This dataset is released for research use. Access is reviewed and granted manually by the maintainers. Please state your name, affiliation, and intended use.

Roles

Roles: canon repo — annot is the source label, kept machine-parseable as the gold for verification and reward parsing; there is no reasoning column and this repo is not itself a training view. Derived repos (-annotated, -grounding, -region, -mcq) each state their own regime on their own card. Geometry for every record lives in metadata.geometry (below).

210

Outdoor maintenance-inspection anomaly detection over 7 asset scenarios (SYNTHETIC, 3D-rendered; binary 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

105,000 records (test=35000 · train=70000). 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); for multi-image rows, a preview of the first view
`images`	list[Image]	(multi-image rows) all input views / modalities for the row, bytes embedded
`annot`	str	the answer — for this dataset: plain-text `{label, defect_type}` — `{good, null}` or `{anomalous, <defect>}`, the defect name from THAT scenario's own closed set (enumerated in the query), following D20/D22. The binary `mask` column is deferred localization GT
`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]	(multi-image T-B1 / D21) per-view masks aligned with `images` (None where a view has no defect), or multi-region masks

Task, mask & split

What this is. MIAD — Maintenance Inspection Anomaly Detection (Bao, Chen, Li, Wang, Fei, Wu, Zhao, Zheng, arXiv:2211.13968; ICCV 2023 Workshop) — 105,000 512x512 images across 7 outdoor maintenance-inspection scenarios: catenary_dropper, electrical_insulator, metal_welding, nut_and_bolt, photovoltaic_module, wind_turbine, witness_mark. The design is perfectly regular: every scenario ships 10,000 good training images and a 5,000-image test split (2,500 good + 2,500 defective) with 2,500 pixel masks.

These are outdoor assets in service — overhead lines, turbine blades, PV modules — not factory production-line parts, which is what most of this corpus holds.

⚠ THE IMAGERY IS SYNTHETIC. MIAD is generated with 3D graphics software, not photographed. That is the point of the dataset — it buys free variation in viewpoint, weather and lighting together with exact pixel ground truth — but a model trained on it learns rendered appearance, and any claim about real-world transfer needs a real-image test set. Every record carries metadata.synthetic = true so a training mixture can weight or exclude it. The only other synthetic member of this corpus is 182 (Eyecandies); everything else is photographic.

Task & answer. Anomaly detection with defect naming. query is our own template (the source ships no natural-language question): it names the asset and asks whether it is good or anomalous. On the four scenarios whose defect set has two or three members it also asks for the defect type from that scenario's closed set (enumerated in the query). On the three single-type scenarios the closed-set question is not asked — see below. annot is plain text {good, null} / {anomalous, <defect>} on every record, and the gold carries its type token under both forms.

Defect vocabularies differ per scenario, and three are effectively binary. electrical_insulator (broken), wind_turbine (crack) and witness_mark (looseness) have exactly one defect type; metadata.single_defect_type marks those 45,000 records. The four richer scenarios are catenary_dropper (broken / looseness / miss), nut_and_bolt (looseness / missbolt / missnut), photovoltaic_module (broken / foreign_body / miss) and metal_welding (weld_beading / weld_pit).

Single-type scenarios: the closed-set type question is not asked there (2026-09-17). With a set of one the query hands the answer over, and the type slot's accuracy is 100% by construction however the sentence is worded — measured on the previous revision, all 45,000 of those records' queries named their own lone type. So those 45,000 queries now ask for the verdict, and ask the defect type to be named from what is seen rather than chosen from a list. What does not change: the answer form stays {label, defect_type} and the gold stays {anomalous, broken} / {anomalous, crack} / {anomalous, looseness}. This repo has no reasoning column, so annot is the output-format target and the query must request the form annot holds (see Roles above); a bare verdict ask over a two-slot gold would be a query/gold mismatch. All three are real names — each states a kind or a state that the verdict does not — so the type slot is kept; a scenario whose only "type" merely restated the verdict would drop the slot instead. Type accuracy is scored only where the scenario's defect set has two or more members. This rule is why this revision's query differs from the previous one on exactly those 45,000 records, and on no others.

