Toward Edge-Level Comparison of Sewing Patterns: A Construction-Aware Edge-Role Representation

A sewing pattern is a manufacturing specification that describes how a garment is cut and assembled, rather than a collection of similar polygons. Small changes in a numerical feature such as edge length can alter garment size, or style. Although the amount varies across garment parts, the average length adjustment applied to an individual edge for size variation is approximately 3–4cm. Even a dimensional change of this scale can affect the compatibility of connected edges and the resulting garment fit. To address this, recent learning-based methods reconstruct structured sewing patterns instead of representing garments only as final draped surfaces [1]–[7]. Recent related models generate sewing patterns from images, sketches, text, body measurements, or multimodal prompts [1]–[7]. Their outputs can appear plausible yet still require manual correction when edge dimensions are wrong - here, an edge refers to an individual straight or curved segment forming part of a pattern outline. Previous evaluation metrics for sewing-pattern generation have focused on whether the generated panels have overall outline shapes similar to those of the reference pattern. However, these metrics have limited ability to determine whether the individual edges used for garment construction have appropriate lengths, directions, and curvature. Improving edge-level accuracy may require changes to model architectures or post-processing methods, but such work first requires a consistent way to identify edges that perform the same construction function. This study takes a foundational step toward that goal through three contributions. First, this study introduces a unified interval representation in which every source edge is stored as one or more records. Second, it defines a construction-aware set of 26 roles covering independently finished and stitched sewing-pattern edges. Third, it provides a preliminary application of the proposed roles to 200 GarmentCodeData patterns and reports their role coverage and frequency distribution.

This section defines panels, edges, stitches, and normalized edge intervals. It also reviews existing reconstruction methods and their limitations in identifying edge roles and correspondences.

Following the sewing-pattern representation used by NeuralTailor [1], a sewing pattern consists of multiple fabric pieces, referred to as panels, together with their geometric placement and stitch relationships. Figure 1 illustrates the three basic elements of this representation: panels, edges, and stitches. In the example, the upper garment consists of three panels: a front bodice, a back bodice, and a sleeve. An edge is an individual straight or curved boundary segment connecting two consecutive vertices of a panel. The illustrated front bodice consists of 12 boundary edges, including the segments created by the side and waist darts. In the middle-left example of Figure 1, the stitch marks highlight the front-bodice shoulder edge, which is stitched to the corresponding back-bodice shoulder edge. An edge may be represented geometrically as a straight line, circular arc, or Bézier curve and may be characterized by its length, direction, endpoint tangents, and curvature. Edge correspondences are not necessarily one-to-one. Depending on the garment structure, one edge may correspond to multiple partner edges. For example, the sleeve-cap boundary connects to both the front- and back-bodice armhole edges, producing a one-to-multiple construction relationship. A stitch specifies which boundary edges of different panels are sewn together. In the underlying sewing-pattern representation, stitch information is stored in JSON using panel and edge references. The right-hand example in Figure 1 shows a stitch between a sleeve-cap edge and a front-bodice armhole edge. Although each stitch record represents a pairwise connection, the complete construction relationship may include multiple partner edges, as demonstrated by the sleeve-cap example.

To describe edge roles and stitch relationships consistently, each edge is parameterized by t ∈ [0, 1]. An edge or part of an edge is represented as ((t_start, t_end), role), where 0 ≤ t_start < t_end ≤ 1. A role covering the complete edge uses the interval ((0, 1), role). When the construction role or stitch partner changes along an edge, the edge is divided at the corresponding change boundaries into multiple non-overlapping intervals. Each stitched interval retains its own stitch_id and reciprocal partner-edge and partner-interval references. In a one-to-one correspondence, one complete edge interval is connected to one complete partner interval. For example, the complete front-bodice shoulder interval may be connected to the complete back-bodice shoulder interval. In a one-to-multiple correspondence, the original edge is not duplicated. Instead, it is divided into non-overlapping intervals according to its partner edges. Thus, the sleeve-cap edge may be divided into a front interval stitched to the front-bodice armhole and a back interval stitched to the back-bodice armhole. The intervals assigned to an edge do not overlap, and their union covers the complete normalized range [0, 1]. Stitch connectivity identifies which intervals are sewn together but does not independently define their construction functions. Therefore, a semantic role is assigned to each interval. Intervals that are not connected to another panel, such as independently finished boundaries, retain null stitch_id and partner-reference fields.

