[Task]: Is it possible for data without labels?

[korkard]

What is the motivation for this task?

If the training data has a myriad of normal data and a small amount of abnormal data, and very small labelling test data

Describe the solution you'd like

idk..

Additional context

No response

[blaz-r]

Hello. You can still train without the labels, but you won't be able to test the model.

[korkard]

Thanks for your reply.
Are you saying that I can train in a state where normal and abnormal are mixed?
I think good, bad is labels.

[blaz-r]

Sorry I understood the question incorrectly. All models implemented in Anomalib only utilize normal data for training, so you should be able to train the model using your good data, which does not require labels. The abnormal data can be used as part of the test set and shouldn't be mixed with good data.

[aradwan777]

hi to clarify, all models that exist in anomalib are unsupervised? Or in this case semi-supervised, I am not sure, but the learning is only done on the good images, that is the only data required, correct? Thank you, as I was confused if these were supervised(requiring masks for training, thus labeling for custom data) or not

[blaz-r]

Hello, that is correct. All models in currently Anomalib are unsupervised, using only good data for training. Labeled anomalous images are only used in test phase, where you require ground truth annotation to tell how good the model is performing.

[Abhijeet241093]

Is it possible, to test operation without using ground truth ?

[Sj-Yuan]

If you just need the result of anomaly detection but not localizaition, the ground truth is not required, you can just prepare some black image as the fake ground truth for test phase.

[samet-akcay]

@Abhijeet241093, you could refer to this section in the documentation
https://anomalib.readthedocs.io/en/latest/markdown/guides/how_to/data/custom_data.html#with-validation-and-testing-via-synthetic-anomalies