Reference Glossary
Object recognition
Computer vision technique that detects and labels physical objects within an image — a car, a chair, a product box — typically returning a category and a confidence score, sometimes a bounding box.
Why it matters in a DAM
Object recognition is what lets a product photography library be searched by 'chair' or 'backpack' without anyone typing that tag by hand, and it's the piece platforms like Cloudinary hand off to services such as Amazon Rekognition at ingestion. But recognition confidence for niche or branded objects — a specific SKU rather than the generic category 'shoe' — is often much lower, because the underlying models are trained on general-purpose object datasets, not a company's own product catalogue.
Common mistake
Teams trust object recognition tags for compliance-sensitive labelling — like confirming a recalled product doesn't appear in any active marketing asset — without a human check, when the model's confidence on that specific SKU can sit far below what a person would assume from a returned 'match'.
Object recognition identifies and labels physical things within an image, distinct from facial recognition (which is specifically about people) and from broader scene classification (which labels the whole image, like ‘outdoor’ or ‘kitchen’). A single photo can return several object labels at once, each with its own confidence score.
In a DAM, this is the mechanism behind searching a product photography library by object rather than by whatever a photographer happened to type into the filename. Cloudinary‘s integration with Amazon Rekognition is a straightforward example: an uploaded image can be run through Rekognition’s object and scene detection, and the results attached as tags automatically, subject to a confidence threshold the team sets.
The gap is between generic and specific recognition. A general-purpose model reliably tells ‘shoe’ apart from ‘bag’, because both are well-represented categories in the datasets these models are trained on. It has much less basis for telling one shoe SKU apart from a visually similar one in the same product line — that’s a finer distinction than the model was built to make, and treating a low-confidence match as a reliable identification is where these systems get misused.
Frequently asked
What does object recognition do, and how is it distinct from facial recognition or scene classification?
It detects and labels physical objects within an image — a car, a chair, a product box — typically with a category and confidence score, distinct from facial recognition (about people specifically) and scene classification (labeling the whole image, like 'outdoor').
How does Cloudinary's integration with Amazon Rekognition use object recognition?
An uploaded image can be run through Rekognition's object and scene detection at ingestion, with results attached as tags automatically, subject to a confidence threshold the team sets.
Why is object recognition confidence often much lower for niche or branded objects?
The underlying models are trained on general-purpose object datasets, not a company's own product catalogue, so they reliably tell 'shoe' apart from 'bag' but have much less basis for telling one shoe SKU apart from a visually similar one in the same product line.
What's the risk of trusting object recognition tags for compliance-sensitive labeling?
Using it without a human check — for example, to confirm a recalled product doesn't appear in any active marketing asset — is risky because the model's confidence on that specific SKU can sit far below what a person would assume from a returned 'match.'
What does object recognition let you search a product photography library by?
By object rather than by whatever a photographer happened to type into the filename — searching 'chair' or 'backpack' directly, without anyone typing that tag by hand.
Can a single photo return multiple object recognition labels at once?
Yes — a single photo can return several object labels simultaneously, each with its own independent confidence score.
Sources
- Cloudinary can automatically classify and tag uploaded images by object and scene category using Amazon Rekognition, with confidence thresholds controlling which detected labels are kept. checked 2026-08-07 — Cloudinary Documentation — Amazon Rekognition Auto Tagging Add-on