AI Tech Inspire

Kanops Retail Scenes: 10k Supermarket Photos for Planogram and Shelf AI (Eval-Only)

Retail AI has plenty of clean benchmarks for products and barcodes — but far fewer real-world store scenes with messy fixtures, weird seasonal displays, and questionable signage placement. At AI Tech Inspire, a new dataset caught our eye because it leans into that reality: authentic supermarket photos intended for evaluating shelf and planogram reasoning rather than polished showroom samples.

Quick facts at a glance

  • ~10k+ retail photos from UK/US supermarkets; includes fixtures, shippers, pumpkins/seasonal displays, and signage.
  • Faces are blurred; EXIF/IPTC metadata carries provenance.
  • Gated for evaluation use only: no redistribution and no model-weight redistribution.
  • Hosted on Hugging Face; dataset path structured under train/{2014, FullStores, Halloween2024}/Retailer/Subcategory/*.jpeg.
  • Repo includes MANIFEST.csv, metadata.csv, checksums.sha256, LICENSE, and README.md.
  • Intended tasks: scene understanding for retail — bay detection, planogram reasoning, signage classification, seasonal cues, OCR on shelves, and shelf-fill adjacent use cases.
  • Roadmap (v1): weak labels for orientation/aspect/season and CVAT tags.
  • Quick load via datasets library using the imagefolder builder with an hf:// path.
  • Contact: happytohelp@groceryinsight.com for questions and task suggestions.

Why this matters for developers and engineers

Most open retail datasets focus on products in isolation (think controlled lighting, single-shelf crops, or box-level annotations). Real stores are different: mixed fixtures, temporary shippers, seasonal chaos, and signage that tells you more than pixels alone. If you’re working on shelf compliance, planogram checks, or store analytics, you need evaluation scenes that stress-test models as they’ll actually be used.

That’s the gap this collection targets. The images are built to probe scene understanding: finding where one bay ends and another begins, reasoning if the display matches expected planograms, and interpreting environmental cues like Halloween promotions. It’s especially useful if you’re benchmarking vision systems built with PyTorch, TensorFlow, or multimodal stacks that combine VLMs (e.g., APIs based on GPT) with classic CV.

Key takeaway: This set is not for training or redistribution — it’s a reality check for your models on retail scenes you’re likely to encounter in the wild.

What’s inside the folder structure

The structure hints at both historical and seasonal coverage:

  • train/2014: older content, useful for domain-shift evaluation and testing model robustness to vintage fixtures or legacy signage.
  • train/FullStores: broader store-scene coverage beyond just tight shelf crops.
  • train/Halloween2024: seasonal variance (pumpkins and promos) to check how models adapt to fast-changing visual context.

Beneath those, folders are split by Retailer/Subcategory, which can be repurposed for weakly supervised tasks. Faces are blurred, and EXIF/IPTC carries provenance — helpful for auditability, but also a reminder to treat metadata carefully in your pipelines.

Licensing and constraints (read this before you test)

The dataset is gated for evaluation use, with explicit restrictions: no redistribution and no model-weight redistribution. In practice, that means:

  • Use it to evaluate model performance internally.
  • Do not redistribute the images or any derivative that violates the terms.
  • Avoid training on the images and then open-sourcing the resulting weights.

Always confirm the exact terms in the included LICENSE and README.md. When in doubt, reach out to the maintainers at happytohelp@groceryinsight.com.

Getting started quickly

The images are on Hugging Face with a straightforward path for datasets:

# pip install datasets
from datasets import load_dataset
ds = load_dataset("imagefolder", data_dir="hf://datasets/dresserman/kanops-open-access-imagery/train")
print(len(ds["train"]))

Expect the CSVs to be useful for bookkeeping and provenance checks. The checksums.sha256 file is there for integrity; verify it if you’re running critical evaluations or sharing results internally.

