Contents
If you’re working in the CPG industry, you’re drowning in shelf photos, cooler shots and display videos. But what you really need is clean, objective data and the ability to act with confidence. This piece will explain what a computer vision ontology is, why it's more than a taxonomy and how the Neurolabs Ontology transforms visual data into retail execution insights.
What Is an Ontology?
Put simply, an ontology is a set of concepts and categories in a subject area or domain that shows their properties and the relations between them. In the world of computer vision, it's the master plan that translates raw visual information into a machine-readable, actionable representation of the real world. Think of it as the shared “language” that allows humans and machines to describe a scene and reason about it.
The Problem an Ontology Solves
Siloed visual data like photos and videos can't be directly queried. Taxonomies can label things, but they don’t explain the relationships between them (e.g., “this price tag belongs to that Stock Keeping Unit (SKU) on this display”). This semantic gap is exactly what most teams are missing when they try to compute KPI automation at scale.
An ontology solves this by creating a unified layer of business meaning [1]. This shared data model across teams leads to consistent KPIs, reusable models and faster onboarding of new products or promotions. The same ontology lets analysts, engineers, field teams and vendors "speak" the same objects, attributes and rules. For a broader view on how this meaning layer fits into the wider CPG technology ecosystem, see our blog From Silos to Synergy: The CPG Tech Stack and the Role of Visual AI.
Here are some of the KPIs you can automatically derive from images and videos using an ontology:
- Share of Shelf (SoS)
- On-Shelf Availability (OSA)
- Price & Promo Compliance
- Planogram vs. Realogram compliance
- Facings count
- Asset execution
- POS display effectiveness
Ontology vs. Taxonomy: A Key Distinction
This is a critical point that differentiates a truly powerful system from a simple one.
- A taxonomy is a list of named buckets and hierarchies: Beverages → Soft Drinks → Cola. It's great for classification, but it has limited expressive power [2].
- An ontology takes that taxonomy and adds relationships and constraints. For example, a rule like PriceTag → belongs to → SKU or SKU → is displayed in → Cooler Bay 3. This shift from simple labels to rich relationships is the shift that unlocks operational use.
The leap to an ontology is a shift from merely organising data to creating a dynamic knowledge graph. It allows you to infer new knowledge from existing facts.
Ontology vs Database: Different but Complementary
A relational database is where your data lives; rows, keys and joins that store facts with governance and performance. An ontology, by contrast, is the “meaning layer” that defines how those facts relate and what rules apply. In practice, the two work together: your master data (e.g., UPCs, brands, locations) remains in the database, while the ontology references those records and applies relationships like:
“promo applies to these SKUs”
or
“this price tag belongs to that product in this cooler bay.”
This separation makes the system more flexible: you can evolve rules and relationships without rewriting schemas or pipelines, while still compiling meaning back into the database when speed or reporting requires it.
The Lego Analogy for an Ontology in Retail
Here’s an easy way to picture it: imagine your ontology as a giant box of Lego.
- Bricks = Your spatial primitives. These are the items detected in images, like SKUs, price tags and regions of interest (ROIs) such as a shelf or display.
- Sets = Your Catalogue Digital Twin. Each brick connects to a product master record (UPC, brand, size). This is how a generic "bottle of soda" becomes your specific SKU with all the right attributes.
- Instructions = Your Logic Rules. These are the rules that allow you to assemble the bricks into meaningful KPIs. For example: "Associate price to nearest SKU within regions of interest (ROI); group SKUs by size; compute SoS for a target brand."
Once your bricks, sets and instructions are defined, you can build new analyses quickly without manual relabelling or bespoke pipelines, simply by changing the instructions.
How We Use an Ontology at Neurolabs
At Neurolabs, our ontology is the backbone that turns pixels into decisions. We use a three-layer design to process complex visual data and produce actionable insights.
