1
$\begingroup$

I'm carrying out training/testing of a convolutional neural network for facial expression recognition with various datasets - all labelled by 7 emotion classes.

For other datasets, there are a large number of mostly unique subjects so I randomly split. In this case, however, there are only 6 subjects but a large number of images for each subject in each class. Randomly splitting seems ineffective because of the similarity in images - think of how an emotion changes per frame.

Is the best method to separate an entire subject for testing? Or something else?

I did run the network with random splitting and achieved 100% validation accuracy so I believe that is unlikely to be the best method. Thanks for your time.

$\endgroup$
1
$\begingroup$

I think you're hitting on the fact that by training and testing on the same subjects your model is not going to be able to generalize to new subjects very well. If you're only interested predicting emotions for these subjects, you are taking the right approach. However, if you want to generalize your model to new, unseen subjects, you should split your training and testing sets such that subjects in the training set are not in the test set, and vice versa to get a more accurate test score. Most likely it will not perform as well and you should consider collecting more data on new subjects if possible.

Here is a similar question on Stack Exchange.

$\endgroup$

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.