Abstract

After seeing a car multiple times, we all can recognize new ones from almost all possible angles and distances. Deep learning models are not there yet, but we can achieve human-level performance in many tasks with the help of big datasets. But what if we don’t have millions of labeled images to feed into our network?

Data augmentations can help us to expand our datasets and provide the network with diverse samples to learn from. But are the default augmentations like horizontal flipping and rotation enough? Should we perform the same augmentations to all datasets or are they heavily dataset dependent?

We will go over multiple augmentation techniques and see which of them have an actual impact on the task of image segmentation.