Computer vision, as a key part of AI, is constantly impacting and transforming major industries such as security, retail and healthcare. According to research conducted by McKinsey, the adoption of AI and computer vision is "rapidly taking hold across the global business." Recognizing the potential for tremendous growth in computer vision technologies, major companies including Amazon, Tesla, Apple and Google are dedicating significant resources to different projects in this field.
As the CTO of a company that has developed an innovative gun detection system, I found that the creation of a computer vision solution able to monitor countless data streams in real time is a nonstop process involving constant efforts to further improve your technology in order to bring it to perfection.
However, there are some important lessons I learned in the process of creating and constantly improving our AI. Here are five key issues you should be looking at when working on a computer vision-centered system.
Creating A Successful Strategy
You shouldn't underestimate the importance of strategy in any area of your life, but it gets even more important when we're talking about computer vision. Considering the complexity of any computer vision system, you must carefully structure the strategy behind it to mitigate the risk of failure right from the beginning.
The three main components of a successful computer vision system are data, infrastructure and talent — and they're all equally important. You might think technology is the key to success, but underestimating people and your overall company culture would be a mistake. Due to the complexity, computer vision programs invariably require several groups of specialists to work together, and weaknesses in some of these groups — as well as unsatisfactory cooperation culture — will result in poor performance of the overall system.
Datasets For Machine Learning
Computer vision models are only as good as the data they've been trained on. According to a report from the BCG Henderson Institute, many companies do not understand the importance of data and training to make a good computer vision model successful.
You should put as much effort as possible into the creation of a proper dataset. Do not rely on a synthetic dataset, and make sure to augment your new data, simulating different conditions. When we were building our computer vision gun detection system, we needed a lot of data to train our model to be able to detect all kinds of firearms that could be seen on security cameras from various distances and in different light conditions.
It required enormous volumes of data, and we soon realized that using pictures and videos available on the internet was not enough to properly train the computer vision. We had to hire a film crew and actors to shoot on different locations and from various angles to produce more than 50% of the data we needed internally.
Diversify Your Testing As Much As Possible
Testing is the next important step, and diversification is key in testing. In our case, we needed to perform multiple tests, which included changing the distance from camera to object as well as trying different types of cameras, various angles, resolutions and so on. Today, we have more than 100,000 tests done to verify accuracy and results.
You should perform at least 1,000 tests as a base to tell how your computer vision is doing, what its weaknesses are and how to fix them.
The importance of testing takes us back to the previous point: Testing results provide you with lots of data, which is crucial to properly weigh countless correlations and connections when building a computer vision system.
Infrastructure
Computing infrastructure, both software and hardware, must be in place in order to run these models. The amount of data and the need to be able to handle it promptly means a lot of computing power is required. In 2018, research published by OpenAI (via VentureBeat) showed a rapid increase in compute power since 2012, which is driving advances in computer vision and AI.
This trend represents an increase of roughly a factor of 10 each year, and the research shows there are multiple reasons to believe that the trend could continue. No matter what kind of computer vision you're working on, it will require enormous computing power.
Talent
According to Gartner, AI will generate 2.3 million jobs by 2020, exceeding the 1.8 million it will remove. But AI is a new and still emerging technology, so finding specialists with appropriate skills and training most likely won't be easy. One way to address this problem is looking outside of the U.S. There are a number of locations globally that are popular nowadays as sources of qualified yet less expensive outsource experts for U.S.-based companies, including those specializing in AI development.
India has long been and still is a prime location for all kinds of IT outsourcing, while Eastern European countries such as Ukraine and Poland are getting increasingly more competitive in this field. Among other locations that offer qualified AI talent are China, Taiwan and the Philippines. Allocating more of your budget toward AI development training for your current employees is another option.
Most organizations working with AI do not limit themselves to one direction, taking an "all of the above" approach instead by hiring external talent, building capabilities in-house and buying or licensing capabilities from large technology firms. AI talent is a significant problem in the industry, so finding capable specialists will be a challenge at least for the decade ahead.
When talking about building and perfecting a computer vision system, the complexity of this process is something I cannot stress enough. That's why it is so important not to overlook any of the key aspects while keeping the focus on maintaining the balance, which is essentially the key to successful and high-functioning AI.
