ModelPublicadoExterno

GaugeAnything

Masks in, millimeters out. These are the trained task heads of GaugeAnything — a promptable quantitative inspection pipeline for industrial micro-vision (SAM 3 backbone + metrology core).

Fuentehuggingface.co/James-joobs/GaugeAnything/tree/392076242ccff0cf7ae0df6712c3f4719edc20d2

TipoOtherVersiónDraem unevenCcalidad

Fuente del modelo

Tipo: Other
Tamaño del archivo: 452 KB–14,8 MB
Versión: vDraem uneven
Actualizado: 1 nov 2026
Fuente: Hugging Face

Descripción de la fuente

GaugeAnything — task heads for promptable quantitative inspection

Masks in, millimeters out. These are the trained task heads of GaugeAnything — a promptable quantitative inspection pipeline for industrial micro-vision (SAM 3 backbone + metrology core).

🌐 Project page: https://falcons-eyes.github.io/GaugeAnything/ · 📊 All numbers below are audited (held-out splits, multi-seed where applicable, protocols in the repo).

Checkpoints

File	Task	Audited result	Training data	Use
`profile_width_cnn.pt`	1-D crack-width regression from a 501-px brightness profile (the "signal for HOW WIDE" head)	table test MAE ≈18.6 μm (~1 μm GPU run variance); end-to-end promptable 39.9 μm MAE / 23.2 μm median (localization-gated)	krkCMd, 14,424 profiles (CC BY 4.0 — license-clean)	✅ commercial OK
`gaugehead_tiny_width.pkl`	Tiny owned crack-width specialist over SAM-mask/image statistics	held-out source rel.err 0.472 vs 5-bin quantile 0.480 and old neural M2 0.564; worst source still 0.720	CrackSeg9k M2 cache	⚠️ research (subset licenses vary)
`gaugehead_tiny_width_conformal.pkl`	GaugeHead-Tiny + 90% conformal interval (log cross-conformal; μ + σ-diagnostic + q)	keeps rel.err 0.4724 with per-source coverage 0.91/1.00/0.95 @90%; adaptive variants collapse on the worst source (0.21/0.11) — see repo `experiments/results/m2_uncertainty_conformal.json`	CrackSeg9k M2 cache	⚠️ research (subset licenses vary)
`m2_refiner.pt`	Measurement-aware crack mask refiner (UNet, 1.9M)	superseded baseline: a logit-threshold + quantile calibration beats it (0.437 vs 0.564 rel. err) — kept for reproducibility	CrackSeg9k train sources	⚠️ research (subset licenses vary)
`matte_fray_directional.pt`	Alpha matting head for fuzzy-boundary (fray) defects, directional synthesis v2	real MT-fray preservation IoU 0.949 vs classical guided filter 0.860	synthetic compositing over Magnetic-Tile free images	⚠️ research (MT license unstated)
`matte_fray.pt`	v1 (blob synthesis) — kept as the honest negative: real-transfer failure 0.483	see repo progress logs	same	⚠️ research
`draem_uneven.pt`	DRAEM-lite reconstruction head for boundaryless (uneven/mura) defects	test AUC 0.636 (classical illumination-residual baseline: 0.669)	synthetic mura over Magnetic-Tile free images	⚠️ research

The SAM 3 backbone is not redistributed here — get it at facebook/sam3 (separate license, gated).

Usage (profile width head)

import torch, numpy as np

ckpt = torch.load("profile_width_cnn.pt", map_location="cpu")
# architecture: see experiments/krkcmd_signal_width.py::build_1d_net in the GitHub repo
from gaugeanything_repo.experiments.krkcmd_signal_width import build_1d_net, norm_profile
net = build_1d_net(); net.load_state_dict(ckpt["model"]); net.eval()

profile = np.asarray(...)          # 501 samples of image brightness across the crack
x = torch.from_numpy(norm_profile(profile)).view(1, 1, -1)
width_um = float(net(x))            # crack width in micrometers

Full pipeline (prompt → SAM 3 localization → perpendicular profile → width) lives in the GitHub repo — see experiments/krkcmd_signal_width.py and docs/WIDTH_BOTTLENECK_ANALYSIS.md for why width is read from the signal, not from mask geometry.

Honest limitations

profile_width_cnn is trained on 6400-dpi scanner profiles of concrete (krkCMd); transfer to other resolutions/materials is not yet validated — scale-normalize inputs.
End-to-end accuracy is localization-gated: coverage 46–66% on the scanner domain; points failing the gate are reported as "not measurable", not guessed.
Heads marked research await upstream dataset license clarification before commercial use.

