About Services Tools Blog Contact us

Product Identity from Stills

Reads grouped product photographs, including label shots, and fills in the identity fields for that colour: what the product is and how it is described.

  • Visual Question Answering
  • Footwear retail
  • Pipeline
  • Productised
  • Hybrid
  • 2026

The problem

A delivery of photographs does not name itself. Before any catalogue text could be written, staff had to open each group and decide what the product was, which colour it was and whether the label agreed.

What it does

The job picks up styles nobody has looked at yet. It sends the photographs, including label shots, to Google Gemini, with any text already read from the label, and Gemini fills in the identity details. A style that fails three times is left with the reason recorded. The same pass can be started from the import page.

Outcomes

  • A photo group gets identity fields without anyone opening each frame.
  • Label text and pictures are read together.
  • A style that will not read is stopped with a recorded reason.
  • The import page can start the same pass.

Simplified mockup

Interactive preview with sample data. Nothing is sent to a server.

Styles not yet read

AL-4418 Ash Lane
Not looked at
Side still of a tan oxford
Side
Top still of a tan oxford
Top
Label still reading Ash Lane, oxford, tan
Label

OCR on the frame: ASH LANE · TAN · 4408

Pictures, then the label, then the fields.
TypeNot read
ColourNot read
LineNot read
PL-1902 Pewter Last
2 failures
Blurred side still
Side
Dark top still
Top
Label still too soft to read
Label
  1. 1 No fields returned
  2. 2 Label could not be read
Two failures stored. A third stops the style.

Import page

The same vision pass can be started here. Styles already stopped are left alone.

StyleNameStillsLabelStatus
AL-4418Ash Lane3YesNot looked at
PL-1902Pewter Last3Soft2 failures
BR-2201Briar Row4YesNot looked at
KC-1180Kelp Court2NoNot looked at
Nothing queued yet.

How it works

  1. Pick up styles that have not been looked at.
  2. Send the grouped stills, including any label photograph, to Google Gemini.
  3. Include text already read from the label with the pictures.
  4. Write the identity fields: what the product is, the colour and a short description.
  5. Stop the style after three failures and store the reason.
  6. Start the same pass from the import page when needed.

Pipeline

flowchart LR
  A["Grouped stills"] --> B["Vision pass"]
  C["Label OCR text"] --> B
  B --> D{"Fields filled?"}
  D -->|Yes| E["Identity record"]
  D -->|No| F["Count failure"]
  F --> G{"Third failure?"}
  G -->|Yes| H["Store reason and stop"]