When I actually ran it, there were three idols, and I followed the code provided by the three idols in boxes. So, I confirmed that I can recognize the members I want, and one of my recent hobbies is watching idol stages at a live indie event called Under. I have been watching idol stages there. So, I created a service related to these idols, and I thought it would be nice to add some functionality later. I thought it would be nice if I could extract highlight videos by member from the full videos of the idols that are uploaded like this, without having to create separate cam videos. I used Cloud, and first, to use it, I had to install OpenCV version 5. I made it so that I could install OpenCV, a computer vision library, in Python, so that I can watch videos.

Now I know, and Yellow detects what's in the video, tracks it based on recognizing the office, and understands that it is used to create highlights or avoid it. So, I installed it thinking that it is used for creating highlights or avoiding what I want. Therefore, in the prompt cloud, it says that it is the world's best Python developer, and now I have minimized the laptop resources, having only Python and PyCharm. I also have idol videos, so I need to produce edited videos extracting each idol member. I did it this way, so ultimately I followed Cloud's instructions and now I have put it in. Actually, I don't fully understand it, but when I tried it first, there was one issue. First, I will show you the results. So, I just copied and pasted the value, and if you look, when I actually ran it, there are three idols, and I followed the code provided by Cloud, boxing these three idols and keep following them.

So, first of all, recognizing the members I want was confirmed, but initially, it didn't work. The problem was that there was an SSL issue. Claude didn't even know what the SSL issue was at first, and later found out that it was a security authentication issue. Claude informed me about this, and there were three methods. One is to temporarily bypass and do SSL, and the other is to directly update or install something in the terminal. Among these three, the last two methods didn't work for updating, only bypassing worked. So, up to this point, it was a bit disappointing because each member's video should appear, and if it appears exactly as it is cut here, it doesn't look good. So, zooming in and out for each member to show a real video.

I wanted to download it as if it were taken with a camera and left a question while doing this. Now, there was a part about errors, problems with actual video tracking, and consuming too many resources, wondering if there is a way to reduce it. I watched this for a few seconds, and Junggi also gave ideas, and Kuri left a comment. Later, it was mentioned that it would be nice to use it as a YouTube content highlight service. Actually, I was planning to end it here, but as I continued, it became a bit fun, and the questions were recharged. Today, I played with it a bit more and uploaded another follow-up. So, while proceeding, I came to know about something called Colab.

Before I explain, the program I wanted to create was like this: when uploading an idol pool video to Google Drive, highlight videos are extracted for each member. If there are three members, three highlight videos should be extracted. These extracted videos needed to be saved in a specific folder on Google Drive. For example, if there is a highlights folder, three member-specific highlight videos should be saved there. Previously, I was running it on PyCharm, but it was a bit slow and my laptop struggled, as you may have noticed earlier. So, the insight I gained was that Jin-gi also provided an ID, and it seems like using something called Google Lab allows for a development environment that can be used on the web by Google instead of PyCharm.

It became more convenient to link, and similarly, if you write a code here, you can borrow a GPU provided here for free up to a certain extent, so there seems to be no reason not to use it, so I tried it again with this. So, you, as a partner, also showed the results while giving feedback, and explained it a bit, like how to integrate with Google Drive, load the path, detect people, and show it again. But one thing is, when using Colab and running it in PyCharm, one difference is that in PyCharm, there was something that showed in real-time from the openCV installed earlier, but here, it shows the images frame by frame in a rigid manner. I felt like it didn't show in a continuous manner, and also, another thing that was helpful.

I was tracking three people at once, but I made some modifications because I couldn't understand it with my code, so I decided to track only one person. The person on the right is the future of the Proxima Club, whom I like. Anyway, so I tried to progress and was told to try it with Colab, so I rewrote the program and tried it again. I mentioned that I need to develop using Frappe and felt a bit disappointed because I couldn't zoom in and out like I wanted in the video I showed before. So, I felt that the edited video might be a bit disappointing, and I got a bit greedy, so I immediately integrated it with Google Drive this time, and I had to include the path, saying that I should do the Hyreight video here. So, the result is the same. There is a limit to this. Even if I received the limit, I couldn't do it. I ran this for over an hour, but.

It didn't apply well, so I was a little disappointed. The fact is, I don't know the code well, so I'll move on quickly. What I gained insight from here was when I tried it again with PyCharm on my local machine. I tried it again with PyCharm after Colab, you know. The reason is that there is a limit to it, so I couldn't just keep running it. If it breaks and I can't continue, I had to run it again with PyCharm. One thing I felt was that there was a problem with the continuity of idol object recognition, so I requested the application of a Kalman filter. Luckily, I happened to learn about the Kalman filter a while ago when I touched a bit on the positioning side. It was about adjusting the 5th order from the coordinates. So, how did I perceive this problem? I came back to PyCharm again. When I ran it here, I created highlight videos for each member.

I requested a change. Oh, it's me. So, originally, three should appear when each member appears in the video. I thought this was normal, but it's not worldwide. It seems like it recognizes based on the number of people. Just differentiating this, without continuity, recognizing each object as a person in each frame, I thought there might be a problem. So, I requested a Kalman filter to correct it. I used up my Claude quota, so I couldn't proceed further. It's kind of interesting anyway. That's it for now. Also, the service I mentioned earlier that I'm trying to apply was originally a service developed with a bit of MBP, but I made it public with a no-code tool. Now, briefly, a community.

I created a way to cook, then made a room where we can chat. Later on, if I add one more menu here and upload a video, we can create highlight videos just with full idol videos, or we can directly provide content like this. I thought about trying to utilize it in this way. That's all for the case presentation. Thank you. Minji, you prepared many interesting cases, so I really appreciate it. If you could provide the filter you mentioned earlier, I actually found Minji's work quite interesting. I have two things to mention. When running the code, even though it's free, Google may not allow it to run continuously as a public service. So, if you run it continuously for about one or two hours, it will cut off in the middle of the runtime. When it cuts off, you have to start over. So, before it cuts off, you should save something.

I have to do that, so there are some limitations. However, if you subscribe, you can expand. You can increase the duration by paying money, and you can use the GPU. Earlier, when running Kalma Peter in PyCharm, now it recognizes people and seems to want to track three people individually. But, it cuts people off well, but it seems to have difficulty distinguishing when it goes up to a thousand people. So, as I mentioned, the function of tracking a specific person and classifying if it is the same person, there are some tools to overcome that hurdle technically. I know a few of them, so by utilizing them, if we go through the codes one by one, interesting results might come out. It was a very good plan.

You are amazing. You originally did this in the field of soccer. Soccer is played by 11 people, and each of those 11 people now has highlights, and that's what you originally did. Now, you have applied it to underground idols, so wherever you do something, it could become a great idea in the future if combined with something else. So, I really enjoyed watching it. Yes, Dr. Minji, who presented, please continue to do well.

직캠 영상 없어도 AI로 내 최애 아이돌 직캠 영상 만들 수 있을까?

ChatGPT와 같은 AI를 사용하는 사례를 함께 공유하는 GPTers 커뮤니티에서 운영하는 유튜브 채널입니다. 


