DemonstrationEvery number on this site is computed from real pipeline output at build time.
Alert evidence
Every alert can persist an annotated snapshot. A stratified
sample was labelled by hand: 815 verdicts across three
soak runs. These are the frames the reviewer saw, true detections beside the
false alarms that measured and then improved each detector.
near-miss
Figure 1. Camera 3238, gap px 8.5. Verdict: marked real by the reviewer.Figure 2. Camera 3252, gap px 1.7. Verdict: marked real by the reviewer.Figure 3. Camera 3238, gap px 15.8. Verdict: marked not real; alarms like it set the detector gates.Figure 4. Camera 3252, gap px 1.5. Verdict: marked not real; alarms like it set the detector gates.
tailgating
Figure 5. Camera 3238, time gap s 0.51. Verdict: marked real by the reviewer.Figure 6. Camera 3252, time gap s 0.7. Verdict: marked real by the reviewer.Figure 7. Camera 3238, time gap s 0.77. Verdict: marked not real; alarms like it set the detector gates.Figure 8. Camera 3252, time gap s 0.55. Verdict: marked not real; alarms like it set the detector gates.
pedestrian or cyclist in the roadway
Figure 9. Camera 3252, cell vehicle hits 4985. Verdict: marked real by the reviewer.Figure 10. Camera 3379, cell vehicle hits 48. Verdict: marked real by the reviewer.Figure 11. Camera 3238, cell vehicle hits 38. Verdict: marked not real; alarms like it set the detector gates.Figure 12. Camera 3252, cell vehicle hits 62. Verdict: marked not real; alarms like it set the detector gates.
wrong-way
Figure 13. Camera 3238, heading share 0.0028. Verdict: marked real by the reviewer.Figure 14. Camera 3252, heading share 0.0022. Verdict: marked real by the reviewer.Figure 15. Camera 3238, heading share 0.0167. Verdict: marked not real; alarms like it set the detector gates.Figure 16. Camera 3252, heading share 0.0097. Verdict: marked not real; alarms like it set the detector gates.