Case study 002
Photo AI. An archive that tags itself.
LiveStage one is live. Stage two, mobile capture and trends, is scoped.
- Coverage
- 100%of new images tagged and normalised on ingest
- Manual effort
- 0hand-tagging steps between capture and searchable
- Stage
- 1 of 2tagging is live. Capture and trends come next
01Context
A photo archive nobody could search
Years of store and shelf photography, shot on cameras by researchers in the field, loaded onto laptops and pushed up in batches. Inconsistent framing, lighting and naming meant the archive was only useful to the person who took the pictures.
The frontier report had tracked vision-language models crossing from demo to deployable for retail imagery. Photo AI was the concept that followed.
02What we built
A pipeline that normalises and tags on ingest
Every image is normalised as it lands: orientation, exposure, cropping and naming brought to a common standard.
Then a vision model tags it against a taxonomy IGD's researchers already use, so retailer, category, fixture type and promotional mechanics are searchable across the whole archive rather than remembered by whoever was in the store. Basic by design in the first stage: reliable tags first, clever analysis second.
03How
Multi-model, tested against the archive first
Candidate vision-language models were benchmarked in the lab against a hand-labelled slice of IGD's own images before anything was built.
The pipeline uses more than one: a fast model for normalisation and coarse tags, and a stronger one where confidence is low. Storage and processing run on Edge alongside igd.com, so the archive never leaves the platform that serves it.
04What changed
The archive became a dataset
Because every image now carries structured tags, the archive is for the first time a dataset that can be analysed over time. That's what the second stage is for.
Before
Findable by who took the picture, or when. Useful once someone had sorted it.
Now
100%of new images tagged on ingestSearchable by what's in the picture. Usable the moment it's uploaded.
05What's next
Capture on a phone, trends across the archive
Two moves, both scoped in IGD's quarterly report.
- 01A mobile capture app that streams straight into the pipeline, so a shelf is tagged before the researcher leaves the aisle
- 02Trends read across the archive: share of shelf, promotional intensity and range changes across retailers and quarters
Run by Edge Expert Services. Extended by Labs when IGD says go.
Engagement
A rapid build with a second act designed in.
Photo AI was deliberately scoped small: get tagging reliable and useful, then earn the right to do trend analysis. It's a good example of a Labs concept that starts as a fortnight's work and grows into a roadmap.
Photo AI 002
Live- Client
- IGD, a customer of Edge since 2024. Edge hosts and runs igd.com.
- Sector
- Food & grocery. Membership and research body.
- Status
- Live, stage one. Stage two scoped.
- Labs' role
- Frontier report
- Concept and spec
- Rapid build
- Expert Services
- Now run by
- Edge Expert Services
- Stack
- Multi-model vision pipelineTaxonomy-constrained taggingEdge StorageEdge Compute
- Related
- AskIGDThe IGD sample report
More work
All seven buildsData you already own, made useful
Most organisations are sitting on an archive like this. If yours is images, documents or recordings that only their author can find, tell us about it.
Or write to labs@edge.network