Image Augmentation: A Practical Guide to Prevent Overfitting in Computer Vision

What you'll learn

For machine learning models, we all know more data is better. For convolutional neural networks, image augmentation provides a straightforward way to expand your training dataset, by applying simple transformations to the images you already have. In this talk I'll start by demoing the basic standard transforms using the albumentations python package, and work up to some more advanced strategies like CutMix and mixup. I will also discuss some findings of the RxRx1 kaggle competition that Recursion ran last summer, and how this demonstrated the power of these techniques when applied to our cellular image data.

Table of contents