pattern (c) and the physical garment (d). (a) (b) (c) (d)
Fig. 14. A dress created from two poses (a). We show the final rest shape(b), the sewing pattern (c) and the physical garment (d). the more folds can be expected. This can be seen when comparingthe shirt in Fig. 11, which is designed for the T-pose, and the shirtin Fig. 13, which is designed for the A-pose. There is a tradeoffbetween fit and movement range that has to be considered whenusing our system. The more poses the garment needs to fit tightlyin, the less the garment can fit each individual pose. For example,traditionally designed shirts usually do not accommodate a pose
D Custom Fit Garment Design with Body Movement • 11 (a) (b) (c) (d)(e) (f)
Fig. 15. The jumpsuit constitutes a challenging full-body example, where we use four poses as input (a). Note that in the first pose the arms are loweredhalfway at the sides, whereas in the last pose the arms are held halfway to the front. We use our toolset to create an initial garment (b) and compute theadjustment of the rest shape to all four poses (c), which we drape on all four avatar poses (e) and use to create a sewing pattern (d). The physical jumpsuit fitsall poses comfortably (f). with arms raised high above the head. By discarding this pose, wecan achieve a tighter fit with a limited movement range, which mostpeople are already accustomed to.
We presented a method to design and optimize the shape of a gar-ment for a range of 3D scanned poses, such that the garment fitscomfortably but tightly in all poses. We demonstrated the capabil-ities of our method by sewing different garments for a range ofindividual body shapes.
Limitations.
Since the input of our method is a number of reg-istered 3D scans, we are limited by the current state of the art inhuman body model capture. This becomes apparent when we scanpeople with proportions strongly deviating from the average, as canbe seen in Fig. 1, where difficulties in registering the scans resultin elongated hands and misplaced heels arise. In theory, we coulddesign garments for people with missing extremities, but currentmethods based on SMPL [Loper et al. 2015] would hallucinate theseextremities, requiring manual fixing of the body meshes. The sameregistration process also creates self-intersections in the avatars,leading to problems in cloth simulation and precluding the usage ofself-intersecting poses as the initial poses for designing a garment.
Future work.
Currently, the final shape of the garment dependson the sequence of poses and especially on the initial one. Thoughthis gives designers the option to choose a main pose in which thegarment is worn, creating a garment independent from the simula-tion sequence might consider the poses more equally. Furthermore,we initiate the rest shape of the garment from the body shape of the initial pose and do not take the draping effect into account in thisfirst step. Generating the initial rest shape such that it physicallydeforms into a tight fit could slightly improve our method in largerconvex areas, like under breasts and the bottom.Our method only allows for the body to influence the garmentshape, but not vice versa. In the future, we would like to incorporatethe physical simulation of body tissue to simulate the influence ofthe garment on the body. Especially in tight areas, body tissue canbe visibly displaced by the garment. Incorporating recent work inbody modeling, like the dynamic body model Dyna [Pons-Moll et al.2015] or the STAR model [Osman et al. 2020], would be a first stepin that direction. We would also like to improve the simulation byincorporating cloth parameters measured from real textiles. Incorpo-rating friction into the simulation would further improve the designof pants or skirts to prevent rubbing against the skin and sliding.
ACKNOWLEDGMENTSREFERENCES
David Baraff and Andrew Witkin. 1998. Large steps in cloth simulation. In
Proc.SIGGRAPH . 43–54.Aric Bartle, Alla Sheffer, Vladimir G Kim, Danny M Kaufman, Nicholas Vining, andFloraine Berthouzoz. 2016. Physics-driven pattern adjustment for direct 3D garmentediting.
ACM Trans. Graph.
35, 4 (2016).Miklos Bergou, Max Wardetzky, David Harmon, Denis Zorin, and Eitan Grinspun. 2006.A quadratic bending model for inextensible surfaces. In
Proc. SGP . 227–230.Floraine Berthouzoz, Akash Garg, Danny M Kaufman, Eitan Grinspun, and ManeeshAgrawala. 2013. Parsing sewing patterns into 3D garments.
ACM Trans. Graph.
Environmental Health
17, 1 (2018), 92.Federica Bogo, Javier Romero, Matthew Loper, and Michael J Black. 2014. FAUST:Dataset and evaluation for 3D mesh registration. In
Proceedings of the IEEE Conferenceon Computer Vision and Pattern Recognition . 3794–3801.
Table 2. Statistics for all garments.
Figure faces poses initial pose tools offset textile pieces sewing time model heightShirt T-pose 10 6 k 1 T-pose 1,4,6 1.5 cm cotton poplin 11 7 h 190 cmShirt 2, 11 6 k 5 T-pose 1,4,5,6 1.5 cm cotton poplin 10 6.5 h 190 cmJumpsuit 15 7 k 4 A-pose 1,2,4,5,6 1.0 cm denim 17 14 h 163 cmPregnancy dress 12 4 k 2 pre-pregnancy 1,2,3,4 0.8 cm lyocell 7 3 h 172 cmYellow dress 1 5 k 5 T-pose 1,2,4,5,6 0.5 cm lyocell 13 2.75 h 125 cmWhite dress 14 9 k 2 A-pose 1,2,4,6 0.5 cm polyester crepe 9 3 h 177 cmBlue Shirt 13 8 k 2 A-pose 1,4,5,6 1.0 cm cotton poplin 8 4.5 h 188 cm
Mario Botsch and Olga Sorkine. 2008. On linear variational surface deformationmethods.
