Examining The Relationship Between Coronavirus and Logo Detection

BY ADAM GROSSMAN

President Donald Trump’s recent diagnosis for the coronavirus has, among other things, shone a spotlight on efficacy for both main types of coronavirus testing - rapid testing and polymerase chain reaction (PCR) testing. We can use probability analysis to understand these issues and see how they apply to logo detection for a use case applicable directly to sports sponsorship.

The foundation for examining coronavirus testing is using Bayesian analysis. Bayesian analysis uses conditional probabilities to examine outcomes across multiple events or elements. From a coronavirus testing perspective, we are most concerned with the following questions:

  • What is the probability that I am negative for the coronavirus given that I tested negative (i.e. test specificity)?

  • What is the probability that I am positive for coronavirus given that I tested positive (i.e. test sensitivity)?

One key element with Bayesian analysis is called a “prior.” In this case, knowing how high of a risk a person had of having COVID prior to taking the test can impact the probabilistic confidence in the test’s results. As of early September, the base rate for the coronavirus in the United States was 1.9% on a per capita basis. 

Test creators consider base rates when trying to find the right balance of sensitivity and specificity to maximize the number of true (or correct) outcomes. The challenge with the coronavirus is that the average person is unlikely to have the disease and having the disease is difficult to detect.

This combination makes false positives unlikely because most people do not have the virus. However, it has also led to low sensitivity (true positives) and false negative rates for PCR tests to be as high as 30% and for rapid tests to be as high as 50%+.

The high false negative rate does not have a significant impact in the confidence of a negative test because the true negatives far outnumber the false negatives in an average or low-risk population. This dynamic changes, however, when a person becomes a higher risk for having the disease like Trump.

We have created B6A’s COVID Model (access the model by clicking on the link) to illustrate this point. The model requires three inputs. They are:

  • Did you take a rapid test?

  • What is your risk profile?

  • What is the quality of the test?

The model uses Bayesian analysis to produce a probabilistic assessment for both positive and negative test results. By changing the inputs to a higher risk or lower quality of test, one can see the increase in false negatives and decrease in the lower likelihood of an accurate negative test result. This arguably is the exact opposite of the desired outcome of coronavirus testing (i.e. having confidence that a negative result is accurate).

That is also one reason to be critical of Trump’s actions with regards to his communication around his testing. Reporting from The Wall Street Journal states that Trump received a rapid test positive prior to early in the evening last Thursday (prior to an appearance on Fox News where he could have said he was confident he was positive). He only communicated he was positive after receiving a positive PCR test result early Friday morning as he appeared to want to be surer that he was positive.

While rapid tests are less accurate than PCR tests, a positive result given Trump’s risk profile (a close aide to Trump had recently tested positive for coronavirus) means he was more likely to be positive. In fact, Trump should have been less confident in a negative outcome even with the rapid test as it was more likely to be a false negative. In other words, Trump arguably took the “wrong” approach to examining testing efficacy which led to “wrong” strategic decision-making and communication strategy.

We can apply the challenges of coronavirus testing to logo detection in sports. Logo detection essentially works by breaking broadcast, streaming, digital, or social media into individual frames for analysis (B6A breaks down media into four frames per second). The challenge is that most frames do not contain a specific partner’s logo and a logo can be difficult to detect.

To address this issue, properties and partners work with companies like ours to use logo detection technology (in our case computer vision and optical character recognition) to identify exposures. Both properties and partners are (understandably) biased against false positives – a system or person stating that a logo is present in media when it is actually absent.

This is a complementary problem to coronavirus testing. Instead of companies optimizing systems for true negatives, however, detection is optimized against false positives. This creates the issue where logos that are present in video will be undercounted given the bias against false positives. This arguably is the exact opposite of the desired outcome of logo detection (i.e. having confidence that a positive result is accurate).

We address this issue in our Media Analysis Platform (MAP) by optimizing our machine learning system against false positive and false negative issues using a two-phased approach. We use a lower positive threshold (i.e. have false positives) in our initial training set for new client engagements to ensure we better know how a logo looks in media.

B6A’s team then uses our Visual Verify platform to remove false positives from these initial results and re-calibrate MAP weightings to best optimize both sensitivity and specificity going forward. This re-calibration is critical to achieving the desired outcome for logo detection and is not done in the same way by other logo detection companies.

Both coronavirus testing and logo detection will likely never achieve perfect results. The goal, however, is to maximize the probability of accurate results. Using probabilistic analysis to understand how and what types of errors can occur can help lead to better outcomes.