Promptable quantitative inspection for industrial micro-vision — masks in, millimeters out.
Foundation models tell you what and where. GaugeAnything tells you how many millimeters, how many instances, and which condition grade — the metrology that actually drives field decisions (crack width, defect size, part spacing), emitted as first-class outputs on top of a promptable segmentation backbone (SAM 3).
Prompt
"crack"→ SAM 3 segmentation → skeleton + EDT width profile → mean width 1.53 mm, length 212 mm, on a real concrete surface, zero-shot. (Assumed scale 0.25 mm/px — real-metric capture with ArUco markers is the next data milestone; see Metrology Rigor.)
📄 Paper: PDF (arXiv pending endorsement) · 🤗 Weights: James-joobs/GaugeAnything
Every "Anything" model stops at perception: Segment Anything gives masks, Depth Anything gives relative depth, Count Anything counts. None emits a measurement. Industrial inspection needs "crack width 0.42 mm ± 0.05, condition: Fair" — and as of mid-2026 no foundation model produces that output. GaugeAnything fills the gap with a metrology core that is honest about its own rigor.
gaugeanything/ # the metrology core library
├── segmenters.py # SAM 3 adapter (+ prompt-set ensemble vs. synonym collapse) + classical fallbacks
├── geometry.py # mask → width profile (skeleton+EDT), equivalent diameter, spacing
├── scale.py # pixel→mm: ArUco / bolt-head specs / PlaneScale homography (tilt-robust)
├── soft.py # soft inspection: illumination-residual (mura), guided matting, Σα measurement
├── router.py # regime router: sharp→binary · fuzzy→matting · field→illumination model
├── pipeline.py # inspect(): image+prompt → Inspection Atoms {mask, count, mm±σ, grade}
├── selftest.py # metrology self-tests (14/14, synthetic GT)
└── soft_selftest.py # soft-measurement math self-tests (11/11)
README.md # this document — the project's single detailed entry point
huggingface/README.md # model card for the released weights
docs/static/pdfs/ # the released paper PDF
This repository is intentionally minimal: the metrology core library, this README, the released paper PDF, and a few result figures. Trained weights live on Hugging Face, datasets are obtained from their official sources (below), and the SAM 3 backbone is gated upstream — nothing internal (servers, audit logs, research scripts) is published here.
| Resource | Where | |
|---|---|---|
| 🧩 | Source (this library) | github.com/falcons-eyes/GaugeAnything |
| 🤗 | Trained weights + model card | huggingface.co/James-joobs/GaugeAnything |
| 🧠 | SAM 3 backbone (gated, not redistributed) | huggingface.co/facebook/sam3 |
| 📄 | Paper (PDF) | gaugeanything.pdf |
git clone https://github.com/falcons-eyes/GaugeAnything.git && cd GaugeAnything
pip install -e . # metrology core (CPU)
pip install -e ".[gpu,bench]" # + SAM 3 backbone & benchmarks
# verify the metrology core (no model weights needed)
python -m gaugeanything.selftest # 14/14: width ±10%, ArUco scale ±5%, e2e ±15%
python -m gaugeanything.soft_selftest # 11/11: soft-area / severity / uncertainty math
# measure something (requires SAM 3 access: accept license at hf.co/facebook/sam3, `hf auth login`)
python - <<'PY'
import numpy as np
from PIL import Image
from gaugeanything import inspect
img = np.array(Image.open("your_crack_photo.jpg").convert("RGB"))
res = inspect(img, "crack", segmenter="sam3", marker_size_mm=20.0) # ArUco 20mm in frame → mm output
print(res.summary())
PYAll numbers were measured on a single modern GPU with the same audit rules used throughout the project: empty-GT excluded, multi-seed reporting where applicable, config selection on validation only, and held-out test sources. Internal audit logs are kept out of the public repository.
Crack segmentation, zero-shot (CrackSeg9k, crack-only, 3 seeds):
| Method | crack mIoU (±std) | non-crack clean rate |
|---|---|---|
| frangi (classical) | 0.115 ± 0.005 | 0.26 |
| adaptive (classical) | 0.181 ± 0.006 | 0.00 |
| SAM 3 zero-shot | 0.442 ± 0.011 | 0.68 |
2.44× the best classical baseline — and it also wins detection. (Supervised U-Nets reach ~0.7+ on this benchmark; the claim is promptability, not SOTA.)
Segmentation ≠ measurement — every method under-estimates crack width (GT 11.3 px):
| Method | width MAE (px) ↓ | width rel. err ↓ |
|---|---|---|
| adaptive | 6.67 | 43.5% |
| SAM 3 | 5.67 | 62.9% |
Best mIoU is not best measurement — the core motivation for measurement-aware refinement.