Efforts to recover detailed sewing-pattern information have progressively moved from panels to their sewing relations and edge geometry. NeuralTailor reconstructs variable-cardinality panel structures from garment point clouds [1], recovering each panel's outline and placement numerically rather than approximating only the final draped surface. However, since the method focuses on panel-level shape recovery, the function each edge performs in garment assembly—such as a seam, hemline, or armhole—must be inferred from its position or connectivity data, leaving edge semantics without an explicit definition. SewFormer predicts panels and stitches hierarchically from a single image using a visual encoder, a two-level Transformer decoder for panel and edge reconstruction, and a stitch-prediction module [2]. PanelFormer uses two sequential components: first, its Panel Transformer estimates the shapes and spatial placement of visible and occluded panels; second, its Stitch Predictor determines which edges of the reconstructed panels should be sewn together [3]. By supplementing panel geometry with predicted stitch-pair information, these methods recover both the panel shapes and their sewing connectivity. However, stitch information indicates only which edges are connected and does not by itself determine their specific construction roles. For example, a predicted connection may indicate that a bodice edge is sewn to a sleeve edge, but additional criteria based on panel identity and construction relationships are required to label the bodice-side edge as the armhole and the sleeve-side edge as the sleeve cap.

An alternative approach represents each panel using its boundary curves, connections to other components, and sewing information. GarmentCode [8] introduced this structured data, and GarmentCodeData [9] extended it to a large garment corpus. Recent work has focused on accommodating diverse input modalities—text, sketches, images, body measurements, and multimodal prompts—and seeks to improve generation accuracy by combining multiple models or increasing parameter scale [4]–[7]. Yet these studies identify numerical errors as an important limitation, underscoring the need for edge-level evaluation beyond panel and stitch reconstruction. Because existing datasets explicitly record stitched edge pairs, stitch-connectivity accuracy can be evaluated by checking whether a predicted edge pair matches its reference pair. However, these datasets do not consistently specify the functional roles of the connected edges or their interval-level correspondences especially in one-to-multiple relationships. Consequently, it is difficult to determine which geometrically corresponding regions should be compared when evaluating differences in edge length, direction, and curvature.

This study defines a role set comprising eight independently finished edge roles and eighteen stitched-edge roles. An independently finished edge is completed without being stitched to another panel, for example through folding, binding, or hemming. Every source edge is represented in the same form. When one role applies to the complete edge, it is stored as ((0, 1), role). When different portions of an edge connect to different panels or perform different construction functions, the edge is divided into multiple non-overlapping intervals, and each interval is assigned its corresponding role.

The candidate roles were collected by reviewing edge-related names, panel information, and stitch descriptions in the sewing-pattern dataset GarmentCodeData [9]. Because the dataset uses different naming conventions, names referring to the same garment-construction function were first standardized and combined. The candidate roles were collected, standardized, and reviewed by the first author, who holds a Ph.D. in fashion and has domain expertise in garment construction, and this domain-informed review served as the basis for selecting the final 26 roles.

The final role set was formed by reviewing edge-role occurrences in the available sewing-pattern data and consolidating sparsely represented or insufficiently supported construction functions into the broader opening and seam_other categories. It contains eight independently finished roles and eighteen stitched roles. The preliminary 200-pattern application showed that every retained role occurred at least once, with observed counts ranging from 76 sleeve_hem intervals to 2,576 waistline_seam intervals. This result demonstrates use of the complete role set within the selected patterns, but it does not validate the original role-selection decision or prove the correctness of individual assignments. The proposed set should therefore be understood as a data-informed, expert-guided starting point for sewing-pattern edge description. Tables 1 and 2 summarize the retained roles and their construction evidence.

Edge-role assignment begins by identifying the source panel and organizing its source edges according to their stitch relations. The panel name is interpreted as a canonical garment part, together with its side and front/back attributes when available. Each source edge is then expressed as an ordered set of role-labeled intervals, regardless of whether the construction relation covers the whole edge or only a subinterval. A representative midpoint is calculated for each role-labeled interval and normalized within the panel bounding box to determine its panel-local region. Construction roles are assigned in priority order using canonical panel attributes, stitch status, partner-panel relations, panel-local position, ordered edge adjacency, and previously assigned neighboring roles. The two intervals forming a stitch are classified jointly so that their roles and reciprocal partner references remain compatible.

The panel name recorded in the source dataset is first interpreted as a canonical garment part, such as bodice, sleeve, skirt, trouser, collar, cuff, waistband, or gore. Side and front/back attributes are retained when they are available. For example, a trouser panel may be represented by the attributes (trouser, left, front), while a bodice panel may be represented by (bodice, right, back). After the panel attributes have been identified, all edges belonging to the panel are collected as role-assignment candidates. For each edge, the source panel ID, source edge ID, position in the ordered panel loop, stitch status, partner edge, and stitch interval are retained when available.