Mask (deferred GT). Binary {0, 255} masks are embedded in the mask column for all 17,500 defective test images; good images carry mask = null.

Lazy-baseline floor. The test split is exactly balanced — 17,500 good vs 17,500 anomalous — so the binary majority floor is 50.0%, and the full {label, defect_type} floor is also 50.0% (answering {good, null} every time). This is one of the cleanest floors in the corpus, a consequence of the synthetic design.

Query text — pooled paraphrases (v2)

This repository ships 2 question forms over the same images, and each draws from its own pool in common/vision_query_pools.json (metadata.query_template is the index within that form's pool; metadata.query_pool says which form a record is):

F2a/label_type — 60,000 records, 39 gate-verified paraphrases (39 in use, top share 2.7%); template 1 is v1's wording byte for byte.
F2b/label_type_open — 45,000 records, 36 gate-verified paraphrases (36 in use, top share 3.0%); this form has never been published before, so it has no earlier wording to reproduce and every template in its pool was gated as new.

The opening role sentence is drawn separately (metadata.query_role, a 10-way hand-written pool _role/sentence; index 0 is this repository's own sentence, index 1 is none); the subject sentence is this repository's own, verbatim, on every record. Role and ask are hashed independently.

Both pools clear the 30-variant floor on their own (39 and 36 gate-verified paraphrases), so neither form's diversity rests on the other's count. The two are separate index spaces: metadata.query_pool is stamped on every record because a template index alone does not say which pool it indexes — 1179 of the re-drawn records land on the same index NUMBER in the new pool as they carried in the old.

Template ↔ gold independence on this build: 105,000 records, 75 templates, worst template p = 0.00899, alpha 1.3e-04, 0 flagged; 10 roles, worst role p = 0.156, 0 flagged → PASS.

Frame-size floor (common/lazy_floors.py, the standing (width, height)-only row): vacuous by construction — all 105,000 images share one frame size.

image, mask, annot, reasoning, cate, task and the split are byte-identical to the previous revision — this revision was issued from the published parquet itself (tools/requery_repool.py re-draws the text, tools/requery_stream.py --push carries every other column out of the live shard). What moved: query on 45,000 of 105,000 records, and metadata on 105,000 (the added query_pool key; query_template on 43,821). The image identities in §8 were carried from the previous pass and re-measured from the metadata.pixel_sha256 this repository already ships — no image was decoded again, because none was touched.

Provenance

Underlying dataset: MIAD (Maintenance Inspection Anomaly Detection). Upstream license: CC BY-NC-SA 4.0 (non-commercial) (this card is license: other; respect the upstream terms). Converted read-only from the raw source into the unified schema; conversion code under 210/ (with publish/push_to_hf.py) in AI4Manufacturing/forge_model.

Overlap / de-duplication (§8)

No overlap with any other dataset in this corpus. ⚠ The imagery is 3D-rendered, not photographed — see the synthetic note below.

Two identities, and they answer different questions. metadata.image_sha256 hashes the file bytes: it finds byte-identical copies and is blind to a re-encode. metadata.pixel_sha256 hashes the decoded image (mode | size | pixels): it finds the same photograph saved twice. Only the second one settles whether an image is duplicated.

Carried forward and re-measured, not decoded again. This revision changed text columns only, so no image was touched; the identities below are the ones this repository already ships in metadata.pixel_sha256, re-counted from them here and asserted equal to the pass that decoded them (revision 47639ac11ecd). A disagreement aborts the build and names the offending records:


images checked	105,000
distinct by decoded pixels	104,996
images carrying more than one record	1
images on both sides of the split	1

⚠ This dataset declares a exempt image-identity policy, so the row above is expected to be non-zero: UPSTREAM (MIAD release): five records share ONE near-black 512x512 frame (pixel values 0/1, mean 0.42/255) — test/good/000476.png and test/good/002343.png of wind_turbine, train/good/008548.png of wind_turbine, train/good/006127.png and train/good/008630.png of nut_and_bolt — so the same blank render is published under two categories, on both sides of the split, labelled good five times. Every record is kept exactly as published; the next data revision should drop all five (a black frame carries no evidence for a 'good' verdict) and any carving must keep the group on one side. Recorded for the next data revision Images are still forbidden from crossing the split — and that rule too is exempted here, which is why the last row may be non-zero.

