AutoSew: A Geometric Approach to Stitching Prediction with Graph Neural Networks

Pablo Ríos-Navarro; Elena Garces; Jorge Lopez-Moreno

cs.CVFeb 25, 2026

AutoSew: A Geometric Approach to Stitching Prediction with Graph Neural Networks

Authors: Pablo Ríos-Navarro, Elena Garces, Jorge Lopez-Moreno

Organizations: Universidad Rey Juan Carlos, Madrid, Spain · Adobe Research, France

Abstract

Automating garment assembly from sewing patterns remains a significant challenge due to the lack of standardized annotation protocols and the frequent absence of semantic cues. Existing methods often rely on panel labels or handcrafted heuristics, which limit their applicability to real-world, non-conforming patterns. We present AutoSew, a fully automatic, geometry-based approach for predicting stitch correspondences directly from 2D pattern contours. AutoSew formulates the problem as a graph matching task, leveraging a Graph Neural Network to capture local and global geometric context, and employing a differentiable optimal transport solver to infer stitching relationships-including multi-edge connections. To support this task, we update the GarmentCodeData dataset modifying over 18k patterns with realistic multi-edge annotations, reflecting industrial assembly scenarios. AutoSew achieves 96% F1-score and successfully assembles 73.3% of test garments without error, outperforming existing methods while relying solely on geometric input. Our results demonstrate that geometry alone can robustly guide stitching prediction, enabling scalable garment assembly without manual input.

Figures & tables

Figure 2: Overview of AutoSew. For each sewing pattern, a graph is constructed with nodes representing stitching edges and edges capturing geometric relationships of the panels. A GNN with message passing learns embeddings for each edge, optimized for stitch correspondence using a differentiable optimal transport loss.

Figure 3: Comparison of sewing patterns and draped avatars before and after multi-edge merging. Top: original dataset with one-to-one stitches, showing artificial seam marks. Bottom: our Multi-Edge dataset with realistic multi-to-one stitches, producing more natural seams and garment reconstructions.

Figure 4: Random results showcasing actual patterns from the M-E.GarmentCodeData .

Config.	TF1 / MEF1	GSP
Full	96.2 / 84.7	73.3
w/o Panel ID	87.7 / 72.4	45.3
w/o Topol. Prop.	78.8 / 73.7	30.5

Table 1: Ablation study on the geometric descriptors. Both topological properties and the Panel ID (unique panel identifier) are crucial for stitching prediction performance.

Config.	TP/MEP	TR/MER	TF1/MEF1	GSP
Full	96.1/80.4	96.3/89.5	96.2/84.7	73.3
w/o GNN	93.2/71.2	94.2/92.9	93.7/80.6	58.6

Table 2: Ablation study results on the test set. Removing message aggregation (GraphSAGE) severely impacts performance, particularly GSP.

GNN layers	Total F1	Multi-edge F1	GSP
$L=3$	94.3	81.8	61.8
$L=5$	96.2	84.7	73.3
$L=7$	96.0	78.9	71.6
5 + Max	96.1	84.8	72.4

Table 3: Impact of GNN depth and aggregation method on stitching prediction performance. Increasing the number of layers improves performance up to

L=5

, but deeper networks (

L=7

) exhibit reduced GSP due to over-smoothing. Using max aggregation instead of mean results in similar performance, with a slight drop in recall.

Configuration	TF1	MEF1	Time
$T=500$	97.1	84.5	26ms
$T=100$	97.0	84.7	7ms

Table 4: Effect of Sinkhorn iterations (

T

) on stitching prediction performance and inference time.

Figure 5: Impact of confidence threshold on AutoSew’s stitching prediction performance.

Appendix figures & tables1 asset

Supplementary material from the paper’s appendix.

Appendix

Geometric Features	Feature	Description
Local Shape	Start vertex $(x_{0},y_{0})$	Coordinates of the edge’s starting point
	End vertex $(x_{1},y_{1})$	Coordinates of the edge’s ending point
	Length $l$	Euclidean distance between start and end
	Orientation $(o_{x},o_{y})$	Unit vector between the initial and end edge point
	Curvature type $k_{t}$	$\{0\dots 5\}$ : straight segment, circular arcs quadratic Bézier, cubic Bézier and B-Splines
	Curvature params $k=(k_{1},\dots,k_{10})$	All zeros for straight segment $(r,d,\theta,0,\dots,0)$ for circular arc $(q^{1}_{x},q^{1}_{y},0,\dots,0)$ for quadratic Bézier $(q^{1}_{x},q^{1}_{y},q^{2}_{x},q^{2}_{y},0,\dots,0)$ for cubic Bézier $(q^{1}_{x},q^{1}_{y},q^{2}_{x},q^{2}_{y},c_{x},c_{y},q^{3}_{x},q^{3}_{y},q^{4}_{x},q^{4}_{y})$ for B-splines

Appendix

Table 1: Geometric edge features composed of Local Shape descriptor and Topological properties.

Manufacturing

Manufacturing research develops methods and systems to design, plan, automate, monitor, and improve industrial production.

See all 77 papers

Learning-based Seam Correspondence Reconstruction in Sewing Patterns
Jul 23, 2026Zhendong Wang, Jintong Wang, Chen Liu +3
Style3D Research, China · Zhejiang Sci-Tech University, China · University of Science and Technology of China, China
Towards automatic digital garment assembly
Oct 12, 2026Alejandro Rodriguez
SEDDI.
NGL: Natural Garment Language for Training-Free Sewing Pattern Estimation
Feb 24, 2026Anna Badalyan, Pratheba Selvaraju, Giorgio Becherini +33D Garment Reconstruction 3D Reconstruction
Max Planck Institute for Intelligent Systems

Jul 23, 2026J/K move · Enter open · S save

Learning-based Seam Correspondence Reconstruction in Sewing Patterns

Zhendong Wang, Jintong Wang, Chen Liu, Yao Jin+2

Style3D Research, China · Zhejiang Sci-Tech University, China · University of Science and Technology of China, China

Open paper AskPDF arXiv

Digital sewing patterns typically consist of disjoint 2D panels without explicit stitch annotations, making downstream 3D modeling reliant on labor-intensive expert specification. In this paper, we present a graph-based learning framework that reconstructs two-level stitching information, coarse panel connectivity and fine-grained seam correspondence, from 2D panel geometry alone. At the coarse level, panel connectivity is inferred by predicting panel semantics associated with anatomical body regions, enforcing consistency with body structure and garment design conventions. Based on the reconstructed panel graph, fine-grained seam correspondences between panel pairs are inferred by learning latent edge representations that jointly encode local seam geometry and global garment context through graph message passing. The resulting edge embeddings are subsequently decoded into detailed seam correspondences. Our method supports complex sewing-pattern topologies, including many-to-one correspondences, intra-panel seams, and curved seams. Experiments demonstrate high stitching accuracy and strong generalization across garment styles.

Abstract

Manufacturing

Explore similar work

Learning-based Seam Correspondence Reconstruction in Sewing Patterns

Towards automatic digital garment assembly

NGL: Natural Garment Language for Training-Free Sewing Pattern Estimation