Hi,
I’m creating an Object Detection project on Roboflow.
Feeling a little bit confused when creating a new version of the images.
On the Maximum Version Size I can see that only the train images are multiplied (by x3 by default) but the Valid and Test stays the same which leads to an unbalanced dataset.
What can I do to fix it?
Thanks!
This isn’t something that needs to be “fixed”. It’s on purpose.
You don’t want to augment your validation or testing images because their purpose is to have visibility into how your model is going to perform on images from the real world that it has never seen before.
Those images at inference time won’t be augmented, so if you augment your validation set it won’t give you an accurate measurement.
Further reading here: Train, Validation, Test Split and Why You Need It
1 Like
Thank you for explaining Brad!
It makes sense now.
3 Likes