PaperlayerView exact revision on arXiv

Source-linked AI summary

CT2Yarn: Yarn-Level Reconstruction of Crochet from Computed Tomography

Chang Luo, Nobuyuki Umetani

arXiv:2609.06950v1cs.GR

TL;DR

Real crochet yarn paths are difficult to recover because loops occlude one another and fiber orientations do not directly reveal yarn direction. CT2Yarn combines orientation estimation, centerline extraction, topology repair, and interactive correction. It is evaluated on 18 hand-crocheted samples and supports downstream applications, while tightly crocheted samples remain challenging.

  • Problem

    Recovering yarn paths from real crochet is challenging because contacting, occluded loops and twisted fibers make yarn trajectories difficult to distinguish in micro-CT volumes.

    Source: §1 · p. 2

  • Method

    CT2Yarn estimates fiber orientations, aggregates them into yarn-level orientations, extracts fragments, repairs their topology, and lets users correct unresolved connections.

    Source: §5 · p. 10

  • Results

    The method was evaluated on 18 hand-crocheted samples across three stitch types, with visual comparisons, quantitative metrics, and ablations; the recovered paths support simulation, rendering, and stitch-pattern extraction.

    Source: §5 · p. 10

  • Takeaways & Limitations

    The authors release micro-CT volumes and reconstructed centerlines to support future research, alongside yarn paths usable for simulation, ply-level rendering, and stitch-pattern extraction.

    Source: §5 · p. 10

    On tightly crocheted samples, reduced boundary contrast can cause center-point merging, producing shorter curves that require more manual editing.

    Source: §5 · p. 10

Abstract

from arXiv · show

We introduce CT2Yarn, a human-in-the-loop framework for recovering a single continuous yarn path from micro-computed tomography (micro-CT) scans of real crochet objects. Crochet is a craft that creates complex three-dimensional shapes by interlocking loops formed from a single yarn. Recovering the underlying yarn path from external observations is challenging because of severe self-occlusion. While micro-CT reveals the full internal structure of a crochet object, the volumetric scan alone does not explicitly encode how the yarn traverses the object. This challenge stems from the hierarchical structure of yarn: a yarn consists of multiple twisted plies, and each ply itself consists of twisted fibers. Consequently, local fiber orientations observed in micro-CT scans are not aligned with the overall yarn direction. To recover yarn-level orientations from ply-level fiber orientations, we first estimate local fiber directions using Gabor filtering and convert the volume into an oriented point cloud. We then introduce an anisotropic mean-shift procedure that aggregates local fiber orientations within a neighborhood of the yarn radius into yarn-level orientation estimates. Combined with an automatic topology skeletonization strategy, our method extracts yarn-path fragments. Subsequent fragment linking, junction cleaning, and loop detection merge these trees into a small number of long curves. Where the automatic reconstruction remains ambiguous, a sketch-based user interface enables users to interactively complete the single continuous yarn path. The recovered yarn path enables downstream applications including physically based simulation, ply-level rendering, and stitch-pattern extraction.

1. Introduction

Recovering real crochet yarn paths remains difficult because stitches form freeform, occluded structures and local fibers do not reveal the yarn trajectory. CT2Yarn combines micro-CT analysis, automatic reconstruction, and interactive editing to recover a continuous path.

1. Introduction

  • Real crochet yarn-path recovery remains largely unexplored, despite its relevance to digitizing handmade objects and supporting simulation and rendering.Prior yarn modeling largely uses procedurally generated paths for synthetic structures.Source: §1 · p. 2
  • Freeform stitch connections and severe self-occlusion make image-based reconstruction difficult, while micro-CT exposes the interior without directly encoding yarn traversal.Crochet stitches may connect to arbitrary earlier stitches rather than a regular grid.Source: §1 · p. 2, Figure 2 · p. 2
  • CT2Yarn estimates ply orientations with Gabor filtering, aggregates them into yarn-level orientations with anisotropic mean-shift, and extracts centerline candidates.A heuristic linking strategy repairs fragmented candidates, and users resolve remaining ambiguous connections interactively.Source: §1 · p. 2
  • The evaluation covers physical samples, visual and quantitative comparisons, ablations, and stitch-pattern extraction; interactive completion takes a median of four sketch edits and under three minutes per sample.The quantitative comparisons use manual annotations and a synthetic CT volume with exact ground truth.Source: §1 · p. 2
  • The work contributes two-stage yarn-orientation analysis, automatic fragment repair with sketch-based completion, and a public dataset of micro-CT scans and reconstructed yarn curves.The recovered representation supports simulation, ply-level rendering, and stitch-pattern extraction.Source: §1 · p. 2

2. Related Work

Prior yarn-level pipelines largely target regular knitted or woven structures, while crochet's freeform topology and close contacts challenge image- and CT-based reconstruction. CT2Yarn addresses real crocheted samples with a staged pipeline from CT volume to continuous centerline.