Let a source edge e be parameterized by t over the normalized range [0, 1]. The edge is represented as an ordered sequence R(e) = [((t_start(1), t_end(1)), r(1)), …, ((t_start(n), t_end(n)), r(n))], where r(k) denotes the construction role assigned to interval k. The records are ordered according to their starting positions. If a single construction role and stitch partner apply to the complete edge, the representation contains only one record: R(e) = [((0, 1), r)]. If the construction role or stitch partner changes along the edge, the edge is divided at each change boundary and represented by multiple intervals. Every interval satisfies 0 ≤ t_start(k) < t_end(k) ≤ 1. The first interval begins at 0, the final interval ends at 1, and the end of each interval equals the start of the following interval. Consequently, the intervals are non-overlapping and collectively cover the entire source edge. For stitched intervals, corresponding metadata—including stitch_id, partner_edge_id, and partner_interval—are explicitly assigned. For an independently finished interval, these three fields are set to null. If the orientation of an edge is reversed, an interval [a, b] is converted to [1 − b, 1 − a]. This unified representation preserves complete, partial, and one-to-multiple stitch relationships without introducing a separate auxiliary edge object.

A representative midpoint is calculated for each role-labeled interval. Let the interval extend from a = t_start to b = t_end. For a straight interval, the geometric midpoint between its endpoints is used. For a curved interval, however, the arithmetic mean of the endpoint coordinates may not lie at the center of the curve. The midpoint is therefore defined according to arc length. The interval is sampled over the parameter range [a, b], and the distances between consecutive sample points are accumulated to estimate its total arc length L. Let s(t) denote the cumulative arc length measured from the beginning of the interval at a to a parameter value t. The midpoint parameter t_m is defined by s(t_m) = L/2. When L/2 lies between two sampled points, the midpoint coordinates are obtained by linear interpolation between those points. This definition is applied to straight segments, circular arcs, and Bézier curves; therefore, t = 0.5 is not assumed to be the midpoint of a curved edge or interval.

Before the positional ratios are calculated, each panel is placed in its canonical orientation. This step provides consistent upper, lower, left, and right regions even when the source panel has been rotated or mirrored. Let (x_m, y_m) denote the arc-length midpoint of a role-labeled interval, and let x_min, x_max, y_min, and y_max define the bounding box of the canonically oriented panel. The normalized midpoint coordinates are calculated as x_n = (x_m − x_min) / (x_max − x_min) and y_n = (y_m − y_min) / (y_max − y_min). The normalized coordinates are then divided into coarse panel-local regions using thresholds of 0.25 and 0.75. An interval with y_n > 0.75 is assigned to the upper region, whereas an interval with y_n < 0.25 is assigned to the lower region. Similarly, x_n < 0.25 indicates the left lateral region, and x_n > 0.75 indicates the right lateral region. Values from 0.25 to 0.75 indicate the central region along the corresponding axis. These thresholds provide interpretable spatial divisions within a panel. The resulting panel-local region is combined with the panel type, stitch relation, partner-panel information, and adjacent edge roles to determine the role of each interval.

Edge-role rules are applied according to the specificity of their construction evidence. Explicit within-part stitch relations are evaluated first, followed by complementary relations between different garment parts. Independently finished roles are assigned after all stitched relations have been processed, and generic labels are used only when no specific role can be determined. Once a role has been assigned by a higher-priority rule, it is not replaced by a later rule.

Relations between panels belonging to the same garment-part family are evaluated before connections between different garment parts. When two explicitly connected edges form a dart within the same panel, both edges are assigned dart_leg. For trouser panels, crotch_seam is identified before inseam and outseam. A connection between opposite-side trouser panels with the same front/back attribute is assigned crotch_seam. Among the connections between front and back trouser panels belonging to the same side, the edge that shares an endpoint with crotch_seam or continues from it in the ordered panel-edge sequence is assigned inseam. The remaining same-side front–back connection is assigned outseam. Connections between compatible front and back bodice panels are classified as shoulder_seam or side_seam according to their panel-local regions and adjacent roles. An upper stitched edge occurring between the neckline and armhole regions is assigned shoulder_seam, whereas a lateral stitched edge continuing from the armhole region toward the waistline or hemline region is assigned side_seam. Other within-part roles, including sleeve_underarm_seam, cuff_seam, center_seam, center_front_seam, and center_back_seam, are assigned from their corresponding canonical panel types and stitch relations.

