Hello everyone I'm building a license plate detection model in Tensorflow. I built a function that chooses a license plate at random from a collection of ~5000 plates and puts it in a random place in on a random background and saves the coordinates. At first I thought to generate about 40K images this way and train the network on with the generated data. But wouldn't it be a good idea to just continiously keep generating new data to feed to the network and basically eliminate any chance of it getting overfitted?

  • $\begingroup$ Doing data augmentation on the fly is generally good in my experience. I am worried about you just putting a license plate on any old background when in reality license plates appear usually on cars. You may find performance is better training in real data. $\endgroup$
    – kbrose
    Aug 21, 2018 at 3:03
  • $\begingroup$ @kbrose I started off training it on real data. But many times the network learned to identify the fron side of a car and not the license plate itself for somereason. So that's actually one of the reasons why I'm trying something different $\endgroup$
    – ronsap123
    Aug 21, 2018 at 11:27