Overview

AVA counts are conditional evidence, not neutral censuses. This chapter explains how distance, occlusion, detector-versus-tracker choices, frame rate, hardware, demographic classification, and data gaps shape instantaneous and cumulative billboard counts. It uses the benchmark’s tested results—including lower recall for farther and occluded people and changing cumulative error across 0.25–30 frames per second—without treating them as universal rates. It then distinguishes internal benchmark performance from external validity and sets a claim-specific validation standard for live OOH deployments, including model adjustments, bias, privacy, and missing data.

Learning outcomes

Distance and occlusion create structured recall differences, so an overall average can conceal who is missed in an exposure zone.
Detector and tracker choices, frame rate, and hardware affect identity persistence and cumulative counts; more frames do not guarantee better results.
Validation must match the claimed metric and deployment, covering definitions, bias, data gaps, recency, privacy, and external validity.

Keywords

trackingocclusionrecallsamplingbiasvalidation

The Error Is Part of the Number: Tracking, Bias, and Validation

An AVA result is conditional evidence, not a neutral census. This chapter focuses on how distance, occlusion, tracking, frame rate, demographic classification, and deployment context change a billboard count, and on what must be validated before that count is used in a live OOH audience decision. The practical skill is to read a performance result together with its conditions, not as a promise that the same number will hold everywhere.

Failure is structured, not random

A person far from a camera is not simply a nearby person with a little more random noise. The image contains fewer pixels, faces and bodies are harder to distinguish, and small changes in position can change whether the person is detectable at all. Occlusion creates a related problem: a person may be partially hidden by another person, a vehicle, or part of the environment. The failures are concentrated in particular places and conditions, which means that an overall average can conceal exactly the people most likely to be missed.

The benchmark reports this pattern directly. In one GPU system, a cited face detector had median recall of 0.94 for closer faces and 0.68 for farther faces. The person trackers had median recall around 0.91 for closer people and 0.55 and 0.63 for farther people. For the cited algorithms, recall for partial occlusions fell below 0.6 and for heavy occlusions below 0.5. These values belong to the tested systems and scenes; they are not universal AVA rates. Their importance is diagnostic: distance and obstruction create systematic exposure-selection effects in the count. [AVA]

For a billboard, that matters because the people most difficult to localize may not be a negligible fringe. A camera angle can make distant travelers the dominant part of the roadway view. A curbside crowd can create frequent overlap. If an algorithm performs well on clear, close subjects but poorly on far or occluded subjects, the resulting total can be biased toward the easiest people to see. A report that gives only an overall recall or total count may make the result appear more stable than it is across the exposure zone.

Detector and tracker choices create a second structured difference. A detector evaluates frames independently. If the same person appears in ten processed frames, the system may assign ten new identities. That can produce a reasonable instantaneous count and a badly inflated cumulative count. A tracker tries to associate observations over time, which can lower duplicate counting, but it needs more computation and can drift when a person is hidden, exits and re-enters, or is confused with another person.

The benchmark’s measures make this distinction visible. Mean Opportunity Error, or MOE, concerns the number of people with OTS at a given instant. Cumulative Opportunity Error, or COE, concerns the unique cumulative number over a whole video. Temporal Cumulative Opportunity Error, or TCOE, tests unique counts over segments of different durations. An algorithm can therefore be close to the reference in a frame while diverging over a longer interval. The error is not one thing that simply scales with the number of people; it changes with identity persistence and time.

Published by DeepMox

FAQ

Why does distance or occlusion bias an AVA billboard count?

Distance and occlusion bias an AVA billboard count because smaller or partially hidden people are harder to detect and track, so recall can fall for those groups. An overall average may conceal this structured failure across the exposure zone.

Can processing more video frames improve a unique people count?

Not necessarily. More frames can give a detector more opportunities to assign duplicate identities, while tracking and hardware create different cumulative error patterns; the benchmark varied 0.25 to 30 frames per second and found no universal best rate.

What does external validity require for live billboard AVA?

External validity requires evidence that the benchmark rules, scenes, people, conditions, and population resemble the live billboard deployment. Providers must validate the claimed metric and adjustments for relevant distance, occlusion, travel, data coverage, and setting rather than extrapolate confidently.

Previous14 minOOH Billboard Analytics: What AVA Measures

Next11 minOOH Billboard Claims You Can Defend

OOH Billboard Analytics: Tracking Error and Bias

The Error Is Part of the Number: Tracking, Bias, and Validation

Failure is structured, not random

FAQ