0

I know this is a very basic question. I want to train my Custom Vision Model to detect when someone is using the cell phone. So, my question is: must I train the model with cell phones images (only the device) or with images of people using the cellphone?

I am trying to train it with images of people using the cell phone but I am getting a very low performance (Recall of 34%).

Roger
  • 83
  • 1
  • 2
  • 12
  • You have to train it with similar images to the ones you want to detect stuff in. – Cris Luengo Jul 09 '19 at 00:29
  • That's what I am doing, but then... why so low recall? – Roger Jul 09 '19 at 00:31
  • 2
    You probably don’t have enough images to train with. This is not a trivial problem, the difference between a person holding a phone to his ear and a person scratching his ear is rather subtle! Models like these are typically trained with many thousands of images. And then you need just the right model, some NN architectures work better for some problems, some work better for other problems. It’s all educated guesses and lots of trial and error, it seems. – Cris Luengo Jul 09 '19 at 00:35
  • Thanks Chris! I have 250 images, I will throw some more in there. – Roger Jul 09 '19 at 00:42

0 Answers0