Right now
Shopper funnel
This is the number an FMCG brand cares about: of everyone who passed the shelf, how many stopped, touched, and took a product.
Where hands went on the shelf
Every reach into the shelf, on an empty-shelf image the AI built by removing the people. Hot spots = the products that got touched.
Where people walked and stood
How it works
1. Person detector + skeleton (YOLO26s-pose, runs on a plain CPU at ~10 frames a second).
2. Tracker gives each shopper an id and follows them (ByteTrack).
3. Simple rules: still for 1 s = stopped, wrist stretched out from the shoulder = reach.
4. For each shopper who stayed a second or more, 6 to 7 crops go to a vision model that answers: touched? took? put back? carrying what?
5. Only numbers and short text leave the store. No faces stored, no face recognition.
Every shopper
Click a card to jump to that moment.
Proof: AI vs a human
Each clip in the dataset came with a description written by a person. Left: the human. Right: what the AI read, with no access to the human text.