When doing data augmentation in computer vision problems, should you train with the original (un-augmented) data as well or just the augmented data? Are there pros and cons to the two strategies or does it not matter?

  • 1
    $\begingroup$ What is your goal in using data augmentation? $\endgroup$ – Sycorax Jan 20 '19 at 17:38
  • $\begingroup$ @Sycorax Increase the size of the training dataset. What else can I accomplish with data augmentation? $\endgroup$ – fabiomaia Jan 20 '19 at 17:53
  • 1
    $\begingroup$ What problem does increasing the size of the training set solve? One usage of data augmentation is to make an auto-encoder which is robust to a specific perturbation or addition of noise. Another might be to improve classification performance. Another might be to make the model robust to some amount of incorrect label data. $\endgroup$ – Sycorax Jan 20 '19 at 18:06
  • $\begingroup$ I want to improve generalization in an image binary classification problem. Increasing the size of the training set will make the model more robust against overfitting. Those are interesting uses too. $\endgroup$ – fabiomaia Jan 20 '19 at 20:44
  • $\begingroup$ @Sycorax Are there any books or tutorials that discuss data augmentation in detail? I was not aware of some of the use cases you mentioned, so perhaps my understanding of it is very basic $\endgroup$ – fabiomaia Jan 21 '19 at 12:18

In theory, if your augmentation is sensibly chosen and does not really change anything meaningful (e.g. rotation for satellite images), then it should not matter.

But there is certainly no harm in using the originals, too. Just make sure too don't use them so often that the model/ neutral network overfits them (perhaps use the exact originals just once, I'd speculate that you might want to use them in your very final iterations).

| cite | improve this answer | |
  • $\begingroup$ What do you mean by "just once"? Just once every epoch or just once ever? And why might it be a good idea to feed the original last? $\endgroup$ – fabiomaia Jan 20 '19 at 17:55
  • 1
    $\begingroup$ Just in one epoch. The idea with doing out last is that if there's something really subtle about the originals vs. augmented versions, then fine-tuning a pretty good network at the end based on that might be the most useful. $\endgroup$ – Björn Jan 20 '19 at 17:57
  • $\begingroup$ Are there any good papers or sources that discuss this sort of thing in more detail? $\endgroup$ – fabiomaia Jan 21 '19 at 12:57

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.