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?