Augmented images split location and behaviour

  1. When augmenting a dataset subdivided in train, validation e test set, are the augmented images guaranteed to remain in the same set? E.g. the train set augmented images remain in the train set, and so on for the ones of the validation and test set.

  2. Does all the augmentation methods retains the annotations (instance segmentation masks in my case) in the augmented images? I’ve tried a few and it seem that the answer is yes, but I don’t know if there are exceptions.


Hi @Davide_Visentin

I think @Lenny gave a good answer here, but let me clarify and add on:

  1. Yes, if you put an image into a specific split (train, valid, test) your augmented split will stay in that split when you generate and export a version. Do note that we only augment the train split.
  2. Yes, all the augmentations retain the annotations. I’m not entirely sure what kind of exceptions you have in mind, but one might be how we do our bounding box-level augmentations. Those alters the annotations and the image.

This topic was automatically closed 21 days after the last reply. New replies are no longer allowed.