Third-party cross-check / validation-v1

External validation

The detector has no manually labeled test set, so its precision, recall, false-positive rate, and false-negative rate are unknown. Public bridge counts expose a large, direction-dependent undercount; they do not make the camera measurements ground truth.

Analysis pages default to a reference-scaled estimate using the two directional ratios below and provide a Raw detector option. The scaled view is an approximate presentation layer—not ground truth or a calibrated correction. These ratios average across different years, locations, sampling definitions, weather, congestion, and camera conditions.

Direction-specific reference ratios

The difference is too large to use one global scale factor. A rough pre-lights, full-day comparison gives 2.45× for Oakland-bound and 1.36× for SF-bound. Likely mechanisms include deck visibility, perspective, lane occlusion, tracker merging, speed/congestion dependence, and different projected motion on the upper and lower decks. Without labeled frames, their individual contributions cannot be separated.

The camera’s apparent two-to-one SF-bound/Oakland-bound volume ratio is therefore contradicted by the official near-even split. It is principally a measurement response, not a traffic conclusion.

Monthly westbound cross-check

MTC publishes monthly toll-bridge totals for the one-way toll direction. For the Bay Bridge this is westbound/SF-bound, so it provides a useful monthly comparison with the camera’s right series. Across the six complete pre-lights months below, the ratio ranges from approximately 1.23× to 1.43×. That stability is encouraging for relative trend analysis, but it still does not establish accuracy at hourly resolution.

SF-bound daily mean: public count vs camera detections

Pre-lights months only · MTC monthly toll count divided by calendar days

Which conclusions survive

Site observation	External check	Publication boundary
SF-bound has a strong weekday morning peak	Qualitatively corroborated: Caltrans’ 2024 toll-plaza record identifies westbound as the AM peak direction	The camera’s exact peak hour is not externally validated
Oakland-bound has a broad afternoon peak	Qualitatively corroborated: Caltrans identifies eastbound as the PM peak direction, with a 2 PM peak-hour record	Do not generalize the exact camera profile to every day
Camera reports about twice as many SF-bound detections	Contradicted as a traffic split: Caltrans AADT is approximately 51% SF-bound / 49% Oakland-bound	Interpret the ratio as direction-dependent detector bias
Sunday is quieter in the camera series	Plausible but not independently checked with a current day-of-week source	Report only as an internal detector observation
Post-relighting nighttime values are noisy	Strongly supported by the detector’s own discontinuity and known Bay Lights timing	Treat as optical contamination, not a traffic surge

For context, the Caltrans peak-hour reference used here records westbound vehicles at :00 for its AM record and eastbound vehicles at :00 for its PM record. These are reference records, not a simultaneous directional survey or a calibration set.

Public sources

MTC Monthly Transportation Statistics — Bay Bridge monthly one-way toll-direction totals.
MTC monthly statistics workbook — machine-readable monthly values used in validation-v1.
Caltrans Traffic Census Program — definitions and access to annual traffic volumes.
Caltrans 2024 AADT workbook — directional west-span reference at Treasure Island.
Caltrans 2024 peak-hour workbook — AM/PM peak directions at the toll plaza.

DETECTOR BASELINE: 191 COMPLETE PRE-LIGHTS DAYS · PUBLIC REFERENCES: MTC 2025–2026 / CALTRANS 2024 · SNAPSHOT: validation-v1