---
license: cc-by-4.0
pretty_name: Bodnár-Huawei Syn10k — synthetic + real barcode images with masks
task_categories:
- image-segmentation
- object-detection
tags:
- qr-code
- barcode
- segmentation
- synthetic
size_categories:
- 10K<n<100K
configs:
- config_name: warp2014
  data_files:
  - split: train
    path: data/warp2014/train*.parquet
- config_name: huawei
  data_files:
  - split: train
    path: data/huawei/train*.parquet
---

# Bodnár-Huawei Syn10k

A **mirror** of the `syn10k_plus_huawei` barcode dataset from the Szeged group
(Bodnár, Grósz, Tóth), repackaged as parquet with each image joined to its
ground-truth mask in the same row.

Released under **CC BY 4.0**, so this mirror is permitted with attribution.

## Configs

| config | rows | content | with mask |
|---|---:|---|---:|
| `warp2014` | 10,000 | synthetic warped barcode images | 10,000 |
| `huawei` | 98 | real photographs | 98 |

**Every image has its mask** — the pairing was verified by stem, and zero images
in either config are missing one.

## Fields

| field | notes |
|---|---|
| `image` | the picture (JPEG) |
| `mask` | the ground-truth mask (PNG) |
| `file_name`, `stem` | original filename and its stem |
| `subset` | `warp2014` or `huawei` |
| `has_mask` | always true in this build; kept so the schema stays honest if a future revision has gaps |

## Two notes on how to read this

**The split name is a container, not a protocol.** Both configs expose a single
split called `train` because HuggingFace requires a split name — it does **not**
encode the authors' intended train/test partition. `warp2014` is synthetic and
`huawei` is real photographs, which is the natural synthetic-train / real-test
arrangement, but this mirror does not assert that the authors used it that way.
Consult the source publication before reporting numbers as theirs.

**This is a barcode dataset, not a QR-only one.** It is included in a QR
collection because it is standard in the barcode-localisation literature and is
one of the few sets shipping pixel masks. Check the masks for what you need
rather than assuming every symbol is a QR code.

## Usage

```python
from datasets import load_dataset

syn  = load_dataset("devmandan/syn10k-huawei-barcodes", "warp2014", split="train")
real = load_dataset("devmandan/syn10k-huawei-barcodes", "huawei",   split="train")
syn[0]["image"], syn[0]["mask"]      # PIL images, aligned
```

## Citation

Original source:
<https://www.inf.u-szeged.hu/~bodnaar/barcode_database/syn10k_plus_huawei.zip>

```bibtex
@misc{bodnar_syn10k,
  title  = {Syn10k + Huawei barcode database},
  author = {Bodn\'ar, P\'eter and Gr\'osz, Tam\'as and T\'oth, L\'aszl\'o},
  note   = {Synthetic and real barcode images with ground-truth masks},
  url    = {https://www.inf.u-szeged.hu/~bodnaar/barcode_database/}
}
```

Licence: CC BY 4.0 — <https://creativecommons.org/licenses/by/4.0/>
Please cite the authors above, not this mirror.

## Derived ground truth (qr-decode-bench)

`derived_gt/derived_gt.jsonl` was ADDED to this repository on 2026-09-01; no existing file was modified (this section was appended to the card). It records, for each image whose payload this dataset does not publish and on which at least two independent QR decoders agreed, what every decoder of a panel decoded on that image — as sha256 digests of the payload bytes, with the payload text beside them only where no identifier detector fires on it — so that a benchmark can score a decoder against what the OTHER decoders established (per-image leave-one-out, quorum 2), never against its own output.

This is truth DERIVED by decoder consensus, not published by the dataset's authors; the benchmark that uses it marks it with †. Agreement between independent implementations is strong evidence, not proof of what a symbol encodes.

