0
$\begingroup$

I have a bunch of 2D grayscale images for which I want to train a (multi-label) segmentation model. What is the simplest way to interactively train such a model?

I.e., I want to:

  1. Label a (very) few images manually
  2. Train / finetune a model on those segmentation masks
  3. Iteratively refine / retrain that model by going through more images + correcting the predicted masks

I know of methods such as DeepEdit, DeepGrow etc. and I technically understand how they work.

However, I am looking for the simplest (most time-efficient) way and tool to do this in practice without having to code things myself.

Active learning would be a plus but isn't required.

Improve this question
$\endgroup$

1 Answer 1

Reset to default
0
$\begingroup$

Here are three similar software for interactively training segmentation models:

Improve this answer
$\endgroup$
2
  • $\begingroup$ Thanks! The first two kind of look like what I was searching for, but from the tutorial pages I'm not yet sure they allow for refining the classifier over multiple different images? I.e., it looks as if the classifier is only trained based on scribbles on a single image? $\endgroup$
    – Eike P.
    Commented Aug 29 at 14:28
  • $\begingroup$ Using ilastik, for example, you can open many images at once, say, all the images in a directory (see here ilastik.org/documentation/basics/dataselection). It may probably interpret them as frames in a movie, but it will not be a problem if the interpretation is wrong. What is important is that they will all be open at once, and adding annotations on any image will update the model, and the live/interactive prediction for all the open images.f $\endgroup$
    – bogovicj
    Commented Aug 29 at 23:16

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.