IEEE Trans. Vis. Comput. Graph.
14, 1 (2008), 213–230.Robert Bridson, Sebastian Marino, and Ronald Fedkiw. 2005. Simulation of clothingwith folds and wrinkles. In
ACM SIGGRAPH 2005 Courses . 3–es.Remi Brouet, Alla Sheffer, Laurence Boissieux, and Marie-Paule Cani. 2012. Designpreserving garment transfer.
ACM Trans. Graph.
31, 4 (2012).Yanqing Chen, Timothy A Davis, William W Hager, and Sivasankaran Rajamanickam.2008. Algorithm 887: CHOLMOD, supernodal sparse Cholesky factorization andupdate/downdate.
ACM Trans. Math. Software
35, 3 (2008), 1–14.Kwang-Jin Choi and Hyeong-Seok Ko. 2005. Stable but responsive cloth. In
ACMSIGGRAPH 2005 Courses . 1–es.Gabriel Cirio, Jorge Lopez-Moreno, David Miraut, and Miguel A Otaduy. 2014. Yarn-level simulation of woven cloth.
ACM Trans. Graph.
IEEE Computer graphics and applications
23, 1 (2003), 38–48.Leonardo Dagum and Ramesh Menon. 1998. OpenMP: an industry standard API forshared-memory programming.
Computational Science & Engineering, IEEE
5, 1(1998), 46–55.Philippe Decaudin, Dan Julius, Jamie Wither, Laurence Boissieux, Alla Sheffer, andMarie-Paule Cani. 2006. Virtual garments: A fully geometric approach for clothingdesign.
Comput. Graph. Forum
25, 3 (2006), 625–634.Rony Goldenthal, David Harmon, Raanan Fattal, Michel Bercovier, and Eitan Grinspun.2007. Efficient Simulation of Inextensible Cloth.
ACM Trans. Graph.
26, 3 (July 2007),49–es. https://doi.org/10.1145/1276377.1276438Yuki Igarashi, Takeo Igarashi, and Hiromasa Suzuki. 2008. Automatically adding seamallowance to cloth pattern. In
ACM SIGGRAPH 2008 posters . ACM, 15.Alec Jacobson, Daniele Panozzo, et al. 2018. libigl: A simple C++ geometry processinglibrary. https://libigl.github.io/.Il-Hoe Jung, Sang-Bin Lee, Jong-Jun Kim, Han-Na Ryu, and Hyeong-Seok Ko. 2016.Modeling the non-elastic stretch deformation of cloth based on creep analysis.
TextileResearch Journal
86, 3 (2016), 245–255.Jonathan M Kaldor, Doug L James, and Steve Marschner. 2010. Efficient yarn-basedcloth with adaptive contact linearization. In
ACM SIGGRAPH 2010 papers . 1–10.Michael Keckeisen, Matthias Feurer, and Markus Wacker. 2004. Tailor tools for in-teractive design of clothing in virtual environments. In
Proceedings of the ACMsymposium on Virtual reality software and technology . ACM, 182–185.Tsz-Ho Kwok, Yan-Qiu Zhang, Charlie CL Wang, Yong-Jin Liu, and Kai Tang. 2016.Styling evolution for tight-fitting garments.
IEEE Trans. Vis. Comput. Graph.
22, 5(2016), 1580–1591.Minchen Li, Alla Sheffer, Eitan Grinspun, and Nicholas Vining. 2018. FoldSketch:Enriching Garments with Physically Reproducible Folds.
ACM Trans. Graph.
37, 4(2018). https://doi.org/10.1145/3197517.3201310Tiantian Liu, Adam W Bargteil, James F O’Brien, and Ladislav Kavan. 2013. Fastsimulation of mass-spring systems.
ACM Trans. Graph.
32, 6 (2013), 1–7.Matthew Loper, Naureen Mahmood, Javier Romero, Gerard Pons-Moll, and Michael J.Black. 2015. SMPL: A Skinned Multi-Person Linear Model.
ACM Trans. Graphics(Proc. SIGGRAPH Asia)
34, 6 (Oct. 2015), 248:1–248:16.Yuwei Meng, Charlie CL Wang, and Xiaogang Jin. 2012. Flexible shape control forautomatic resizing of apparel products.
Computer-Aided Design
44, 1 (2012), 68–76.Meshcapade. 2020. meshcapade.com. https://meshcapade.com.Eder Miguel, Derek Bradley, Bernhard Thomaszewski, Bernd Bickel, Wojciech Matusik,Miguel A Otaduy, and Steve Marschner. 2012. Data-driven estimation of clothsimulation models. In
Computer Graphics Forum , Vol. 31.Juan Montes, Bernhard Thomaszewski, Sudhir Mudur, and Tiberiu Popa. 2020. Compu-tational design of skintight clothing.
