[Editor's Note] Artificial intelligence (AI) technology is becoming increasingly sophisticated and is on the rise in terms of commercialization. Autonomous driving has received considerable attention and is currently at the center of the trend, but various services such as facial recognition, object detection, and AI chatbots are becoming more sophisticated and are making their presence felt in the market. We met with AMC PLANET CEO Kim Hyeon-seok, an AI expert, to ask about fire detection solutions using computer vision and hear about the related market outlook and AI development.

■ What level do you think the current technology is at in the market?

As of September 2022, the recognition rate of papers published in Korea up to that point was around 97-98%.

Until then, there had been no recognition rate of over 99% like ours. It's been 4 or 5 months, so I don't know now.

■ Is a difference of 1 or 2% in recognition rate a huge deal?

That's huge. Isn't that a fire hazard? Even 0.1% is dangerous.

■ Is object detection technology the core of computer vision, such as fire detection solutions and facial recognition?

First of all, the fire detection solution is object detection. Object detection is, to be precise, looking at the entire photo and going through a labeling process in advance. Using the label tool, it is processed into a form that can be learned and recognized by AI. When this type of feature appears, you ask us to recognize this as the object we have designated, and so it continues to learn. It looks at various types of photos, but this is continuing to learn the characteristics of AI.

Neural networks include MobileNet and ResNet. Let me explain based on this. MobileNet has very few parameters. It has about 5 to 10 million parameters, and ResNet has about 30 to 80 million parameters.

Parameters for as little as 5 million and as large as 80 million are combined with each other. In this photo or this video, this part is the object we designated, and the result of sending out the result with all 80 million trained is now object detection.

Then, recognition is a bit different story.

It is true that detection must be done first. Detection must come first, and in the case of fire detection, the scope can be limited to object detection, so the network can be made small.

However, in the case of facial recognition, the object must first be detected, and then the features of the face itself must be further refined. The designated points must be further refined and recognized through a separate network.

When looking at the features comprehensively, the recognition that “Oh, this is person A” can be made only when it is probabilistically accessible.

■ What should you pay attention to when implementing object detection technology?

When doing people counting, object detection is done first. Then, the people on the screen must be followed. At this time, a square follows along.

At this time, the mistake that AI engineers who are just starting out with AI development or are intermediate level AI engineers make is that they only recognize one screen at a time. If you recognize one, the object will continue to be detected. So if you look at this continuously, it looks like it is following. Contrary to what you see, it is not actually following.

Now, surveillance, or in other words tracking, is usually called surveillance in the AI field, but it is called object tracking, and it is not tracking, it is just detecting a screen and showing it to the user.

One way to tell the difference is to assign an ID to each object and see whether it actually goes with that ID or not.

However, when I went to the AI competition in Korea in December of the year before last, most of them seemed to be doing object detection but not tracking well. Because sometimes the screen blinks with a square or there is no ID, that means tracking is not working.

The difference in application between tracking and not tracking is that when there are multiple cameras and a person is passing from one side to the other, when the person goes in this direction, he or she goes from the range of camera A to the range of camera B.

At this time, this person may have ID a on camera A, but the moment he switches, he may become person b on camera B.

■ We will share more detailed stories at the e4ds special webinar AI Tech Concert . Lastly, please say a word to E4ds readers.

When I first encountered artificial intelligence, it was a little difficult to learn. As you may know if you have studied it a little, we hardly do any math after graduating from school.

When you study computer engineering, you focus on a specific area of mathematics. That's why when you first come across artificial intelligence, you come across differentiation, integration, probability, and statistics, and I think probability and statistics are the most difficult things in mathematics.

Since there are so many of these, it is very difficult to approach them and it is also difficult to understand the symbols. You may feel that the barrier to entry is quite high.

But I think that may not necessarily be the case. When software develops, an OS comes out, a framework comes out, and then apps come out, and app developers start to appear. In the past, people who developed systems like the OS also developed apps.

At that time, it was very difficult to come up with even one software or application. So at that time, people like us were very popular. The price was high, but not anymore. People who develop apps appeared separately.

Likewise, AI is now evolving into a framework form. Then, if the framework for AI is perfect, if a platform like a system framework is fully established like now, anyone with creative ideas about AI can create their own AI.

I hope you don't think too hard about it. Rather than thinking hard about it, I think if you approach AI by thinking about how it can make our lives richer and more fun, you will be able to develop it more enjoyably.

Thank you.