AdPixTechnologies

Computer vision · Advertising infrastructure

Ad systems monetise media they cannot see.

Most ad stacks still lean on URLs, keywords and history — everything around the image, and nothing inside it. AdPix reads the picture itself and returns structured signal you can match demand against.

See a worked example Talk to us about your inventory

01 The gap

Two engines, one pass over the image.

Recognition establishes what the media contains. Recommendation decides what belongs beside it. Take both, or take the signal and rank it yourself.

  • Input

    The media already passing through you

    Editorial imagery, user uploads, video frames — whatever runs across the properties you monetise. No tagging work for your publishers, no schema for them to adopt first.

  • Recognition

    Objects, scenes, materials, brands

    Models locate and label what is actually in frame — garments, furniture, food, logos, setting, on-image text — with a confidence value on every detection.

  • Recommendation

    Ranked against live demand

    Those labels are matched and ranked against the campaigns you already have booked, so the creative that serves follows from what the viewer was looking at.

02 Worked example

One image, end to end.

This is an illustration, not a live demo. The annotations and payload below were prepared by hand to show the shape of the output. To see our models run on your own media, get in touch.

A person photographed from behind on a city pavement, wearing a backwards cap, a hand-painted orange patchwork jacket, painted light trousers and white trainers.
Fig. 1 — boxes drawn from the bbox values opposite
Response application/json 
{
  "media_id": "img_8f21c0",
  "scene": "street / daytime / urban",
  "detections": [
    { "label": "jacket",   "conf": 0.97,
      "bbox": [0.234, 0.148, 0.468, 0.374],
      "attrs": ["outerwear", "patchwork", "orange"] },
    { "label": "trousers", "conf": 0.95,
      "bbox": [0.318, 0.494, 0.309, 0.396],
      "attrs": ["denim", "printed", "loose"] },
    { "label": "cap",      "conf": 0.94,
      "bbox": [0.352, 0.038, 0.146, 0.064] },
    { "label": "trainers", "conf": 0.91,
      "bbox": [0.382, 0.851, 0.264, 0.114] },
    { "label": "car",      "conf": 0.88,
      "bbox": [0.000, 0.264, 0.206, 0.118] }
  ],
  "faces": 0,
  "logos": []
}

bbox — normalised [x, y, width, height], origin top left

Matched against a sample demand pool

  • Outerwear advertiser — autumn flight Object Matched on jacket + outerwear, 0.97
  • Footwear brand — new season creative Object Matched on trainers, 0.91
  • Streetwear retailer — always-on line item Scene Matched on scene and combined garment attributes

Home & furniture

A bright modern living room with two grey velvet sofas facing each other across a marble coffee table.
sofacoffee table rugfireplacewood flooring

Grocery & kitchen

A home kitchen counter with a gas hob, olive oil bottle, pepper mill, stone mortar and pestle, and folded tea towels.
gas hobolive oil mortar & pestlepepper milltea towel

And the rest of it

Apparel, home, grocery, travel, automotive — the label set is tuned to the categories your demand actually buys.

your categoriesyour taxonomy

03 Approach

Off-the-shelf labels rarely match how demand is bought.

A generic model returns “shoe”. Your advertisers are bidding on forty kinds of shoe. The work is in the gap between those two facts.

  1. 1

    Start from your supply

    We take a slice of the media running across your properties and the categories your demand is organised by, and establish what a generic model already gets right — and where it falls short.

  2. 2

    Train to your taxonomy

    Models are tuned to the labels your business actually uses, at the granularity your ad ops and campaign teams need to act on.

  3. 3

    Ship, then correct

    Wired into your pipeline and measured against whatever you were doing before. The labels that come back wrong get corrected, and those corrections are the training data for the next pass.

Delivery

Hosted API
A REST endpoint. Send media, receive structured JSON.

Delivery

Batch
Backfill an existing library once, then process new media as it arrives.

Delivery

In your infrastructure
Deployed inside your own environment where media cannot leave it.

Scope

Signal only
Take the recognition output and match it to demand yourself.

04 Contact

Tell us what you are sitting on.

If you monetise media you did not create — at volume, across properties you do not control — and you know far more about the page than the picture on it, there is probably something here worth a conversation. Tell us roughly how much media moves through you, and what you would want to do with it.

Write to

contact@adpix.co.in