Boundary regimes — binary segmentation collapses to chance on fuzzy/boundaryless defects (IoU ≤ 0.03, AUC ≈ 0.50); continuous representations recover signal (val/test protocol):
| Defect | SAM 3 binary | classical soft (test) | learned (test) |
|---|---|---|---|
| Uneven (field) | 0.499 | 0.669 | 0.636 (DRAEM-lite) |
| Fray (fuzzy edge) | 0.526 | 0.644 | — |
Metrology rigor — two silent measurement killers, quantified and fixed:
| Failure mode | naive | fixed |
|---|---|---|
| Camera tilt 50° (scale error) | 19.3% | 0.7% (PlaneScale homography) |
| Prompt synonym collapse ("fracture"/"pit") | mIoU 0.000 | 0.374 / 0.352 (prompt-set ensemble) |
Dynamic / uncontrolled scenes — first evidence that the metric signal survives moving cameras:
| Track | Data | Result |
|---|---|---|
| TUM handheld checkerboard | 160 gated frames | 1.06% / 2.60% median/p90 relative error |
| ADT egocentric walkthroughs | 2 sequences, 480 frames, 229 objects | 8.7% median 3-D dimension error; 9.1% in the 0.5m/s+ speed bin |
ADT is an oracle-depth upper bound using GT object volume/pose gates, not SAM3 promptable performance yet. The ROI-only negative control collapses to 316% median error, so the next real model problem is replacing the oracle gate with segmentation or promptable masks.
Honest negative results (we publish these too): a matting head that wins 20× on synthetic fuzzy boundaries failed to transfer to real fray (mask IoU 0.48 vs guided filter 0.86) — synthetic blob distribution ≠ directional texture. Production keeps the classical guided filter; the learned head ships only after real-distribution synthesis passes.
This project audits itself before reviewers do: no test-set tuning (val/test splits), empty-GT separated from IoU, multi-seed reporting, prompt-sensitivity sweeps, checkpoints saved for every trained artifact, and negative results documented internally before public release.
- Metrology core (width / diameter / spacing / severity / uncertainty) + self-tests
- SAM 3 integration + classical baselines + cross-source benchmark
- Regime router (sharp / fuzzy / field) + soft measurement
- PlaneScale (tilt-robust mm) + prompt-set ensemble
- Measurement-aware refinement head (M2 v1) — superseded: a 5-number quantile calibration beats it (0.480 vs 0.564 rel. err, held-out); M2 v2 bar = 0.480 + per-source worst-case
- Real-metric substitutes — coins LOO 1.74% real-photo consistency; krkCMd profile-level crack width MAE 27.8±2.5μm (5-fold; single split 25.9) vs manual GT
- Dynamic metric evidence — TUM handheld 1.06% gated error; ADT oracle RGB-D multiview 8.7% over 2 sequences/480 frames/229 objects
- Owned measurement head (M2 v2-a) — GaugeHead-Tiny rel. err 0.472, first learned rung past the quantile bar (0.480)
- Uncertainty audit (M2 v2-b) — 90% conformal intervals keep 0.4724 rel. err with per-source coverage 0.91/1.00/0.95; adaptive variants collapse on the worst source (0.21/0.11) and no difficulty signal flags it (concept shift — honest negative)
- Real-metric ground truth capture (ArUco/caliper field protocol)
- Counting & spacing validation (fastener datasets) — next: ROI-1555 density/centroid head (Count v1, target MAE < 5)
- HuggingFace weights release — https://huggingface.co/James-joobs/GaugeAnything (task heads + audited model card)
- Paper draft v2 — PDF on the project page (arXiv submission pending endorsement; v2 adds owned-model ladder rung, conformal audit, dynamic-scene section)
- Code: Apache-2.0.
- SAM 3 weights: separate SAM License (commercial use permitted with restrictions) — gated on HuggingFace, not redistributed here.
- Trained task heads: released on Hugging Face (model card lists per-head licenses); not stored in this repo.
- Datasets are obtained from their official sources, each under its own license: CrackSeg9k, krkCMd (crack-width profiles, CC BY 4.0), T-LESS / BOP, Magnetic-Tile-Defect, TUM RGB-D, Aria Digital Twin. This repository ships no dataset files.
@misc{gaugeanything2026,
title = {GaugeAnything: Promptable Quantitative Inspection for Industrial Micro-Vision},
author = {Joo, Hyunwoo},
year = {2026},
url = {https://github.com/falcons-eyes/GaugeAnything}
}Part of the Industrial Anything research program. Contributions welcome — especially real-world measurement ground truth (photos with ArUco markers + caliper readings), fastener/counting datasets, and regime-router edge cases. Open an issue first for anything substantial.