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).

211

Multi-modal connector/component anomaly detection, 20 categories, official 3-way split (2D + pseudo-3D view of a 2D/3D release). 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

10,874 records (test=4450 · train=4000 · validation=2424). 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, <type>}` over the category's own closed set (enumerated in the query), following D20/D22. Ten defect types across the release: scratch, pit, impact_damage, deformation, hole, damage, porosity, oil_stain, exposed_surface, metallic_contaminant. Each record's `images` holds [RGB, photometric-stereo render]; the 2D `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, modalities, mask & split

What this is. Real-IAD D3 (Zhu, Wang, Zhou, Wang, Pan, Zhang, Chen, Cheng, Gao, Zhang, Gan, Wang, Chen, Qian, Chi, Peng, Ma, "Real-IAD D3: A Real-World 2D/Pseudo-3D/3D Dataset for Industrial Anomaly Detection", CVPR 2025), copyright Rongcheer Industrial Technology (Suzhou) Co., Ltd., CC BY-NC-SA 4.0. 10,874 captures over 20 connector / electronic-component categories — audio_jack_socket, common_mode_filter, connector_housing_female, crimp_st_cable_mount_box, dc_power_connector, ethernet_connector, ferrite_bead, fork_crimp_terminal, fuse_holder, headphone_jack_socket, humidity_sensor, knob_cap, lattice_block_plug, lego_pin_connector_plate, lego_propeller, limit_switch, miniature_lifting_motor, power_jack, purple_clay_pot, telephone_spring_switch.

Unlike plain Real-IAD, D3 ships an official three-way split, preserved here: train 4,000 / validation 2,424 / test 4,450. Ten defect types across the release: scratch (1,350), pit (1,249), impact_damage (1,150), deformation (450), hole (350), damage (300), porosity (225), oil_stain (150), exposed_surface (75), metallic_contaminant (75).

⚠ THIS REPO CARRIES THE 2D MODALITIES ONLY — a deliberate, stated cut. Each capture in the source names four co-registered files: an RGB image, a photometric-stereo (PS) render, an XYZ .tiff height map and a .pcd point cloud. The XYZ and PCD are roughly 95% of the release's 277 GB and cannot be read by a vision-language model, so this repo embeds RGB + PS + the 2D mask (images = [RGB, PS]; every record has both) and preserves the upstream-relative paths of the 3D files in metadata.xyz_path_upstream, pcd_path_upstream and voxel_mask_path_upstream. metadata.modalities_embedded states exactly what each record carries. This is the 2D/pseudo-3D view of Real-IAD D3, NOT a substitute for the source release if you are doing 3D anomaly detection — fetch the gated source for that.

⚠ 64 upstream entries name no image at all. The JSONs list 10,938 entries, but 64 of them have a null image_path (audio_jack_socket 1, crimp_st_cable_mount_box 4, miniature_lifting_motor 25, power_jack 25, purple_clay_pot 9). A record needs an image, so those are dropped — hence 10,874, not 10,938. They are counted and reported by the converter rather than silently skipped. Every remaining record resolves: 0 missing RGB files and 0 missing masks, verified against the manifests.

Mask (deferred GT). The 2D binary mask rides in the mask column for defective captures; good captures carry mask = null. The voxel mask is a 3D artefact and is referenced by path only.

Grouping. metadata.sample_id (e.g. S0043) identifies the physical object. Group by it if you resample the splits, so different captures of one object cannot straddle a split.

Lazy-baseline floor. The test split is heavily anomaly-skewed — 1,000 good vs 3,450 anomalous — so always answering anomalous scores 77.5% on the binary question. The full {label, defect_type} floor is 22.5% (always {good, null}). Report against these, not against 50%.

Provenance

Underlying dataset: Real-IAD D3. Upstream license: CC BY-NC-SA 4.0 (Rongcheer Industrial Technology (Suzhou) Co., Ltd.; gated upstream) (this card is license: other; respect the upstream terms). Converted read-only from the raw source into the unified schema; conversion code under 211/ (with publish/push_to_hf.py) in AI4Manufacturing/forge_model.

Overlap / de-duplication (§8)

Distinct from 192-single/192-object (plain Real-IAD) — different capture campaign and categories — but same publisher, so keep the Real-IAD family on ONE side of any split. ⚠ This repo carries the 2D modalities only; see the modality note below. Each record carries metadata.image_sha256 so overlapping images can be kept entirely on one side of a train/eval split.

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`	`1afcea4b37335769`

Provenance and verification


records	10,874
carrying a geometry block	10,874 / 10,874
instances per record	`0`: 5,504, `1`: 4,895, `2`: 314, `3`: 106, `4`: 21, `5+`: 34
total instances	6,132
image dimensions	5328×3040 (225), 2658×2658 (54), 2657×2657 (50)
`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)	932×932 → 924×924, 933×933 → 924×924, 934×934 → 924×924
shipped boxes	6,132
legible at that render (>=16px there)	6,072 (99.0%)

⚠ 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 932×932 is rendered 924×924 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: 31

Total file size:

20.5 GB

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