Every existing retail CV dataset (SKU-110K, RP2K, Freiburg Groceries, MVTec D2S) captures a single moment in time. Kanops captures 15 years of change — how stores, shelves, packaging, and merchandising actually evolve.

This open sample contains ~10–11K retail imagery files from UK and US stores (2010–2024). The full Kanops archive spans 1.135M+ retail images across 2009–2025, covering 20+ retailers, 325,000+ seasonal event images, and documented shifts through Brexit, COVID-19, and the cost-of-living crisis.

Faces blurred. Provenance embedded (SHA-256 checksums + EXIF/IPTC/XMP). Evaluation-only license.

How Kanops compares

Feature	Kanops	SKU-110K	RP2K	Freiburg Groceries	MVTec D2S
Images	10K sample (1.135M+ full)	11,762	200,000	5,000	21,000
Time span	2010–2024 (15 years)	Single capture	Single capture	Single capture	Single capture
Temporal data	✅ Year, season, event	❌	❌	❌	❌
Retailers	20+ named UK/US chains	Anonymous	Anonymous	Single store	Synthetic
Scene types	Full store, aisle, shelf, seasonal	Shelf only	Product only	Shelf only	Tabletop
Seasonal events	✅ Christmas, Easter, Halloween	❌	❌	❌	❌
Real-world conditions	✅ Lighting, layout changes	Limited	Studio	Limited	Synthetic

What you can build

Kanops enables research and applications that static, single-capture datasets cannot support:

Planogram compliance monitoring — detect shelf arrangement drift over time
Packaging evolution tracking — trace how product packaging, branding, and label design change across years
Seasonal merchandising analysis — model how retailers deploy and rotate seasonal displays (Christmas, Easter, Halloween)
Domain shift research — study how real-world retail environments change across lighting, layout, and fixture conditions
Price and promotion cycle modelling — analyse promotional patterns including inflation effects pre- and post-COVID
Product recognition and detection — train and benchmark shelf-level product identification models
Store layout and fixture analysis — study how store formats and merchandising strategies differ across retailers and evolve over time
Competitive intelligence benchmarking — compare visual merchandising approaches across 20+ named retailers

Trademarks. All retailer names, store marks, and logos referenced in this dataset are the property of their respective owners. Kanops is not affiliated with, endorsed by, or sponsored by any retailer. Images are provided for evaluation and research under the included license.

Access (gated)

This dataset is gated on Hugging Face. Request access on the repo page. By requesting and using the data you agree to the Kanops Evaluation License (see LICENSE at the root).

Academic researchers: A free research license is available for academic institutions with a citation requirement. Contact hello@kanops.ai for details.

What's included

Images under train/:
- train/2014/<Retailer>/*.jpg
- train/FullStores/<Retailer>/**/*
- train/Halloween2024/<Retailer>/**/*
Control files at the root:
- MANIFEST.csv — file list + basic attributes
- metadata.csv — enriched attributes (e.g., collection, dims)
- checksums.sha256 — SHA-256 for all images in this release
- LICENSE — evaluation-only terms
- README.md — this file

Images live only under train/. All control files sit at the repository root.

How to load

Hugging Face Datasets (ImageFolder)

from datasets import load_dataset

ds = load_dataset(
    "imagefolder",
    data_dir="hf://datasets/dresserman/kanops-open-retail-imagery/train",
    split="train",
)
img = ds[0]["image"]  # PIL.Image

Read metadata

import pandas as pd
meta = pd.read_csv("hf://datasets/dresserman/kanops-open-retail-imagery/metadata.csv")
meta.head()

Starter Notebooks

Explore the dataset with our Kaggle notebooks:

Data schema (minimum)

metadata.csv columns (you may see additional fields):

file_name — path relative to dataset root (e.g., train/2014/Aldi/IMG_1234.jpeg)
bytes — file size
width, height — image dimensions (if available)
sha256 — content hash (provenance/integrity)
collection — one of {2014, FullStores, Halloween2024}

If present:

retailer — inferred from path
year — inferred from path

Dataset Creation

Collection methodology

Kanops retail imagery has been collected through systematic weekly store visits since 2009 by Grocery Insight, a UK retail consultancy. Trained photographers visit stores across the UK and US, capturing shelf layouts, seasonal displays, promotional setups, and full-store walkthroughs. At peak collection, the team captures 3,600+ images per week across multiple retailers.

This open sample (v0) draws from three collections within the archive: a 2014 cross-retailer survey, full-store photography spanning multiple years, and a Halloween 2024 seasonal event capture.

