---
license: cc-by-4.0
pretty_name: SBD QR Subset (low resolution) — barcode ROIs with instance masks
task_categories:
- image-segmentation
- object-detection
tags:
- qr-code
- barcode
- instance-segmentation
- synthetic
size_categories:
- 10K<n<100K
configs:
- config_name: default
  data_files:
  - split: train
    path: data/train/*.parquet
  - split: validation
    path: data/val/*.parquet
  - split: test
    path: data/test/*.parquet
---

# SBD QR Subset — low resolution

A **mirror** of the low-resolution ROI split of the Synthetic Barcode Dataset
(Quenum, Wang, Zakhor), repackaged from **749,682 loose files into parquet**.

| split | ROIs | instances |
|---|---:|---:|
| train | 80,000 | 439,731 |
| validation | 10,000 | 55,072 |
| test | 10,000 | 54,876 |
| **total** | **100,000** | **549,679** |

## Why this repackaging exists

Upstream, this split is three-quarters of a million individual PNG and JPEG
files. That is unpleasant to move, impossible to browse, and slow to load. Here
each ROI is **one row** carrying its image, its combined mask, all of its
per-instance masks, and its bounding boxes — so the whole thing loads with one
call and renders in the dataset viewer.

## Fields

| field | notes |
|---|---|
| `image` | the 400×400 grayscale ROI |
| `mask` | the combined mask for the ROI |
| `instance_masks` | list of per-instance masks, one per barcode |
| `boxes` | list of `[x1, y1, x2, y2]`, aligned index-for-index with `instance_masks` |
| `n_instances` | number of barcodes in the ROI |
| `roi_id` | upstream ROI index (`roi<N>.png` ↔ `img_<N>` in `all_bboxes.json`) |

### The box↔mask alignment was verified, not assumed

`boxes[k]` and `instance_masks[k]` are the same object. The packer checked the
count of boxes against the count of instance-mask files for **every one of the
100,000 ROIs and found zero mismatches**, and no ROI was missing its combined
mask. If they had disagreed, the join would have been silently wrong in a way no
loader would flag — so it is checked rather than trusted.

## Usage

```python
from datasets import load_dataset

ds = load_dataset("devmandan/sbd-qr-subset", split="train", streaming=True)
row = next(iter(ds))
row["image"], row["mask"], row["instance_masks"][0], row["boxes"][0]

# crowded ROIs
full = load_dataset("devmandan/sbd-qr-subset", split="test")
crowded = full.filter(lambda r: r["n_instances"] >= 8)
```

`streaming=True` is worth using here — the train split is ~3 GB.

## Scope, stated plainly

This mirror carries the **`low_resolution` ROI split only**. The upstream SBD
release also has ultra-high-resolution splits and full-scene imagery that are
**not** included here. Do not describe results on this mirror as results on SBD.

It is also a **barcode** dataset covering multiple symbologies, not a QR-only
one; it appears in a QR collection for its scale and its instance masks.

## Citation

```bibtex
@misc{sbd_synthetic_barcode_dataset,
  title  = {Synthetic Barcode Dataset (SBD)},
  author = {Quenum, Jerome and Wang, Kehan and Zakhor, Avideh},
  note   = {Synthetic barcode detection and segmentation dataset}
}
```

Licence: **CC BY 4.0** as recorded in the dataset registry (the authors' code
repository is separately MIT). Attribute the authors above, not this mirror.