0
$\begingroup$

I'm training a segmentation model (computer-vision). Thus, my dataset contains images and masks (binary segmentation of objects).

I'm augmenting the training dataset (applying random crop, rotation or shift etc.) to get a larger dataset. I don't apply augmentation on test and validation dataset.

Should I use augmentation on the validation dataset or the test dataset too ?

$\endgroup$
1
  • $\begingroup$ My view is that the test data should be as ralistic as possible, i.e. should be as much as the data you will predict in the end. $\endgroup$ – Peter Nov 12 '20 at 12:21
1
$\begingroup$

Your test and validation dataset should reflect the type of data you would expect when you deploy your model in the actual setting. So usually you do not apply augmentation to the validation and test dataset, since in the real setting you will not receive some strange augmented images.

Another way to think of it is if you apply augmentation to your validation dataset then you will actually measure how your model will perform on augmented data and not on 'real' data. You will use this to inform the best settings for your model so you will end up with a model tuned to perform well on augmented data.

$\endgroup$
1
  • $\begingroup$ Thank you for your answer! $\endgroup$ – Nicolas Nov 13 '20 at 14:54

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.