Geometry (`metadata.geometry`)

Every record carries a geometry block inside the existing metadata JSON string, so that its gold can be re-derived at any render size. No schema column changed; existing loaders are unaffected.

Coordinates are native pixels of the image in that record (coords_frame: "record_image"). scale is 1.0 throughout — this repo publishes at source resolution, nothing was downscaled at publish time.

"geometry": {
  "image_wh":  [W, H],        // dims of the image in THIS record
  "source_wh": [W, H],        // dims of the original source image
  "scale": 1.0,               // image_wh / source_wh; < 1.0 would disclose a publish-time downscale
  "n_instances": 2,
  "instances": [
    { "instance_id": 1, "bbox_xywh": [x, y, w, h], "min_side_px": 65, "class": null }
  ],
  "n_dropped_subminimum": 0,  // components removed by the filters below
  "union_box_fallback": false,// true => boxes are per-class unions, NOT real instances
  "conventions": { ... }      // see table
}

instances is present even when empty. [] means the record genuinely has no defects; an absent block would mean geometry could not be recovered. Those are different states and are never conflated.

Conventions used to derive it

There is no universal definition of "one defect instance" — it depends on the mask the source shipped. This repo's is stated, not implied:

field	value
`algorithm`	`dilate_cc`
`binarisation`	`gt:0`
`connectivity`	`4`
`merge`	`mask_dilate:1pct`
`min_area_px`	`15`
`max_instances`	`None`
`artifact`	`fine`
`fill_floor`	`None`
`legibility_floor_px`	`None`
`min_side_floor_px`	`None`
`spec_sha`	`41d7fab342f60aca`

Provenance and verification


records	105,000
carrying a geometry block	105,000 / 105,000
instances per record	`0`: 87,500, `1`: 11,395, `2`: 4,143, `3`: 958, `4`: 606, `5+`: 398
total instances	27,172
image dimensions	512×512 (105,000)
`scale` values present	[1.0]

Computed from this repo's own masks and verified against this repo's own published answers before it was written — a recomputation that disagreed with the shipped gold would have aborted the update rather than overwritten it.

⚠ The 16px floor applies at the RENDER, not at native

min_side_px is in native pixels. The model does not see native: Qwen2-VL caps by megapixels AND snaps each dimension to a multiple of 28. So min_side_px >= 16 is the floor tested in the wrong frame. Measured on this repo:


native → rendered (qwen2_vl @ 2.36MP)	512×512 → 504×504
shipped boxes	27,172
legible at that render (>=16px there)	18,339 (67.5%)

⚠ An earlier version of this section reported the inverse — boxes clearing 16px at native and failing at the render — and that number was misleading. It is frame-relative: publishing at a larger native size lets more boxes clear 16 in the published frame, so more can "fail", which penalises exactly the choice that helps. Measured on 179: publishing native (3024) means a box needs >=32px native to be legible at the render and 86.7% qualify; the previous 1024 publish needed >=47px native and only 69.5% qualified. The native republish improved rendered legibility by 17 points while the old metric scored it as 12.5% "broken". The figure above is the comparable one.

Nothing in the data is frame-dependent — geometry is native and complete. Use forge_model/common/adapt_engine.py, which applies the floor at whatever size the consumer renders.

Using it

Coordinates only stay correct if they are rescaled with the image. A patch-based VLM does not render at native size: Qwen2-VL's processor snaps both dimensions to a multiple of 28, so this repo's 512×512 is rendered 504×504 and native-pixel boxes are then wrong by a few pixels. forge_model/common/adapt_engine.py regenerates coordinates for a target render size, re-derives counts, and drops records whose gold no longer holds there.

Downloads last month: 130

Total file size:

31 GB

Paper for AI4Manufacturing/210

MIAD: A Maintenance Inspection Dataset for Unsupervised Anomaly Detection

Paper • 2211.13968 • Published Nov 28, 2022

You need to agree to share your contact information to access this dataset