Making diverse fabrics usable for learning

As materials and processes combine in more ways, fabric properties become more varied. Measurement standards do not make measurement easy, and describing everything consistently remains difficult. For a model predicting KES measurement attributes, I started by making the fabric data comparable.

Bidirectional estimation between materials, processing and blend proportions, and warp/weft bending stiffness, shear stiffness, warp/weft tensile stiffness and drape. KRR feeds MLP; the forward path regresses physical properties and the reverse path uses Softmax for blend proportions.
The updated concept links materials, processing and blend proportions with physical properties in both directions. KRR feeds MLP; Softmax normalizes blend proportions to a total of 100%. — Illustration: Enzoviz — fabric research, 2026 redraw

Aligning to a common topology

The 3D data was recorded as point clouds and converted into geometry through Poisson reconstruction. Topology and data size differed between models. For learning, I needed a common representation while preserving good geometry and interpolating where needed—the central challenge in alignment and shape transfer.

From scan to training model

From 3D scans to learning data: remove invalid geometry, align in a reduced space, transfer shapes to a common topology and correct excessive deformation. Severely damaged data is excluded.
From invalid-geometry removal to common-topology transfer and deformation correction. Severely damaged data is excluded. — Illustration: Enzoviz — fabric research, 2026 redraw

Recognizing invalid geometry

Problems left by Poisson reconstruction disrupted alignment in the reduced-dimensional space. I implemented detection and removal of the following mesh defects.

  • Tunnels
  • Overlaps
  • Holes
Meshes with tunnels, overlaps and holes
Meshes with tunnels, overlaps and holes — Illustration: Enzoviz — fabric-scanning-ml

From left: tunnels, overlaps and holes.

Understanding the limits of reconstruction

If the original scan was too damaged—for example, with major losses at wrinkle peaks—interpolation could not recover the missing basis for the shape. I excluded those samples. In the data I worked with at the time, I reconstructed just over 97% of the 3D fabric models, helping expand datasets combining 2D and 3D information for training.