2
$\begingroup$

I am currently working on a machine learning project where I use the YOLO Algorithm to detect an object inside of a picture or video. The problem I face is that the specific image set (side-scan sonar) that I am working with is mostly classified, thus there is not a wide range of images available to the public to be used for training. Would I be able to implement a GAN to produce a larger data-set of side-scan sonar imagines to be used for training an image detection Algorithm? I understand that there may be image distortion from using a GAN, but for the purpose of having a larger training set would this be possible?

$\endgroup$
3
  • 1
    $\begingroup$ Are you looking for additional labeled examples (specifically getting additional "positives"), or are you more looking to better suppress the background (reduce false-positives, potentially at the cost of increased false negatives)? $\endgroup$
    – R.M.
    Nov 19 '18 at 21:05
  • $\begingroup$ @R.M. I am looking to get additional positives i currently have roughly 200 images how ever would like to make that close to 2000 $\endgroup$
    – James
    Nov 20 '18 at 14:45
  • 1
    $\begingroup$ Then I agree with the answer you got from Unknown Coder -- It's exceedingly unlikely you'll be able to train a GAN to make decent mock positives from only 200 examples. (Personally, I'd be wary of using GAN output as mock positives even with a large amount of training data.) $\endgroup$
    – R.M.
    Nov 20 '18 at 15:27
2
$\begingroup$

In short: no

If you don't have enough images now then you almost certainly dont have enough to successfully train a GAN (it takes more than you think). If you can't train a GAN to reproduce good images, then you don't have images that are going to give you high accuracy for your YOLO effort. The most likely scenario is that you would end up with an algorithm that is really, really good at detecting GAN images that looking nothing like your side scan sonar validation data :)

$\endgroup$

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.