What to evaluate (and how)

Even without dense labels, there’s a surprising amount you can evaluate with minimal annotation or weak supervision:

  • Bay detection: Use heuristic or geometric detectors to split bays (e.g., identify vertical dividers). Evaluate IoU against a small hand-labeled subset.
  • Planogram reasoning: Convert planogram JSONs (if you have them) into expected patterns and ask a VLM pipeline to identify mismatches; measure precision/recall on a limited set of human-reviewed scenes.
  • Signage classification: Zero-shot classify signage types using CLIP-like embeddings; sanity-check with a small validation set.
  • Seasonal cues: Test models on domain-shift robustness by comparing performance between FullStores and Halloween2024.
  • OCR on shelves: Benchmark OCR robustness for price tags and promos. Use your favorite OCR stack and evaluate on curated snippets.

If you already have retail-trained detectors from internal data, this collection provides a clean way to validate generalization. For teams that rely on synthetic generation (e.g., using Stable Diffusion for augmentation), it’s a reality check to see where synthetic-trained models struggle in actual stores — without touching training on these images.

How it compares to other retail datasets

Public sets like SKU-110K or Retail-Product-Checkout (RPC) emphasize product-level detection and counting with strong annotations. Kanops Retail Scenes swings the pendulum toward store context and messy real-world fixtures. It’s less about “is that a cereal box?” and more about “is this bay compliant with the planogram, and what does the signage say about current promotions?” That shift is essential for end-to-end retail analytics, where decision-making depends on scene-level understanding.

Suggested pipelines and baselines

  • Zero-shot retrieval: Use CLIP-like embeddings to retrieve scenes with similar fixtures for qualitative evaluation. If you’re running on GPUs, make sure your stack is ready with CUDA acceleration.
  • Hybrid VLM + CV: Combine a VLM (API-based or open-source) with classic detectors; for example, detect signage regions, then prompt the VLM for planogram consistency checks. Keep prompts and heuristics versioned for reproducibility.
  • Weak supervision: Leverage folder names and EXIF/IPTC to bootstrap labels; validate on small subsets to mitigate noise.
  • Robustness testing: Compare performance across 2014 vs. Halloween2024 to quantify temporal and seasonal drift.

Training baselines (if using your own data) can be assembled with PyTorch or TensorFlow; evaluate those models on Kanops to measure generalization without violating the dataset’s terms.

Roadmap and what to watch

The maintainers note a v1 roadmap to add weak labels (orientation, aspect, season) and CVAT tags. That’s significant: even coarse labels can unlock structured evaluations and speed up semi-automatic annotation. If you’re planning downstream benchmarks, consider suggesting label schemas or providing feedback on the categories that would best reflect your production tasks.

CVAT support also signals a move toward community-in-the-loop improvements. If certain tasks (e.g., bay boundary definitions) are crucial for your org, aligning on tag conventions now may save you hours of rework later.

Practical tips before you run

  • Respect the license: Treat this as an evaluation set. Don’t redistribute images or trained weights derived from them.
  • Maintain metadata hygiene: EXIF/IPTC is present; strip or preserve thoughtfully depending on your compliance requirements.
  • Cache smartly: If you’re using a shared runner, local caching with the datasets library reduces repetitive pulls.
  • Version your evals: Keep a manifest of which subset you used and the exact commit/hash from the dataset page.

Final thought

Real stores are beautifully chaotic. A dataset that leans into that — even as an eval-only resource — is useful for anyone building shelf intelligence, planogram checks, or signage-aware assistants. If your models only look good on clean product shots, they’re not ready. Use this set to find the cracks before your customers do.

Questions or task ideas? The maintainers invite suggestions at happytohelp@groceryinsight.com. And if you publish internal findings, consider sharing methodology notes with the community. The more we collectively stress-test scene understanding, the faster retail AI moves from the lab to the aisle.

Recommended Resources

As an Amazon Associate, I earn from qualifying purchases.