Week 5 Post 3

Today I learned a lot about the side of AI that I had never thought of. It is the AI's weaknesses and failures. As I am progressing on my project of extracting the centroid of every single video of my hand movements and interpolating the videos to 20 frames, I started noticing the way YOLO works. Overall, the idea of YOLO seems to be great, and it is the leading object detection neural network that exists. However, there are too many issues with it, and it will take many years, in my opinion, to develop a better version of it. The first thing that I would say is that YOLO is bad at detecting objects unless I follow very strict rules for recording the videos. It runs out that in all of my videos, my hand was too close to the camera. Basically, it was too close for YOLO to recognize it, so the confidence level was pretty low. As well as this, YOLO tried to detect other objects in the frame, such as a couch or a drawing on the whiteboard, or even my face as a hand. Because of these issues, I had to rerecord all of my videos to make them according to YOLO standards. MY hand could not be too far from the camera either; otherwise, YOLO would have stopped recognizing it. So I had to rerecord every single video, and as I was doing that, I checked it with the algorithm to see if the confidence level was good enough, which is above 0.65 out of 1. It took me two days to finish that task as there were 80 videos in total that I had to rerecord many times. Finally, I got a good dataset with my hand gestures. Now YOLO recognizes my hand in all the videos. That is my issue with YOLO. The requirements are too strict (not too far, not too close, minimum of background objects, no other people) for YOLO to actually work. It sounds better in theory than it is in actuality. Now that I expressed my concern, I would like to update you on my involvement with YOLO. I am very interested in object detection and recognition and computer vision, and I am thinking now that in the future, I might be involved in a project on the development of convolutional neural networks. I now start feeling passionate about getting into academia and research fields to explore new territories of AI that no one ever worked with. I can see myself creating a new network that would do a much better job in object detection and improve the usability of such a network. I think it is very important because of the progress in technology that humans are making these days. We have cars that can drive themselves, and such creations for sure had to use very well-developed networks and be tested many times. For now, I am trying to narrow my future career path, and that seems to be something that I would be interested in. I am looking forward to keeping working on the project and seeing other advantages and disadvantages of this field.