Annotation and quality control

Over 58,000 images in the broader Kanops archive have been human-verified with retailer, date, and scene-type labels. Metadata is inferred from folder structure (retailer, year, collection) and enriched with image dimensions, file hashes, and embedded EXIF/IPTC/XMP provenance data.

Quality control includes automated duplicate detection (perceptual hashing), SHA-256 integrity verification, and manual review passes for face blurring and metadata consistency.

Source data

All images are original photographs taken by the Kanops / Grocery Insight team. No images are scraped, crowdsourced, or sourced from third parties. Stores are publicly accessible retail environments.

Baseline Model Performance

Early-stage models trained on the broader Kanops archive demonstrate the dataset's utility for retail computer vision tasks. These results are indicative and drawn from internal development; independent benchmarks are welcomed.

Model	Task	Metric	Notes
Model A	Shipper/display detection	~99.5% mAP	Trained on a larger annotated subset; high confidence on a well-defined object class
Model B	Seasonal event classification	~77.7% accuracy	Multi-class classification across Christmas, Easter, Halloween, and non-seasonal imagery
Model C	Product category classification	~61.3% accuracy	Fine-grained categorisation task; performance reflects the inherent difficulty of the problem

These figures are preliminary and based on internal validation splits. We encourage researchers to publish independent evaluations using this open sample.

Considerations for Using the Data

Geographic and retailer coverage

The dataset has a strong UK focus, reflecting the collection team's base in West Yorkshire, England. US retail imagery is included but coverage is more limited in both the number of retailers and image volume. Researchers should account for this geographic imbalance when generalising findings.

Retailer representation is not uniform — some chains (e.g., Tesco, M&S, Aldi) have substantially more imagery than others. This reflects real-world collection patterns rather than deliberate sampling.

Temporal distribution

Images are not evenly distributed across years. Earlier years (2010–2013) have fewer images than later years as the collection programme scaled. Seasonal event coverage is concentrated in Q4 (Halloween, Christmas) with lighter coverage of spring/summer events.

Environment and conditions

All images are captured in real retail environments with natural variation in lighting, camera angle, and store layout. This is a feature for domain-shift research but may require preprocessing for tasks that assume controlled conditions.

Ethical considerations

Images are captured in publicly accessible retail spaces. All visible faces are blurred via automated detection with manual review. No customer tracking, behavioural, or transaction data is included. Retailer names are used descriptively — the dataset does not imply endorsement.

Licensing

The evaluation-only license permits research, benchmarking, and prototyping. Redistribution of images or derivatives, public model-weight distribution, and commercial training require a separate commercial license. Academic institutions can request a free research license with citation requirement — contact hello@kanops.ai.

Intended use

Evaluation/benchmarking of shelf/fixture detection, retrieval, layout analysis, merchandising.
Research and prototyping under evaluation terms.

Not permitted: redistribution of images/derivatives; public model-weight redistribution; production or commercial training without a separate commercial license. See LICENSE.

Privacy & provenance

Faces blurred via automated pass + manual review.
Rights/usage embedded in EXIF/IPTC/XMP.
checksums.sha256 provides per-file SHA-256 for integrity.
For takedown or corrections, contact hello@kanops.ai with file_name and SHA-256.

Maintenance commitment

Kanops commits to maintaining this dataset for a minimum of 5 years (2025–2030), with planned annual updates reflecting new seasonal retail imagery. The DOI via Zenodo ensures long-term citability regardless of hosting changes.

Versioning

v0 — initial release; faces redacted; metadata + checksums included.
Future versions (v1, v2, …) will be published immutably.

Full Dataset

This is a ~10K sample of retail imagery (2010–2024). The full Kanops archive contains:

1.135M+ images spanning 2009–2025
20+ UK/US retailers with extensive coverage
Seasonal collections: 325,000+ seasonal event images
Temporal depth: Brexit, COVID-19, cost-of-living crisis documented

For commercial licensing: hello@kanops.ai

Citation

APA Dresser, S., & Dresser, K. (2025). Kanops — Open Access · Imagery (Retail Scenes v0) [Dataset]. Kanops.ai / Grocery Insight. https://doi.org/10.5281/zenodo.18071325

BibTeX

@dataset{dresser_kanops_retail_scenes_v0_2025,
  title   = {Kanops — Open Access · Imagery (Retail Scenes v0)},
  author  = {Dresser, Steve and Dresser, Katie},
  year    = {2025},
  doi     = {10.5281/zenodo.18071325},
  url     = {https://doi.org/10.5281/zenodo.18071325},
  note    = {Evaluation-only license. Faces blurred.}
}

Contact

Questions, broader licensing, or takedowns: hello@kanops.ai / www.kanops.ai

Downloads last month: 18