Is there a technical term for an image classifier that classifies on a single class but is classifying on an amount like how full a glass of water is rather than different classes?

  • $\begingroup$ Do you mean class probability distribution? $\endgroup$ – Maxim Oct 7 '17 at 8:12

Training a system to predict real numbers instead of classes is called regression. If you had a training data where the inputs were images of water glasses and the targets were the amount of water in each glass, then this would be regression task from images of water glasses to the amount of water in each glass. You could call it image regression in the same way you would say image classification, but beyond that, I don't think there's a special term for regression where the inputs are images.

  • $\begingroup$ Ok, thanks. I've seen it called multi label before but that just doesn't sound right. The problem with calling it regression is it sounds like its using the old style linear method of image classification eg with bag of words, SIFT and SURF etc instead of using a convolutional network. I'm using a convolutional network with a euclidean loss instead of softmax to get a single value. $\endgroup$ – Forth Temple Oct 8 '17 at 1:29
  • $\begingroup$ @ForthTemple Regression, classification, and multi-label classification refer to supervised learning problems--predicting real numbers, classes, and binary vectors, respectively, from annotated examples. Linear models and convolutional nets are two types of models that could be used to tackle any of these problems.Yes, you're right, often you can convert a classification model to a regression model by just changing the loss and the output transfer. $\endgroup$ – Dustin Morrill Oct 9 '17 at 7:29

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.