ACM Trans. Graph.
39, 4 (2020), 105–1.A Morlet, R Opsomer, S Herrmann, L Balmond, C Gillet, and L Fuchs. 2017. A newtextiles economy: redesigning fashion’s future.
Ellen MacArthur Foundation (2017).Rahul Narain, Armin Samii, and James F O’brien. 2012. Adaptive anisotropic remeshingfor cloth simulation.
ACM Trans. Graph.
31, 6 (2012), 1–10. Vidya Narayanan, Kui Wu, Cem Yuksel, and James McCann. 2019. Visual knittingmachine programming.
ACM Trans. Graph.
38, 4 (2019), 1–13.Rajkishore Nayak and Rajiv Padhye. 2017.
Automation in Garment Manufacturing .Woodhead Publishing.Occipital. 2020. structure.io. https://structure.io/structure-sensor.Optitex. 2020. optitex.com. https://optitex.com.Ahmed AA Osman, Timo Bolkart, and Michael J Black. 2020. STAR: Sparse TrainedArticulated Human Body Regressor. arXiv preprint arXiv:2008.08535 (2020).Gerard Pons-Moll, Javier Romero, Naureen Mahmood, and Michael J. Black. 2015. Dyna:A Model of Dynamic Human Shape in Motion.
ACM Trans. Graph.
34, 4 (2015).C. Robson, R. Maharik, A. Sheffer, and N. Carr. 2011. Context-Aware Garment Modelingfrom Sketches.
Computers & Graphics (Proc. SMI 2011)
35, 3 (2011), 604–613.Kenneth Rose, Alla Sheffer, Jamie Wither, Marie-Paule Cani, and Boris Thibert. 2007.Developable surfaces from arbitrary sketched boundaries. In
Proc. Symposium onGeometry Processing . Eurographics Association, 163–172.Nicholas Sharp and Keenan Crane. 2018. Variational surface cutting.
ACM Trans. Graph.
Proc.Symposium on Geometry Processing . 109–116.Robert W Sumner and Jovan Popović. 2004. Deformation transfer for triangle meshes.
ACM Trans. Graph.
23, 3 (2004), 399–405.Rasmus Tamstorf and Eitan Grinspun. 2013. Discrete bending forces and their Jacobians.
Graphical models
75, 6 (2013), 362–370.Min Tang, Tongtong Wang, Zhongyuan Liu, Ruofeng Tong, and Dinesh Manocha. 2018.I-cloth: incremental collision handling for GPU-based interactive cloth simulation.
ACM Trans. Graph.
37, 6 (2018), 1–10.Joseph Teran, Eftychios Sifakis, Geoffrey Irving, and Ronald Fedkiw. 2005. Robustquasistatic finite elements and flesh simulation. In
Proceedings of the 2005 ACMSIGGRAPH/Eurographics symposium on Computer animation . 181–190.Emmanuel Turquin, Jamie Wither, Laurence Boissieux, Marie-Paule Cani, and John FHughes. 2007. A sketch-based interface for clothing virtual characters.
IEEE Com-puter Graphics and Applications
27, 1 (2007).Nobuyuki Umetani, Danny M Kaufman, Takeo Igarashi, and Eitan Grinspun. 2011.Sensitive couture for interactive garment modeling and editing.
ACM Trans. Graph.
30, 4 (2011), 90–1.Pascal Volino, Frederic Cordier, and Nadia Magnenat-Thalmann. 2005. From earlyvirtual garment simulation to interactive fashion design.
Computer Aided Design
Computer-aided design
37, 7 (2005), 675–691.Huamin Wang. 2018. Rule-free sewing pattern adjustment with precision and efficiency.
ACM Trans. Graph.
37, 4 (2018), 1–13.Huamin Wang, James F O’Brien, and Ravi Ramamoorthi. 2011. Data-driven elasticmodels for cloth: modeling and measurement.
ACM Trans. Graph.
30, 4 (2011), 1–12.Tuanfeng Y. Wang, Duygu Ceylan, Jovan Popovic, and Niloy J. Mitra. 2018. Learninga Shared Shape Space for Multimodal Garment Design.
ACM Trans. Graph.
37, 6(2018), 1:1–1:14. https://doi.org/10.1145/3272127.3275074Amy Wibowo, Daisuke Sakamoto, Jun Mitani, and Takeo Igarashi. 2012. DressUp: a3D interface for clothing design with a physical mannequin. In
Proc. InternationalConference on Tangible, Embedded and Embodied Interaction . ACM, 99–102.Kui Wu, Hannah Swan, and Cem Yuksel. 2019. Knittable stitch meshes.
ACM Trans.Graph.
38, 1 (2019), 1–13.Dong Xu, Hongxin Zhang, Qing Wang, and Hujun Bao. 2006. Poisson shape interpola-tion.