I am playing with the transforms from Torchvision.
There are plenty of different kinds of these like:
ResizeRandomCropColorJitterBlurring- ...
These are some cases of Resize for a given image:
ColorJitter
RandomAffine
The main purpose of the augmentation procedure is to prevent overfitting, extend the training dataset in a certain way.
Some transformed images still look like the image from the original dataset, since a small change in contrast or brightness still makes the image look real.
For the other, there can be a significant change in color, or the original image occupies only a small fraction of the resulting image.
In all cases, one can still classify this object as a dog. However, many of these would not lie on the manifold of real images, and the classification for these may not make sense.
Are there some papers or research discussing the issue of the augmentation choice, and the correspondence of these to the directions along the manifold of real images?
