Mission Impossible, Bourne Supremacy, and Minority Report are movies that we can relate to the application of Facial Recognition (FR). Closer to us, the ability of Facebook to identify friends when we uploaded images and tag them is one of the applications. Facebook gives a successful practical demonstration of FR, most of the time, without manually tagging those images.
Previously, FR used to be a novelty features available to high-end CCTV cameras. FR is becoming a standard feature, and manufacturers can incorporate face finding algorithm at the camera to facilitate the FR function at the server end or even at the cloud. In recent times, technological advancements help bridge the data pipeline’s limitation and allow such architecture located in the cloud.
For FR solution implementors, the performance of the engine tends to be the critical focus. However, in a real-world application, is the engine the only factor determining the entire solution’s performance?
I have been dealing with FR since 2009, started as a novice that focuses solely on engine performance. Seeing the operational environment’s challenges and how other factors affect the entire solution helps shifted our attention away from just looking at the engine.
In this series on Facial Recognition, I like to share our field experiences over the last ten years. Anonymous case will relate to various topics: from dealing with uncertainties, failures to awareness and seeing positive results on FR. Besides sharing what I have seen, I want to take this opportunity to articulate out my thoughts on this particular area. By putting these thoughts into writing, I hope to see things from other angles and probably help someone when they are implementing their own FR project.
Firstly, why do we need FR?
To answer this, we to understand the person’s perspective and industry. From a law enforcement point of view, FR enables the possibility of matching someone within their wanted database. Depending on the circumstances, it could either be a 1:1 or 1:N matching. One of the most challenging situations is performing FR in an uncontrolled outdoor environment. What works in a laboratory environment might not always work in an operational environment.
FR can identify wanted people in an indoor controlled environment and flag them out as top match, almost without fail. Especially during a system demonstration where most of the perimeters are tested and scripted for marketing. This superior demonstration gives false expectations on the FR’s performance and ability to operate in an operational environment where conditions could be dynamic and robust.
Managing expectation is a critical factor in deploying an FR solution, and one must be dare to give the reality instead of promising that everything works. Realistically, FR is a tool that provides the operator with possible resemblances to those people in their wanted database. Depending on direct and indirect factors, the correct person might not appear as the first or second hit on the list.
“In this case, why do we need a machine to do this? Can our operator perform the same task better?” Yes, there are a small group of people that can recognise people accurately, called Super Recogniser. They are very effective in remembering faces, but their number remains relatively small and being human, they need to rest. A more effective way is to enable operators to leverage such technology with proper procedures, integrating Man, Method & Machine. Such an approach is similar to how I approach the digital transformation project by focusing on the Human & Process and leveraging Technology to solve their problem.
FR is not always about finding a wanted person or someone who is already wanted or known by law enforcement. FR can also keep an archive of faces captured by the system for use during a post-incident investigation. Such a feature potentially saves valuable time in going through all video footage and consolidating face images.
Law Enforcement is not the only user of FR, and its potential enables other sectors to use it to solve their needs. Marketing uses FR to identify VIP customers and provide targeted marketing, and the best commonly known example is in the Minority Report. *Healthcare can deploy FR, and in a pandemic situation, use it for contact tracing and cutting down the time taken to identify people for quarantine purpose. Mobile device makers can use FR as one of the means to unlock their devices. Banks are using FR to authenticate transaction for fraud prevention, maybe as one factor to gain ATMs access to mitigate card scamming. Companies and schools are using FR to take attendance, preventing “buddy-punching”. Facility owners are using FR to gain access to the building or a room within the building.
Each scenario above calls for a different setup due to different requirement. However, there are many similarities across all scenarios that work with the fundamental of FR. One important thing to note is that FR does not perform wonders, and it is not without any flaw. When one accepts the conveniences, one must also accept the inevitable shortfall that comes with it.
For example, there will not be a face found by the system when there is no face detected. When the lighting is low, the face obstructs by an object or when the angle is too steep. All these prevents the engine from recognising that as a face. There could also be someone that looks like one of the people in the database, and it triggers a false positive, enabling the person to enter the facility. Such an incident is rare when one is looking at a database of a few hundred in size. However, when a database is multiple million in size, the situation changes.
Advancement in technology does help tackle some of the previous issues, but it is unlikely to eliminate them purely from a technological approach. When something is countering the action, there will be a counteraction that counts the counteraction. I am drawing this example from the sector of Electronics Warfare. When a new function is available to counter a measure, a counter-counter measure will be available.
I will share more on this topic in the following article on facial recognition.
*This article was first published on LinkedIn on 5 June 2017.