A crocheted object is a topological curiosity: a three-dimensional form built from a single continuous strand of yarn, looped through its own previous loops until it holds a shape. Unlike knitting, which carries many live stitches at once, crochet advances one stitch at a time along one thread. The whole object is one line.
Recovering that line from a finished object is the problem CT2Yarn solves. The paper, by Chang Luo and Nobuyuki Umetani, was submitted to arXiv on September 7, 2026 and accepted to Pacific Graphics 2026.
The obstacle is self-occlusion
You cannot photograph your way to the answer. The yarn path is tightly packed and passes repeatedly through and behind itself; from any exterior view, most of the strand is hidden by the rest of the strand. Surface scanning gives you the shape and nothing about the structure.
So the authors scan the objects with micro-computed tomography — X-ray CT at fine enough resolution to resolve individual fibres inside the volume. That gets you past occlusion entirely, at the cost of handing you a dense noisy voxel grid with no notion of what a stitch is.
From voxels to a strand
The pipeline the paper describes:
- Gabor filtering estimates local fibre direction throughout the volume, converting raw density into an oriented point cloud — points that each carry a sense of which way the fibre runs there.
- An anisotropic mean-shift procedure aggregates those fibre-level orientations upward into yarn-level estimates, collapsing many fibres into the ply they belong to.
- Automatic topology skeletonization, followed by fragment linking, junction cleaning, and loop detection, assembles those local estimates into long continuous curves.
Then the honest part. Where the automatic reconstruction is genuinely ambiguous — and in a densely packed crochet object it will be, because two strands passing at a junction can be locally indistinguishable — a sketch-based user interface lets a person draw in the correct continuation and complete the single yarn path by hand.
Calling the system “human-in-the-loop” rather than pretending to full automation is the right design, and the right claim. The algorithm handles the volume; the person handles the handful of places where the physics doesn’t disambiguate.
What you can do once you have the path
The recovered strand is not an end in itself. The paper points at three downstream uses:
- Physics-based simulation — a yarn-level curve is exactly the input a cloth or yarn simulator wants, so you can simulate how the real object deforms.
- Ply-level rendering — photorealistic rendering of textiles is bottlenecked on having accurate fibre geometry, which is normally synthesized rather than measured.
- Stitch-pattern extraction — recovering the instructions. Given the path, you can work out the sequence of stitches that produced it, which means reading a pattern off a finished object that nobody wrote down.
That last one is the quietly remarkable capability. Craft traditions transmit through hands and demonstration, and enormous amounts of technique exist only in finished objects in museum collections and in the homes of people who are no longer around to explain them. A pipeline that turns an object back into its instructions is, among other things, a preservation tool.
The broader point
Computational fabrication research usually runs forward: design a thing, then compute how to make it. CT2Yarn runs backward — take an existing handmade object and recover the process. There is not much work in that direction, and there should be more of it, because the world contains a great many objects whose making is better documented in the object than anywhere else.