The right infrastructure, architecture and policies can accelerate progress.
In computer vision, government IT leaders have a head start (at least for non-classified workloads), insofar as all the major cloud providers have tools in place to support such an implementation. For example, AWS, Google, IBM and Microsoft have APIs to make ingesting your visual data affordable and scalable, freeing IT leaders to focus on the business value of the effort.
Internally, there may still be some technical hurdles to cross. The cloud can eat all the video you want to feed it, but you’ll need sufficient internal resources to take in what computer vision has to offer in return. These (possibly sizable) outputs will require scalable systems, as well as the ability to integrate the newly generated intel into your existing systems. What processes will you need to put in place in order to ingest, store and ultimately utilize the new information?
These are the technical issues. At a deeper level, IT leaders and business line owners need to go beyond the mechanical details, to think about the AI itself and consider what it is they are asking computer vision to do, and why.
First there’s the question of return on investment: What will be the actual business benefits of automation and will they justify the cost? If an agency already has video in its toolkit, what will be the practical benefits of leveraging AI against that data? If you aren’t collecting video already, will the outputs of computer vision justify the expense of a video deployment? Few can afford to do video for its own sake; you’ll need to build the business case.
Then there are issues around personnel and processes. Who’s to take on this new task? In all likelihood, responsibility for computer vision won’t fall to the IT leaders alone. In addition to business line leaders, who will need to make the operational case, computer vision will require the active participation of subject matter experts.
If you’re going to be reviewing images of a domestic situation, especially in jurisdictions that require advance permission prior to being videotaped, the proper person to train the AI is the social worker who visits that home, as this person will best understand the complexity and potential risk involved. This holds true with virtually any potential use of computer vision: The human expert has to be at the center of the situation. If we want the computers to see what we see, that information can be conveyed most effectively by those who are most experienced at looking.
Security considerations also factor in. Do the right analysts have access to the right intel? Are the outputs of computer vision being appropriately classified? If you’re handling secret or top-secret imagery, does that put added constraints on your processing capacity? Similar restrictions exist around the use of personally-identifiable information (PII), especially for minors.
What this points to is the need for responsible AI policies within agencies employing computer vision. These policies should promote the ethical, transparent and accountable use of the technology. For example, the potential for bias in facial recognition technology is receiving fair scrutiny. Furthermore, there is limited consensus about the public use of computer vision to identify individuals without their consent.
Overall, experts advise patience in any initial foray. It isn’t hard to teach a computer to know what a tree looks like, generally speaking. But there are many different kinds of trees and they will look wildly different in still photos and in videos, in sunlight and shadow, on the plains or on a mountainside. Even learning at the accelerated speeds made possible by AI, it will take the computers some time to go beyond the general, to be able to see with the kind of detailed specificity that will unlock the full potential of computer vision.
Regardless of use case, this will be a moving target. Learning is evolutionary. We, as humans, keep getting better the more we do a task, and we never stop learning. Machines should also have these capabilities. Agencies likely will need to periodically adjust the training models for computer vision as new targets of interest emerge or new business goals come into focus, without forgetting what has already been learned. Even once it’s up and running, computer vision will require ongoing care and feeding.
Market analysts say there’s considerable economic energy behind the development of computer vision. Researchers at MarketsandMarkets say the computer vision market will top $17.3 billion by 2023, while Grand View Research puts the figure at $18.2 billion by 2025.
Federal agencies can begin to embrace the new capabilities by starting with what they have on hand. They can use computer vision to index, highlight and annotate existing images in order to make them searchable; in doing so, they’ll be training the AI on how to read and interpret relevant future imagery. In practical terms, object detection and change detection will likely be the first applications of computer vision. They’re achievable even with present-day technology.
For those who can successfully cross the threshold and implement video analytics, the next few years promise to bring significant advances.
In competitions, computer vision can recognize and classify objects with a 3 percent error rate, already better than the typical 5 percent error rate among human analysts. Especially for mission-critical uses, the technology will be getting even more precise. The engineers also are looking to further spoof-proof computer vision, training the machines to distinguish a fake image from the real thing.
All this fine tuning will help to enhance what is already a powerful capability. With its speed, accuracy and persistence, computer vision promises to lighten the load on human analysts while producing higher-quality intelligence.