About Services Tools Blog Contact us

Webcam Shopper Style Profile

Looks at a webcam photo of a shopper and returns an age range, how they present, which part of the body is in view and one sentence a product search can use.

  • Visual Question Answering
  • Footwear retail
  • Pipeline
  • Productised
  • Local/private
  • 2026

The problem

A fitting booth that only knows which button was pressed recommends the same rail to everyone. The useful cue is what is in frame: feet, torso or full length, and the style of what the shopper is already wearing.

What it does

The booth sends a webcam photo to a model running locally. The reply is a short set of labelled details: age range, how the person presents, the visible area (upper body, lower body, feet or full length), style tags and one sentence limited to what can be seen. If the preferred model is unavailable, a lighter one is tried. The result stays with the booth's current visit.

Outcomes

  • The booth knows whether it is looking at feet, torso or a full figure.
  • Product search gets one sentence in place of a raw photo.
  • The profile is produced on the machine.
  • A lighter model covers the gap when the preferred one is unavailable.

Simplified mockup

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

Booth 2

Webcam still
Webcam crop of white trainers with gum soles, cut off at the ankle
Local vision
Crop stays on this booth

Earlier reads on this session

10:41 · Feet
The feet crop from earlier in the session
30 to 40 Feminine Feet
White leather trainer, gum sole, cropped at the ankle.
10:44 · Torso
The torso crop from later in the session
30 to 40 Feminine Upper
A navy overshirt, no footwear visible.
Search uses the sentence for the crop in front of the camera. The other read stays on the session until the shopper leaves.

Model fallback

Preferred model is down
A missing tag falls back to a plain default: adult, a neutral presentation, casual, and the full region. Garments outside the frame are not invented.
OrderModelResult
1Preferred vision modelnot loaded
2Lighter local modelstanding by

How it works

  1. Send a webcam photo to the model.
  2. Ask it to reply with a short set of labelled details and nothing else.
  3. Read the age range, presentation, body area, style tags and one sentence from the reply.
  4. Limit the sentence to the area actually in view.
  5. Replace a missing detail with a plain default.
  6. Try a lighter model if the preferred one is down.

Pipeline

flowchart LR
  A["Webcam crop"] --> B["Gemma on Ollama"]
  B --> C["XML block"]
  C --> D["Age band"]
  C --> E["Body region"]
  C --> F["Style sentence"]
  F --> G["Product search"]
  B --> H["Lighter model if needed"]

Output