I’m working on a new version of my object detection. I thought that I could improve on my previous model’s performance, which was achieving 60 - 90% accuracy, very inconsistently.
To update the model, I decided to spend the extra effort on polygons. After exporting my annotation data set to Create ML and training a new model, I was very sad to see accuracy was hovering closer to 10 to 20% now.
I think I’m missing something fundamental here and I’m not seeing this in the knowledge base … so, here I am.
I’ve attached a screenshot from the mobile app using the Create ML model (grid size is 25 x 25, loss was ~1.6) … I was expecting this to be performing much better / more reliably than my previous model because I added a nice black box around the bubbles, which I thought the model would be able to detect more reliably. Then, adding polygon detection, I thought would improve it further.
Can anyone tell me where I’m going wrong? I’m thinking about going back through the data set and removing anything that has significant angle or barrel distortion and going back to a simple rectangle. Thoughts?