0
$\begingroup$

We are trying to classify damaged cars versus undamaged cars using MS Azure Computer Vision service. The problem is that the model was performing better when it was trained on lesser data compared to more data. Any idea what is happening?

$\endgroup$

1 Answer 1

0
$\begingroup$

Custom Vision can behave this way if the increased data is confusing the underlying model.

It can be due to wrongly labeled data. You can do a manual quality check of the newer data being added by creating a sample set. Compare this manually evaluated sample set with labeled set you have in Custom Vision.

Another check is to remove near similar images in both classes. If features are similar with very minute differences, it can make it harder for models to identify confidently.

Mode of Training can also have an impact. Use automated vs custom training time to validate this.

Hope that helps !

$\endgroup$

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.