Layer 1 — Spatial Understanding: We first detect and localise the primitive concepts in an image: regions of interest (shelf, cooler, display), SKUs, price tags and promo assets.
Layer 2 — Catalogue Digital Twin: This is where we connect the raw detections to your product master data. This binding is how, for example, ‘a can of Red Bull’ becomes your specific SKU with a Universal Product Code (UPC), brand and size.
Layer 3 — Logic-based Reasoning: This is where the real power lies. This layer expresses business rules and computes KPIs automatically. For example: Read price from promo tag → group SKUs by brand → associate price with SKUs/brand → compute price compliance index.
Because our ontology cleanly captures these both logical and spatial relationships, you can recompute KPIs as reality changes, all without having to do a costly data refactor. (See Encord for how labelling ontologies structure objects/attributes/relationships [3]).
Ontology in Action
The three-layer ontology works together to solve complex CPG problems and provides real-world insights.
- Cooler Price Compliance: The ontology recognises spatial concepts like the area of the cooler (e.g. ignoring the window), SKUs, prices and promotional materials within a cooler. It then recognises catalogue properties like Universal Product Code (UPC) and promo information. Finally, the logic layer reads the price from the promo tag and associates the price with the product to ensure price compliance.
- Promotional & Display Compliance: For product displays, the system recognises spatial concepts like SKUs and promotional materials, including both single items and multi-packs. It then recognises the UPC from the catalogue. The logic layer reads the price from the promotion and associates it with the correct SKU. To see how effective display execution translates into measurable business impact, explore our blog How Leading Brands Unlock Significant ROI from POS Displays.
- Asset Execution: The ontology can be used to recognise a specific rack within a region of interest. By recognising the rack from the catalogue, the logic layer can then associate the catalogue item with the ROI to ensure correct placement.
What Comes Out of an Ontology?
By applying the logic layer to the visual data, an ontology produces a stream of structured facts. These facts can be used for KPI automation and provide a single source of truth for different teams.
- Share of Shelf (SoS) by brand, Price per SKU and size
- On-Shelf Availability (OSA) and out-of-stock (OOS) alerts
- Price & Promo Compliance
- Planogram vs. Realogram adherence
- Facings and presence counts
This data feeds directly into BI tools, retail execution apps and ERPs, providing a real-time, objective view of store performance.
What’s Next: Ontologies Meet Agentic AI
The next frontier is combining ontologies with agentic AI. As data platforms like Databricks, Snowflake and Fabric evolve toward conversational and action-oriented layers, ontologies provide the structured facts and rules these AI agents need to act with confidence. Imagine asking, “Which retailers under-delivered promo compliance last week?” and the system not only answers but also suggests reallocating spend or generating field tasks. By grounding generative and agentic AI in an ontology, CPGs can minimise errors, shorten the loop from detection → decision → action and unlock a new level of intelligent automation.
A Note from Patric Fulop, Co-Founder & CTO of Neurolabs
An ontology is not a trendy buzzword, it's a foundational framework that makes visual data machine-readable and business-actionable. With Lego-like building blocks (bricks, sets and instructions), you can assemble any KPI or check you need consistently, quickly and at enterprise scale.
The most powerful solutions are simple. We've built our engine with this in mind. You can read more about the inner workings of our image recognition technology in our blog, "From Images to Insights". If you're ready to see how a real-world ontology can transform your retail execution, connect with me on LinkedIn or request a demo.
—
Sources
[1] Ontology: Finding Meaning in Data, Palantir, 2024. https://blog.palantir.com/ontology-finding-meaning-in-data-palantir-rfx-blog-series-1-399bd1a5971b
[2] Taxonomies Versus Ontologies: A Short Guide, Fluree, 2024. https://flur.ee/fluree-blog/taxonomies-versus-ontologies-a-short-guide/
[3] Computer Vision Ontology Definition, Encord, 2024. https://encord.com/glossary/computer-vision-ontology-definition/
.jpg)