Retail shelf intelligence starts with annotated shelf images. Not web-scraped product photos, not rendered planogram diagrams - photos of real shelves in real stores, with all the variation in lighting, label wear, shelf organization, and product placement that retail teams deal with every day.
Retail Shelf Observations is the training data underlying Shelf-1, General Machines' shelf intelligence model. It is a collection of annotated shelf photographs across a range of retail categories and store formats, with per-facing product labels, price tag detections, and inventory counts at the SKU level. A subset of records are paired with ground-truth POS data from retail partners, enabling training of models that connect visual shelf state to commercial outcomes - not just detection accuracy.
What is annotated
Each image in the dataset carries five annotation layers:
- Product facing: For every product visible on the shelf, a bounding box with SKU label, product category, and a confidence score. Facings that are partially occluded are annotated with an occlusion flag and estimated visible fraction.
- Price tag: Detected price value, tag format (shelf label, electronic shelf label, handwritten, promotional), and a placement quality score that flags tags that are misaligned, obscured, or associated with the wrong facing.
- Inventory: Unit count per facing and empty slot flags where facings are present in the planogram but no product is visible on shelf.
- Planogram compliance: For images collected with planogram data available, each shelf position is annotated with expected versus detected state: correct product, wrong product, missing product, or extra product outside plan.
- POS link: For records collected with ground-truth point-of-sale data from the retail partner, each annotated SKU carries a sales velocity figure (units sold per day at that store location during the collection window). This pairing makes it possible to train models that predict commercial outcomes from visual shelf state, not just classify what is present.
Coverage
- Retail formats: Grocery, consumer goods, specialty retail, and pharmacy (in progress)
- Price tag formats: 40+ distinct price tag formats across format and retailer variation
- Collection method: Standard smartphone photos taken by store associates during routine store walks - no specialist hardware, no sterile capture conditions
The collection method is deliberate. A model trained on photos taken under ideal lighting with professional equipment will not generalize to photos taken by an associate with a store-issued phone at 7 AM before a shift. This dataset is built to close that gap.
Use cases
- Training retail computer vision models: End-to-end training data for detection, classification, and inventory estimation models across a wide range of retail environments and product categories.
- Fine-tuning detection heads: Category-specific or format-specific subsets are available for teams that need to adapt an existing model to a new retail vertical or tag format.
- Benchmarking shelf analytics systems: The combination of per-facing annotations and POS ground truth makes it possible to evaluate not just detection accuracy but whether a model's outputs predict real commercial outcomes.
Data specifications
- Format: JPEG images with structured JSON annotation per image
- Annotation layers: Product facing (SKU, category, bounding box, confidence), price tag (detected value, format, placement quality), inventory (unit count, empty slot flags), planogram compliance (where planogram data available), POS link (where ground-truth sales data provided)
- Price tag coverage: 40+ distinct tag formats
- Retail formats: Grocery, consumer goods, specialty retail, pharmacy
- Collection method: Smartphone photos by store associates, no specialist hardware
facing
tag
count
compliance
link
layers per image
formats covered
format categories
Access
Retail Shelf Observations is available for partnership-based access. Dataset access is structured as a data partnership - retail partners contribute images and, where available, POS ground truth, and receive enhanced annotation and model access in return. AI labs and computer vision teams interested in direct dataset licensing can also reach us to discuss terms. Contact us at founders@generalmachines.ai.