PKU-GoodsAD: A Supermarket Goods Dataset for Unsupervised Anomaly Detection and Segmentation
Paper • 2307.04956 • Published
You need to agree to share your contact information to access this dataset
This repository is publicly accessible, but you have to accept the conditions to access its files and content.
This dataset is released for research use. Access is reviewed and granted manually by the maintainers. Please state your name, affiliation, and intended use.
Log in or Sign Up to review the conditions and access this dataset content.
D22-goods
The only part of D22 that may enter a training pool: its 3,115 unlocked train-split good records. Everything else in D22 is eval-locked corpus-wide.
query, image and annot are carried verbatim from the parent canon, so this repo is the
canon's trainable subset and mixes with it by construction rather than by a matching rule that could
drift. Nothing is re-rolled.
Records
3,115, all good. Excluded: 22 train
records whose images are eval-locked elsewhere.
Query-pool routing, read off the artifact rather than recomputed:
F2a/label_type 2,932 · F2b/label_type_open 183.
Training status — and the caveat that cost 19 records to find
Predicate, both halves asserted separately: split == "train" AND NOT eval_locked(image_sha256),
and then overlap.assert_train_pool_clean must PASS over the result. That is the opposite
direction from this family's two evaluation rungs, which assert the same gate must RAISE — a gate
checked in only one direction never sees the day it breaks.
Absence from the overlap manifest never meant trainable. Absent from the overlap manifest never meant trainable: it indexes cross-family duplicates only (0 of 9,278 entries name a single repo), so a record that is test in its own repo is locked by its split, not by the manifest.
What the corpus eval-lock excludes from this pool
This repository is the trainable remainder of D22 — the rows below are what the eval-lock takes
out, and this pool is the free to train column. D22 shares 2,968 images with MMAD/D05, where they
are test, and an image that is test or validation in any repo is eval-locked in every repo; that is
why this pool is 3,115 and not the 3,137 train records the source ships.
| split | label | locked by the manifest | locked, any reason | free to train | total |
|---|---|---|---|---|---|
| test | good | 1,309 | 1,328 | 0 | 1,328 |
| test | anomalous | 1,659 | 1,659 | 0 | 1,659 |
| train | good | 22 | 22 | 3,115 | 3,137 |
The 19 test goods the manifest does not hold are missing from it only because D05 does not happen to hold those images — they are locked anyway, by their own test split: the manifest indexes cross-family duplicates, and a record whose own split is an evaluation split is locked before the manifest is consulted. Absence from the manifest was never clearance.
Roles
Answer-only tier: annot is the training target, in the form the query asks for. There is no
reasoning column in this release; a gated goods explain view is a separate, later decision and is
not part of this repo.
Single-type categories
cigarette_box has exactly one defect type in this source (opened). On a category with one type no
closed-set type question is asked. Where a query here names that type it is a statement of what to
find, never a question of which type it is, so the token carries no information for that one
category — the rule permits it and the card says so rather than leaving a reader to wonder.
Loader notes
- 823 of D22's images are MPO (multi-picture JPEG, the 3024×3024 subset). Pillow decodes the first frame at the declared size. The published bytes are the source bytes; this is a loader note, not a defect.
drink_cancontains paper cartons as well as cans. The category token is the source's label, not a description of the object.straw_missinggold is an absent-part region — a band where the straw should be, with no visible damage inside it. A localisation task on that type asks where something is missing.
Provenance
Derived at $0 from the gold of AI4Manufacturing/D22. No teacher model was called; every answer on this
repo is a pure function of the source annotation.
Two revisions are recorded here on purpose. The gold was read from shards at
0cd5ee9942ce; this artifact pins 06116762a9b8. That is only honest because the
columns this build depends on were verified byte-identical between the two, by reading both revisions
from the Hub and diffing them rather than by trusting the account of the change: annot moved on
0 of 6,124 records, query moved on exactly 612 — the
cigarette_box set — and the only metadata keys that moved were
query_pool, query_template. Verified 2026-09-17, read from the Hub at both revisions.
Upstream source: PKU-GoodsAD (Zhang et al., IEEE RA-L 2024, arXiv:2307.04956), GPL-3.0. This card is license: other; the repository is gated
manual and released for research use.
Generator: annotate/D22/build_goods.py (sha16 78aecb4debe5be59)
Overlap / de-duplication (§8)
Cross-family evaluation lock — metadata.eval_lock (stamped 2026-09-20; manifest revision fe6e286912b0, generated 2026-09-08). Every record of this repository, locked or not, carries metadata.eval_lock, computed by forge_model/common/overlap.py::Overlap.stamp_for against common/overlap_manifest.json at that revision — so within this repository the absence of the key cannot occur. Shape: {"locked": bool, "against": [{"repo": …, "split": …}, …], "own_split": …, "manifest_revision": …, "manifest_generated": …}. locked is true when the image is evaluation material anywhere in the corpus; against names every repository and split in which it is (sorted; [] when not locked; it includes the record's own family where that is so); own_split marks a record locked by its own split. The per-record field is the authority — the count here is quoted once, at this revision, and a later manifest may change it: 0 of 3,115 records (0 distinct images) are locked — by column: 0 by the cross-family manifest, 0 by their own split, 0 both ways and counted once; counterparts (records per counterpart; a record can appear under several): none — no record in this repository is locked at this manifest revision. A stamp whose manifest_revision differs from the current manifest is stale, not wrong — recompute it (Overlap.stamp_is_current); a record with no stamp has not been checked against the corpus as it now is. Overlap.partition / assert_train_pool_clean read the field: a train pool built from this repository must exclude every locked record. Checked and free: no record of this repository is evaluation material anywhere in the corpus at manifest revision fe6e286912b0 — the stamp says so on every one of the 3,115 records, which is what distinguishes "checked and free" from "never checked".
- Downloads last month
- 65