Practical Simulations for Machine Learning

Chapter 3. Creating Your First Synthesized Data

This chapter introduces synthesis, the second pillar of this book, as discussed in Chapter 1. Here the focus is on the tools and process you’ll be using to synthesize data for machine learning, and how it ties into the work you’ve done so far for simulation as well as how it’s quite different.

By the end of this chapter, you’ll be generating the world’s most disappointing synthesized data! But you’ll be prepared to make far more interesting data in future chapters. We promise. Stick with us.

As we mentioned in “Unity”, the primary tool we’ll be using for our initial foray into synthesis is a Unity package called Perception.

Note

We’re not going to be doing quite as much synthesis in this book as we do simulation. This is simply because there’s not as much to learn: simulation is a tremendously wide field with many different approaches that you can take, while synthesis with Unity mostly boils down to the different kinds of randomizations that you want to perform in order to generate the data you need. We’ll teach you everything you need to know, but there will be fewer activities.

Unity Perception

Unity’s Perception package turns the Unity game engine into a tool for generating synthetic datasets—primarily images—for use in ML workflows that are primarily outside of Unity.

The Perception Framework provides an array of useful tools, ranging from dataset capture, to object labeling, image capture, and beyond. You can create straightforward ...