Citation

@misc{gaugeanything2026,
  title  = {GaugeAnything: Promptable Quantitative Inspection for Industrial Micro-Vision},
  author = {Joo, Hyunwoo},
  year   = {2026},
  url    = {https://github.com/falcons-eyes/GaugeAnything}
}

Fuentes

1 fuente

Hugging FacePrincipalVersión 392076242ccff0cf7ae0df6712c3f4719edc20d2

huggingface.co/James-joobs/GaugeAnything/tree/392076242ccff0cf7ae0df6712c3f4719edc20d2

Verificado 1 nov

Download files

Artefactos del modelo

6 artefactos

draem_uneven.pt

pt · 14,8 MB · SHA-256 a09353e0c664…5d85 · Hugging Face

Download file

m2_refiner.pt

pt · 7,40 MB · SHA-256 58d96b887e65…1d58 · Hugging Face

Download file

matte_fray_directional.pt

pt · 7,40 MB · SHA-256 536c70e7910e…e11b · Hugging Face

Download file

matte_fray.pt

pt · 7,40 MB · SHA-256 4d392e7403de…a940 · Hugging Face

Download file

profile_width_cnn.pt

pt · 679 KB · SHA-256 de5a6ad730b4…43c5 · Hugging Face

Download file

rebar_density_head.pt

pt · 452 KB · SHA-256 1af3fe7f1330…f357 · Hugging Face

Download file

Extractos de fuentes

2 extractos

Readme18 agohuggingface.co/James-joobs/GaugeAnything

--- license: apache-2.0 language: en tags: - industrial-inspection - metrology - crack-detection - segmentation - measurement pipeline_tag: image-segmentation --- # GaugeAnything — task heads for promptable quantitative inspection **Masks in, millimeters out.** These are the trained task heads of [GaugeAnything](https://github.com/falcons-eyes/GaugeAnything) — a promptable quantitative inspection pipeline for industrial micro-vision (SAM 3 backbone + metrology core). 🌐 Project page: https://falcons-eyes.github.io/GaugeAnything/ · 📊 All numbers below are audited (held-out splits, multi-seed where applicable, protocols in the repo). ## Checkpoints | File | Task | Audited result | Training data | Use | |---|---|---|---|---| | `profile_width_cnn.pt` | **1-D crack-width regression** from a 501-px brightness profile (the "signal for HOW WIDE" head) | table test MAE **≈18.6 μm** (~1 μm GPU run variance); end-to-end promptable **39.9 μm MAE / 23.2 μm median** (localization-gated) | krkCMd, 14,424 profiles (**CC BY 4.0** — license-clean) | ✅ commercial OK | | `gaugehead_tiny_width.pkl` | Tiny owned crack-width specialist over SAM-mask/image statistics | held-out source rel.err **0.472** vs 5-bin quantile 0.480 and old neural M2 0.564; worst source still 0.720 | CrackSeg9k M2 cache | ⚠️ research (subset licenses vary) | | `gaugehead_tiny_width_conformal.pkl` | GaugeHead-Tiny + 90% conformal interval (log cross-conformal; μ + σ-diagnostic + q) | keeps rel.err **0.4724** with per-source coverage **0.91/1.00/0.95** @90%; adaptive variants collapse on the worst source (0.21/0.11) — see repo `experiments/results/m2_uncertainty_conformal.json` | CrackSeg9k M2 cache | ⚠️ research (subset licenses vary) | | `m2_refiner.pt` | Measurement-aware crack mask refiner (UNet, 1.9M) | superseded baseline: a logit-threshold + quantile calibration beats it (0.437 vs 0.564 rel. err) — kept for reproducibility | CrackSeg9k train sources | ⚠️ research (subset licenses vary) | | `matte_fray_directional.pt` | Alpha matting head for fuzzy-boundary (fray) defects, directional synthesis v2 | real MT-fray preservation IoU **0.949** vs classical guided filter 0.860 | synthetic compositing over Magnetic-Tile free images | ⚠️ research (MT license unstated) | | `matte_fray.pt` | v1 (blob synthesis) — kept as the honest negative: real-transfer failure 0.483 | see repo progress logs | s...

Note18 agohuggingface.co/James-joobs/GaugeAnything

Source context: 0 downloads · 0 likes · Pipeline image-segmentation · Repo James-joobs/GaugeAnything