After the within-part relations have been assigned, connections between different garment parts are evaluated. A bodice–sleeve connection receives the complementary roles armhole and sleeve_cap, while a bodice–collar connection receives neckline_seam and collar_attachment. A connection between a sleeve and a cuff, or between a trouser panel and a cuff, is assigned cuff_attachment on the cuff side together with the corresponding attachment role on the sleeve or trouser side. The two connected edge intervals are classified jointly. If the bodice-side interval is assigned armhole, the connected sleeve-side interval is assigned sleeve_cap. Likewise, if the bodice-side interval is assigned neckline_seam, the connected collar-side interval is assigned collar_attachment. A role pair is accepted only when both roles are compatible with the canonical panel types and reciprocal partner references are present.

Independently finished roles are assigned after all stitched relations have been processed. Their roles are determined from the canonical panel identity, normalized panel-local region, ordered edge sequence, and roles of adjacent edges. On a sleeve panel, an independently finished edge located opposite the sleeve_cap region is assigned sleeve_hem. If the sleeve end is stitched to a cuff, the corresponding edge is assigned cuff_attachment instead. On a bodice panel, an arm-opening edge connected to a sleeve is assigned armhole, whereas the equivalent region without a sleeve partner is assigned armhole_edge. An independently finished bodice edge in the upper central region is assigned neckline_opening unless it is connected to a collar. An independently finished edge in the lower region of a bodice, skirt, or trouser panel is assigned hemline. Independently finished edges on collar and cuff panels are assigned collar_edge and cuff_edge when they are not used for attachment or internal construction seams. An independently finished edge in the upper region of a skirt or trouser panel is assigned waistline_opening when it has no bodice or waistband partner. A stitched interval that cannot be assigned a specific role is labeled seam_other, while an unclassified independently finished interval is labeled opening.

A preliminary application examined whether the role-assignment rules in Section 4 could be applied to GarmentCodeData patterns [9].

To examine the applicability of the role set across different garment structures, the evaluation sample included all five garment categories: 64 Top+Skirt, 63 Top, 42 Skirt, 16 Trousers, and 15 Top+Trousers designs. In total, the 200 patterns contained 2,226 panels, 15,076 edge intervals, and 6,291 stitches. Panel names were converted into standard panel types, while available left/right and front/back attributes were retained. Roles were then assigned using panel type, stitch connection, partner panel, edge position, and neighboring edges. Initial applicability was measured using role coverage, adapted from the class-richness concept of OntoQA [10]. The 26 edge roles were treated as classes, and the edge intervals assigned to each role were treated as instances. Role coverage was calculated as the number of observed roles divided by the number of defined roles. All 26 defined roles appeared at least once, resulting in a role-coverage value of 1.00. This result indicates that the complete role set could be instantiated in the selected patterns, although it does not verify the correctness of each individual role assignment.

Role frequencies were uneven. waistline_seam was the most frequent role with 2,576 intervals (17.09%), followed by seam_other with 2,144 (14.22%) and dart_leg with 1,872 (12.42%). Together with hemline, side_seam, and cuff_seam, the six most frequent roles accounted for 61.02% of all intervals. Of all seam_other intervals, 41.14% occurred in Top+Skirt patterns, whereas Skirt showed the highest within-category seam_other rate at 39.41%, possibly reflecting decorative or additional construction elements not covered by the current role set. By contrast, roles associated with specific structures, including sleeve_hem, collar_edge, collar_attachment, and neckline_seam, occurred less frequently. The high proportion of seam_other primarily stems from skirt patterns, which contain a greater variety of design features and specialized edges that are not covered by the current role set. While this does not necessarily indicate misclassification, expanding the role set could enable a more detailed description of these construction features.

This foundational study proposed a construction-aware set of 26 functional edge roles and a unified interval representation for sewing-pattern edges. By expressing each source edge as one or more ((t_start, t_end), role) records, the representation can retain complete, partial, and one-to-multiple stitch relationships together with their partner information. The primary contribution is therefore a basic method for describing edge function consistently across different panel and stitch structures. The proposed roles were applied to 200 GarmentCodeData patterns containing 2,226 panels, 15,076 edge intervals, and 6,291 stitches. All 26 roles appeared at least once, producing an OntoQA-inspired role-coverage value of 1.00. The accompanying frequency analysis showed that common construction roles and more specialized roles occurred at substantially different rates. Although these findings confirm the feasibility of the proposed role set across the dataset, future work should focus on independent expert validation and scaling the analysis to larger, more diverse datasets. The role definitions and assignments require independent review by professional pattern makers, and the analysis should be repeated on larger and more systematically sampled datasets. Future work should also document the pre-consolidation candidate roles, test partial and one-to-multiple relations in greater detail, and evaluate whether the representation supports reliable comparison of edge length, direction, curvature, stitch compatibility, and construction feasibility.