Same person or not? Two crop columns, aligned row-by-row to the SAME wall-clock instant in both cameras. Keys 1 / 2 / 3. The system's own answer is never shown — seeing it would turn your label into agreement instead of truth. This is what replaces the circular metric.
Animated explainer of the next accuracy upgrades: co-location must-link, vanish/appear assignment, constrained clustering — with the safety gates and measured baselines.
Both cameras side by side, synchronized — play or scrub the ENTIRE window freely with G-id overlays, jump to events, and watch everyone's approximate position on the 2D floor map.
Browse every confirmed entry, click a person for the corridor→lobby splice with live bounding boxes. Windows: night, stress, demo, plus the QA spans (qa01 14:21, qa02 11:31). Toggle "boxes: everyone" for the crowd view.
Side-by-side calibration of the inter-camera clock offset (currently −1.6s, measured).
Realtime entry alerts with per-person colored IDs. Start with:
python -m ptrack.live --replay 2026-06-30T08:57:00 (or --rtsp for NVR feeds)
Watch offline analysis run on both cameras side by side, wall-clock synced. Start with:
python -m ptrack.analyze --window stress
start_tracker.bat (or python -m ptrack.server) in E:\gndall\tracker