2.1. Yarn-Level Modeling of Knitted Fabrics

  • Regular row-and-column grids underpin yarn-level knit modeling, but crochet stitches can connect to arbitrary earlier stitches and violate that structural assumption.Prior work spans stitch meshes, simulation, rendering, fabrication, and interactive design.Source: §2.1 · p. 3

2.2. Computational Approaches to Crochet

  • Crochet computational work has focused on synthesis, pattern tools, fabrication, and process tracking; none of the cited works reconstructs real-sample 3D yarn structure.Storck et al. produce a continuous curve from a fixed per-stitch template rather than reconstructing a physical object.Source: §2.2 · p. 3

2.3. Image-Based Yarn and Strand Reconstruction

  • Image-based fabric methods assume regular grids, while hair methods assume visible independent strands; neither assumption fits crochet's occluded interlocking loops.The cited image-based methods address woven fabrics, knit fabrics, or hair rather than crochet yarn paths.Source: §2.3 · p. 3

2.4. CT-Based Yarn and Strand Reconstruction

  • Earlier CT yarn methods rely on regular grids, isolated yarns, or short parallel samples; CT2Hair traces separated strands, unlike crochet's contacting, single-yarn structure.Figure 4 traces CT2Yarn's stages from volume and orientations through center points and fragments to a repaired, interactively completed centerline.Source: §2.4 · p. 3, Figure 4 · p. 4

3. Method

CT2Yarn turns a micro-CT density volume into an ordered, continuous yarn centerline through orientation estimation, center-point extraction, and topology reconstruction. Automatic linking and topology checks assemble fragments; a sketch interface handles connections the automatic stage cannot resolve.

3. Method

  • The pipeline assembles the final yarn path from ordered fragments and supports downstream simulation, rendering, and stitch-pattern extraction.Its input is a micro-CT density volume, and its output is an ordered sequence of center points representing the entire yarn.Source: §3 · p. 4

3.1. Preprocessing

  • The method first thresholds each density volume with Otsu’s criterion, avoiding additional smoothing to preserve fine boundary structures.The threshold suppresses the low-intensity haze around yarn.Source: §3.1 · p. 4

3.2. Orientation Field Estimation

  • Gabor filtering estimates local fiber directions and produces a sparse oriented point cloud, whose directions are converted into yarn-level orientations by anisotropic mean-shift.Gabor kernels respond to yarn-scale ridges rather than unreliable boundaries; the smallest eigenvector of the structure tensor gives the local tangent.Source: §3.2 · p. 4

3.3. Center Point Set Extraction

  • Anisotropic mean-shift moves points toward the centerline using a tangent-elongated neighborhood and shifts only perpendicular to the local tangent.Neighbor relations are frozen early and regularized to prevent points from collapsing into dense clusters; bandwidth annealing tightens the cloud onto a thin centerline.Source: §3.3 · p. 5

3.4. Automatic Topology Skeletonization

  • Topology reconstruction orders center points into fragments, links unambiguous endpoint pairs, cleans ring and twig junctions, and uses the single-strand path constraint to resolve remaining connections.A valid path has one connected component, two endpoints, and no cycle; unresolved cases go to manual repair.Source: §3.4 · pp. 5–6

3.5. Manual Repair

  • The interactive interface lets users connect picked endpoints with an adjustable Bézier curve or a freehand screen-space stroke back-projected into 3D.After edits, loop detection is rerun to verify a single open path.Source: §3.5 · pp. 6–7

4. Experiments

The experiments evaluate CT2Yarn on real crochet scans and synthetic data, using visual comparisons, quantitative metrics, ablations, and downstream applications. Results show strong automatic reconstruction, with limited sketch editing completing single continuous paths across the real samples.

