Minimal number of training images


Is there a rule of thumb on how many images you need to train a good model? I’m having trouble building a classifier to distinguish between my two daughters. I used 30 images of each daughter for the training set, and 10 images for validation. I tried using different learning rates and epochs, but I never archieve accuracy higher then 85%.

Is there anyone who can shed some light please? Is this caused by the number of images?

Thank you in advance!

1 Like

With such a small quantity of images, there is a high chance of overfitting. Inspect your model results and try to understand where the model is doing wrong. If you had not use data augmentation then try using it. You can also click, few more photos :smiley: .

1 Like

Thank you, I will try to use more images. However, I remember from the first lesson Jeremy gave an example where a model was trained using far less images. Data augmentation seems like another possibility indeed.

have you look here?

I’ve also written a blog post for tips on optimizing your model, hope it helps!