Transaction video baseline

The built-in adapter uses OpenCV 5 ArUco markers to associate product instances over time and reconcile the visible bagging inventory with net scan counts at payment.

This is a marker-assisted instrumentation baseline. It does not recognize unmarked grocery products, infer intent, or validate a commercial checkout model.

• Real MP4, up to 60 seconds, 250 MB, 1920 × 1080, and 60 fps.
• Fixed camera, bagging zone on the right, initially empty inventory.
• One distinct DICT_4X4_50 marker per product instance (IDs 0–49).
• All merchandise must be marked. Labels must be independently reviewed.
• Keep product markers visible through payment. Occluded or untracked items may be outside the baseline's ability to detect.
• Use mock scan events, never actual customer payment data.

External model predictions can be imported for scoring. Their inference is labeled unverified.

Image perception

The Real-data lab runs classical OpenCV methods and photo-trained Faster R-CNN on original RPC photographs. SIFT estimates SKU identity from training examples; Faster R-CNN estimates product locations without SKU identity. These experimental methods miss many objects; image results do not establish scan or payment status.

Deployment status

This application is for a trusted local operator. Bind it to 127.0.0.1. For the login-free hackathon demo, use local access or a private AWS Session Manager tunnel. A public write endpoint is not supported.

Contribution protocol

Record only with permission, keep faces and identifying audio out of frame where possible, and review physical inventory separately from the scan log. Evaluation permission does not imply public-release permission.