Custom segmentation using YOLO v8: Ghostbuster
: Ghostbuster
A couple of months ago, when I was learning about computer vision, my mom walked into my room and asked me what I was doing. I replied, “I’m working on object detection, which identifies objects in an image or video using a Reset model.” I turned on my webcam and showed her different objects like a mouse, keyboard, person, ball, etc.
My mom was impressed and asked, “So, this can identify anything?” I replied confidently, “Yes, anything, but the model needs to be trained on a specific dataset.” She then asked, “What about planes and fish?” I said, “Yes, if the model is trained on a dataset that includes planes and fish.”
My mom then asked, “What about water, ghosts, and your girlfriend?” I was caught off guard and stunned, unable to say anything.
Time passed, and when YOLO-V8 came out, I had a weird idea: what if I could use it to identify ghosts in images? I downloaded some images using a Chrome extension, labeled them as ghosts using Roboflow, and trained a YOLO segmentation model for 100 epochs in Colab. The model achieved a Map50 score of 0.87, which was excellent for such a small and bizarre dataset.
I downloaded some videos to test the model, and the results were weird and scary, but I wanted to try it. If anyone wants to try a custom project like this, I suggest following the step-by-step procedure I used for my face emotion recognition project. You’ll be good to go!
If you’re interested in following my work or connecting with me, you can find me on LinkedIn or Medium. My next project might be girlfriend detection, but only if I can gather enough data! (Just kidding, of course.)
That’s all for today see you soon
