Net-takeaway recognition
Hand–object tracking plus an event state machine accumulate takes and put-backs event by event. The net result is fixed at door-close. What gets settled is behavior, not pixels.
Polaris is a vision foundation model built for smart vending cabinets. From door-open to door-close it tracks every take and every put-back, then outputs a net-takeaway list you can invoice — not what the camera saw, but what the customer owes.
Raw in-cabinet footage with Polaris's visualization overlaid: tracking boxes, event states and the running take / putback / net counters. Nothing staged — this is the engine's original output.
Every overlay element corresponds to one core mechanism of the engine.
A product only moves because a hand moves it. The engine tracks hands and products jointly; only hand-driven motion can trigger an event, so rack vibration and lighting shifts never become transactions.
When a product is hidden behind a hand or body, its box is kept alive and the event state machine keeps reasoning — until the item is visible again or has clearly left the cabinet. Out of sight is not out of the order.
Takes and put-backs accumulate in real time; net = take − putback. The count is final the moment the door closes — that one line of numbers is what the settlement is based on.
Once the net count is fixed, SKU recognition only has to answer "which products are these" — on the most confident frames. The output is a settlement list ready for invoicing; uncertain orders route to the cloud audit loop.
Hand–object tracking plus an event state machine accumulate takes and put-backs event by event. The net result is fixed at door-close. What gets settled is behavior, not pixels.
First fix how many items left the cabinet, then identify each one. Recognition effort collapses onto a few key frames and a known count — stable settlement output even against a 10,000-SKU library.
Boxes stay alive through occlusion and identities are never dropped. The physical prior — products move only with hands — filters out rack collapses, lighting jumps and every other non-transaction disturbance.
Low-confidence orders route automatically to an audit channel; cloud review writes the settlement back, and reviewed samples flow into training. Every hard order makes the model sharper.
* Illustrative targets — ask us for the measured evaluation report.
Six stages, fully traceable. Every intermediate conclusion of every order can be reviewed.
The in-cabinet camera records door-open to door-close. One order, one clip.
Hands and products are associated frame by frame: who moved what.
Takes, put-backs and occlusion keep-alive resolve each item's final fate.
net = take − putback. The count is final when the door closes.
Post-hoc recognition confirms each item's identity and price.
The settlement list ships via API; hard orders go to cloud audit review.
Swipe to see all six stages →
An edge-cloud package engineered into the cabinet itself: camera selection, mounting positions and calibration are all part of the delivery, with a clean split between on-device inference and the cloud engine.
The recognition API plugs into your platform's settlement system and serves mixed cabinet fleets with one consistent standard of accuracy and evidence.
Every order carries a net-takeaway list and a video evidence chain. Disputed orders trace back in minutes, and the loss column in your P&L finally itemizes.
If you run your own system, the settlement API drops straight into it — integration measured in weeks, not quarters, with the cloud audit loop included from day one.
A 30-minute technical demo on footage from your own cabinets — accuracy and latency measured live, end to end.