4. Experiments

  • Implementation uses physically grounded yarn-radius parameters where possible, while center-point extraction and topology parameters are chosen empirically.Chamfer distance varied within 0.13 voxels for K from 32 to 256; the authors use K=128 as a trade-off.Source: §4 · p. 7

4.1. Dataset

  • The dataset contains 18 samples across five crochet categories, scanned at full resolution using the same recycled cotton yarn.Volumes span roughly 900–2200 voxels in the in-plane axes and 300–550 slices along the scan axis.Source: §4.1 · p. 7

4.2. Metrics

  • Evaluation averages results over 18 samples and combines visual comparisons with reference-free fragmentation measures and reference-based geometry and tangent metrics.Because real samples lack ground-truth yarn geometry, reference-based metrics use paths manually annotated with the interactive editor.Source: §4.2 · p. 8

4.3. Validation on Synthetic Data

  • On synthetic data, the automatic stage recovered one continuous centerline with 1.1% total-length error and a Chamfer distance of 3.45 voxels.Deviation stayed below a third of the yarn radius along 95% of the path, growing near the free ends.Source: Figure 9 · p. 7, §4.3 · p. 9

4.4. Comparison to Baselines

  • Visual comparisons show faithful reconstructions across twelve samples, while baselines exhibit fragmentation, missing arcs, parallel-rail duplication, or overcontraction.The quantitative comparison averages results over all 18 samples.Source: Figure 10 · p. 8, §4.4 · p. 9, Table 2 · p. 9
  • All 18 samples reached a single continuous yarn path after a median of 4 sketch edits and 2:46 of editing; the automatic pipeline led every baseline on reported metrics.The automatic output averaged 15 curves per sample, while the three chain-stitch samples required no editing.Source: Figure 10 · p. 8, Table 2 · p. 9, §4.4 · p. 9

4.5. Ablation Study

  • The anisotropic kernel mattered most in ablation: replacing it with an isotropic ball increased geometry errors by roughly 40% and produced half again as many curves.Removing topology preservation left point-wise metrics almost unchanged but raised curve count, indicating its main effect is connectivity.Source: §4.5 · pp. 9–10, Table 3 · p. 9

4.6. Applications for the Extracted Centerline

  • The recovered centerline supports ply-level rendering, discrete-elastic-rod simulation, and stitch-pattern fitting against chain, slip, and single-stitch templates.Template matching slides the fitted endpoint to refine a stitch boundary and anchor the next iteration.Source: Figure 11 · p. 9, §4.6 · p. 10, Figure 12 · p. 10

5. Conclusion

CT2Yarn reconstructs a continuous yarn path from crochet micro-CT scans by combining orientation estimation, topology-preserving skeletonization, graph repair, and interactive correction. The authors identify limits for tightly crocheted samples, stitch-pattern extraction, and objects made from multiple yarns, while noting that paired scan-and-centerline data could support future work.

5. Conclusion

  • CT2Yarn recovers a single continuous yarn path from micro-CT volumes by combining Gabor-based fiber orientation estimates, anisotropic mean-shift aggregation, skeletonization, graph repair, and interactive correction.The graph-repair stage uses crochet’s single-yarn constraint to merge fragments into long sections.Source: §5 · p. 10
  • Across 18 hand-crocheted samples covering three stitch types, the recovered paths support simulation, ply-level rendering, and stitch-pattern extraction, and the authors release the scans and centerlines.Evaluation primarily used visual comparison, supplemented by quantitative metrics and an ablation study.Source: §5 · p. 10
  • Tightly crocheted samples can yield much shorter centerlines and require further manual editing because yarn-boundary contrast drops to the noise level.For the same stitch pattern, loose samples were reconstructed without edits.Source: §5 · p. 10
  • Stitch-pattern extraction is limited to approximately planar pieces and basic stitch types; extending it to non-planar layouts and a richer stitch vocabulary remains future work.Source: §5 · p. 10
  • The single-yarn constraint does not cover amigurumi, which typically use multiple yarns and fibrous filling that is difficult to distinguish from yarn in CT.Reconstruction would require reasoning about multiple strands and separating yarn from filling.Source: §5 · p. 11
  • Paired CT volumes and yarn centerlines produced with little human effort could support training data at scale for harder, lower-contrast scans and downstream applications.The authors suggest centerline-aware pattern generation and neural stitch classification as possible uses.Source: §5 · p. 11
Loading 2609.06950v1…