What does it mean the term variation for an image dataset?

Question

I am working with convolutional neural networks, and I have seen that often we need to pre process the images before feeding them to the network. In particular, I have seen that often we have to do image augmentation using an image generator. Now, when looking for a clarification on why we need to do this, I came across an article which says:

Image Augmentations techniques are methods of artificially increasing the variations of images in our data-set by using horizontal/vertical flips, rotations, variations in brightness of images, horizontal/vertical shifts etc.

What I don't understand is what is the variation of a dataset.

Can somebody help me? Thanks in advance.

Answer 1

Assume, you have two different datasets of cat images:

Dataset 1

• Images were taken under full daylight.
• The cat covers always almost the whole image.
• The cat is always in the center of the image.

Dataset 2

• The images were taken under various lighting conditions: some under full daylight, some in rooms, some during the night.
• The cats were differently far away from the camera, e.g. sometimes the cat covers only 20% of the image.
• The cats are in different positions in the images, e.g. sometimes, the cat is in the center but other times, it is in the upper right corner.

In this case, we would say that Dataset 2 has a higher variation than Dataset 1.

Generally, we use data augmentation to simulate missing information in datasets. For example, you could decrease the brightness of some images in Dataset 1 to simulate how a cat looks at night. Or you could zoom out to simulate how a cat looks from farther away. Note that a perfect dataset that contained all relevant information would not require any data augmentation at all. However, gathering data is expensive and data augmentation is cheap. Therefore, we almost always have to augment.