I am trying to classify an image that can represents 3 states. Up, down or Middle.

If the image does NOT represent Up or Down, then it is by default Middle.

Should I train my CNN with a dataset including all three, or just Up and Down?

Which would this make classification more or less accurate?


You won't be able to train your model correctly if you set only two classes. Because for all the samples the model will only predict Up or Down. This will cause even the Middle classes to be predicted as Up or Down, which will reduce your model's accuracy. Hence you need to keep all three classes as labels.

  • $\begingroup$ OK thanks -also the incident of Middle is far more often. Should I allow differing numbers of images for Up, Down and Middle? Or should I get an equal number of each? $\endgroup$ – ManInMoon Feb 6 '19 at 22:15
  • $\begingroup$ There is no need to get equal number of each. It is absolutely fine to have unequal number of samples in different classes and is usually the case. However if there is a significant difference, you should try stratified sampling while splitting your data into train and test set. $\endgroup$ – bkshi Feb 7 '19 at 4:37

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.