| --- |
| tags: |
| - smart-manufacturing |
| - sft |
| - industrial |
| - vision |
| license: other |
| pretty_name: "181" |
| extra_gated_fields: |
| Name: text |
| Affiliation: text |
| Intended use: text |
| extra_gated_prompt: >- |
| This dataset is released for **research use**. Access is reviewed and granted |
| **manually** by the maintainers. Please state your name, affiliation, and intended use. |
| --- |
| |
| # 181 |
|
|
| Textured-surface weakly-supervised defect detection (10 classes; weak masks). Category **B**, task **T-B1**, in the unified Smart-Manufacturing SFT schema. |
|
|
| > The repository name is an internal task code. See **Provenance** below for the underlying dataset. |
|
|
| ## Records |
|
|
| **16,100** records (test=8050 · train=8050). Pixel masks are embedded as a `mask` image column. |
|
|
| ## Unified SFT schema |
|
|
| | field | type | meaning | |
| |---|---|---| |
| | `query` | str | the question / instruction (model input) | |
| | `image` | Image | the input image (bytes embedded) | |
| | `annot` | str | the answer — for this dataset: the plain-text image-level label `good` or `anomalous` (binary per texture class; the defect is unnamed). The WEAK elliptical segmentation `mask` is deferred localization GT, with seg info (`mask_path`, `defect_area_fraction`) in `metadata`; `metadata.category` is the texture class (Class1-10) — see **Task, mask & split** below | |
| | `reasoning` | null | no native CoT in these datasets | |
| | `cate` | "B" | SFT category | |
| | `task` | "T-xx" | unified task id | |
| | `metadata` | str (JSON) | split, provenance, `image_path`, `image_sha256` (dedup key) | |
| | `mask` | Image \| null | *(T-B1/T-B2 only)* the pixel ground-truth mask, bytes embedded | |
| | `masks` | list[Image] | *(D21 only)* multi-region masks | |
|
|
| ## Task, mask & split |
|
|
| **What this is.** DAGM 2007 (Wieler & Hahn, *"Weakly Supervised Learning for Industrial Optical Inspection"*, |
| DAGM 2007 competition / Bosch) — 16,100 synthetic grayscale images across **10 statistically-textured surface |
| classes** (Class1-10). Each class carries a single (unnamed) defect type; most images are defect-free. |
|
|
| **Task & label.** Originally weakly-supervised image-level defect classification (defective vs defect-free), now |
| widely used for anomaly detection + weak localization. Per texture class the task is **binary**. An image is |
| labelled **anomalous** iff the source ships a `Label/<id>_label.PNG` mask for it (defect-free images have no |
| mask). `query` (our template) names the texture class and asks only whether it is **good** or **anomalous**; |
| `annot` is the plain-text answer `good` or `anomalous`. **The query does not ask for a mask.** `metadata.category` |
| records the texture class (Class1-10). |
|
|
| **Mask (WEAK label — deferred GT).** DAGM's masks are **weak labels**: a rough elliptical region around the defect |
| (0/255), **not pixel-precise**. Kept in the `mask` column as deferred localization GT (anomalous images only; good |
| = null); seg info (`mask_path`, `defect_area_fraction`) in `metadata`. Segmentation is deferred (a text model |
| can't emit a pixel mask). |
|
|
| **Split.** Official per-class `Train`/`Test` -> `train` (8,050) + `test` (8,050) = 16,100. Class1-6: 575 images |
| per split; Class7-10: 1,150 per split. |
|
|
| ## Provenance |
|
|
| Underlying dataset: **DAGM2007**. Upstream license: **CC BY 4.0** (this card is `license: other`; respect the upstream terms). Converted read-only from the raw source into the unified schema; conversion script: `181/convert_d81.py`, published with `publish/push_to_hf.py`, both in [`AI4Manufacturing/forge_model`](https://github.com/AI4Manufacturing/forge_model). |
|
|
| ## Overlap / de-duplication (§8) |
|
|
| None notable. Each record carries `metadata.image_sha256` so overlapping images can be kept entirely on one side of a train/eval split. |
|
|
