BRENNEN

Industries / Consumer goods / Textiles and apparel / Leather goods

Cutting

Cutting is laying out the pattern pieces of a product on each hide, around its scars, holes and thin areas, and cutting them out with a hand knife, with a steel die (a blade bent to the shape of the piece) pressed through the leather, or with a computer-guided blade or laser.

4

Milestones achieved, from 2012 to 2021.

1

To watch for: announced, not yet confirmed outside.

2

Open deliverables. Nobody has shipped them.

The road

AchievedWatch forOpen

Achieved

  1. 2012

    A laser cutter for leather

    Lectra, a French maker of cutting systems, introduced its Versalis digital laser cutter for leather in 2012, aimed at the cutting then done with die presses. Five years later WardsAuto reported that 90% of leather cutting was still done with die presses, and that Lectra's car-industry business had grown 60% since 2012.

    wardsauto.com

    Credit

    Lectra

  2. 2014

    Nesting around the flaws

    Roberto Baldacci, Marco Boschetti, Maurizio Ganovelli and Vittorio Maniezzo of the University of Bologna published methods for nesting, the placing of pattern pieces on a hide by software, when the hide has defects, where some flaws may fall inside certain kinds of piece and others must be avoided. The problems came from the leather garment and furniture trades, and the paper appeared in Discrete Applied Mathematics on 30 January 2014.

    elmi.hbku.edu.qa

    Credit

    Baldacci, Boschetti, Ganovelli and Maniezzo

  3. 2019

    A public test for flaws

    MVTec Software, a German machine-vision company, released MVTec AD, more than 5,000 high-resolution photos in 15 categories, leather among them, with defect-free images for training and test images with every defect marked, so that methods for finding flaws could be compared on the same pictures. It is free for research under a non-commercial licence, CC BY-NC-SA 4.0.

    mvtec.com

    Credit

    MVTec Software

  4. 2021

    The leather test runs out

    Karsten Roth, Bernhard Schölkopf, Thomas Brox, Peter Gehler and colleagues reported that their method, PatchCore, scored 100.0 on MVTec AD's leather category by the usual measure (AUROC, a score of how well a method ranks flawed images above good ones, where 100 is perfect). An earlier method in their comparison, MahalanobisAD, had also scored 100.0 on leather.

    LessonOnce methods score 100, the test can no longer tell them apart, and it says nothing about grading whole hides for cutting.

    arxiv.org

    Credit

    Roth and colleagues

Watch for

  1. Inspection in 14 seconds

    CORIUM says its G52 machine is the industry's first AI-driven automatic leather inspection machine: it photographs both sides of a hide at up to 40 micrometres per pixel, maps defects by type, size and seriousness, marks the hide itself, and handles one hide every 14 seconds, loading and unloading included.

    Achieved whena tannery or leather goods maker with no stake in CORIUM publishes how often the machine's defect maps and grades agree with its own graders on a few hundred hides.

    mpastyle.it

    Claimed by

    CORIUM, in MPA Style, October 2025

Open

  1. 01

    A public hide-nesting test

    A free set of at least 200 real hides, each traced as an outline with its flaws drawn as shapes and marked by quality zone, plus five standard piece sets (a card wallet, a belt, a tote, a backpack and a watch strap) and a scoring script that reports the share of each hide's area a nesting program uses. Run three free nesting programs on it and publish their yields beside a human cutter's on 20 of the hides.

    The brief
    Credit · unclaimed
  2. 02

    Whole-hide flaw photos, open

    At least 1,000 photos of whole finished hides, the kind a bag maker cuts from, with every flaw outlined and named (scar, hole, insect bite, wrinkle, stain, loose grain) by two trained graders, released for any use, with a baseline model and a measure of how often the two graders agree.

    The brief
    Credit · unclaimed

Open 01 · the brief

A public hide-nesting test

Credit · unclaimed

The first to publish is credited on this line, and the deliverable moves from Open to Achieved.

What to build

A free set of at least 200 real hides, each traced as an outline with its flaws drawn as shapes and marked by quality zone, plus five standard piece sets (a card wallet, a belt, a tote, a backpack and a watch strap) and a scoring script that reports the share of each hide's area a nesting program uses. Run three free nesting programs on it and publish their yields beside a human cutter's on 20 of the hides.

Why it is still open

Nesting research on real hides, such as Baldacci and colleagues' 2014 methods for hides with flaws, works on problems from the authors' industry partners, and makers of cutting systems publish only their own savings figures. A search found no public set of hide outlines with flaw maps that anyone can use to compare programs.

elmi.hbku.edu.qa

Built with what exists

A camera on a stand over a light table; OpenCV to trace outlines; Label Studio to draw the flaws; the free nesting programs SVGnest and Deepnest; Python to score yields.

  1. First step

    Photograph and trace 20 hides at a leather wholesaler, and lay one wallet pattern on each by hand to set the template.

  2. Done looks like

    200 hides from at least three tanneries in two kinds of leather, released under a licence that allows commercial use, with yields for three programs and one human cutter, and the code to rerun them.

Open 02 · the brief

Whole-hide flaw photos, open

Credit · unclaimed

The first to publish is credited on this line, and the deliverable moves from Open to Achieved.

What to build

At least 1,000 photos of whole finished hides, the kind a bag maker cuts from, with every flaw outlined and named (scar, hole, insect bite, wrinkle, stain, loose grain) by two trained graders, released for any use, with a baseline model and a measure of how often the two graders agree.

Why it is still open

The common public benchmark, MVTec AD, bars commercial use, and methods already score 100 on its leather photos. The AI graders sold today are trained on private photos: Zünd says the network behind its grading workflow learned from tens of thousands of images and detects more than 20 types of defect.

mvtec.comarxiv.orgzund.com

Built with what exists

A camera rig over a table with low-angle LED lights; Label Studio; Meta's Segment Anything model and Ultralytics YOLO to speed up outlining and as baselines; a vision-language model such as Claude Fable 5.2 to draft labels for the graders to correct (see the Note "Three working devices").

  1. First step

    Ask one tannery for leave to photograph 50 finished hides that its grader has already marked.

  2. Done looks like

    1,000 hides from at least five tanneries, under CC BY 4.0, with grader agreement and the baseline's precision and recall reported for each kind of flaw.