* derived at revision `b85ce5ce82d34f4515bf5c90b60d2b680defd401` from the data files' recorded inventory (sha256 `fa6cf1fdb7d7c8d9d91962eb9b9fd02b8206c57c1c405ea26b893053bc2a2ac8`); the data files are unchanged by this commit
* panel: apple-vision (apple-vision macOS 14.3.1 VNDetectBarcodesRequest rev4), boofcv (boofcv 0.43.1 (PyBoof jar; java 22.0.1)), opencv (opencv 4.12.0 (QRCodeDetector)), opencv-aruco (opencv 4.12.0 (QRCodeDetectorAruco)), quirc (quirc 1.0 (vendored 927d680904dc)), wechat (wechat_qrcode (opencv-contrib 4.12.0; models 8ae41acb,cc49b8c9,e5d36889,e8acfc39)), zxing-cpp (zxing-cpp 3.1.1)
* images with derived truth: 6580 of 10098 examined (3518 without a quorum, 0 voided)
* payloads: 6516 established, 0 carried as text, 0 as digest only because an identifier detector fired, 6516 as digest only because the payload is the image's own id (image records are keyed by `sha256` of the id)
* format `qrb-derived-gt-peer/1`: JSON lines — one header (rule, quorum, panel versions, yield, corroboration, privacy scan), one line per payload, one line per image

## Derived ground truth (qr-decode-bench)

`derived_gt/derived_gt.jsonl` was ADDED to this repository on 2026-10-01; no existing file was modified (this section was appended to the card). It records, for each image whose payload this dataset does not publish and on which at least two independent QR decoders agreed, what every decoder of a panel decoded on that image — as sha256 digests of the payload bytes, with the payload text beside them only where no identifier detector fires on it — so that a benchmark can score a decoder against what the OTHER decoders established (per-image leave-one-out, quorum 2), never against its own output.

This is truth DERIVED by decoder consensus, not published by the dataset's authors; the benchmark that uses it marks it with †. Agreement between independent implementations is strong evidence, not proof of what a symbol encodes.

* derived at revision `974c629f99349b3112ad68b88fbbff71c68d1122` from the data files' recorded inventory (sha256 `fa6cf1fdb7d7c8d9d91962eb9b9fd02b8206c57c1c405ea26b893053bc2a2ac8`); the data files are unchanged by this commit
* panel: apple-vision (apple-vision macOS 14.3.1 VNDetectBarcodesRequest rev4), aspose-net (aspose-net 26.8 key=trial decode_type=QR), boofcv (boofcv 0.43.1 (PyBoof jar; java 22.0.1)), dynamsoft-dbr (dynamsoft-dbr 11.6.3000 (11.6.30.0818 (Algorithm 11.6.30.28649)) key=trial template=PT_READ_BARCODES formats=QR), mlkit (barcode-scanning 17.3.0 (bundled); Android API 34; arm64-v8a), mmd-hyb-gen (mmd v0.96 8ead9b0fa1d1 effort=64 hyb=sprint deep=1500@150000kpe view=12000kpe seedcap=5 deep.pyr_cap=1000000 deep.futility=1 deep.spill=0 deep.ccfind=0 prefix=1500@2800kpe prefix.pyr_cap=500000 prefix.futility=1 prefix.spill=0 prefix.ccfind=0 esc=850 det=heatnet_mixe3@ded00e97ea9c hyb_ref=refiner_v0@7e2f3313294b ort=1.30.0 cv2=4.12.0), opencv (opencv 4.12.0 (QRCodeDetector)), opencv-aruco (opencv 4.12.0 (QRCodeDetectorAruco)), quirc (quirc 1.0 (vendored 927d680904dc)), scanbot-linux (scanbot-linux 9.0.0 key=trial symbology=QR:restricted-at-api), scandit-linux (scandit-linux version-unreported key=trial symbology=QR:restricted-at-api), wechat (wechat_qrcode (opencv-contrib 4.12.0; models 8ae41acb,cc49b8c9,e5d36889,e8acfc39)), zxing-cpp (zxing-cpp 3.1.1)
* images with derived truth: 6827 of 6919 examined (92 without a quorum, 0 voided)
* payloads: 6760 established, 0 carried as text, 0 as digest only because an identifier detector fired, 6760 as digest only because the payload is the image's own id (image records are keyed by `sha256` of the id)
* format `qrb-derived-gt-peer/2`: JSON lines — one header (rule, quorum, panel versions, yield, corroboration, privacy scan), one line per payload, one line per image
* SUPERSEDES the earlier derivation at revision `974c629f99349b3112ad68b88fbbff71c68d1122` on the same image bytes: every image it established is established here with a superset of its payloads and no void (6580 kept identical, 0 completed, 247 added). Panel members whose reads are committed benchmark rows rather than live decodes are listed in the header's `panel_sources`; a `null` read means the member never examined that image.