You're reading from Synthetic Data for Machine Learning Revolutionize your approach to machine learning with this comprehensive conceptual guide

Product type Paperback

Published in Oct 2023

Publisher Packt

ISBN-13 9781803245409

Length 208 pages

Edition 1st Edition

Concepts

Data Science

Author (1):

Abdulrahman Kerim

View More author details

Table of Contents (25) Chapters

Preface

1. Part 1:Real Data Issues, Limitations, and Challenges

2. Chapter 1: Machine Learning and the Need for Data FREE CHAPTER

3. Chapter 2: Annotating Real Data

4. Chapter 3: Privacy Issues in Real Data

5. Part 2:An Overview of Synthetic Data for Machine Learning

6. Chapter 4: An Introduction to Synthetic Data

7. Chapter 5: Synthetic Data as a Solution

8. Part 3:Synthetic Data Generation Approaches

9. Chapter 6: Leveraging Simulators and Rendering Engines to Generate Synthetic Data

10. Chapter 7: Exploring Generative Adversarial Networks

11. Chapter 8: Video Games as a Source of Synthetic Data

12. Chapter 9: Exploring Diffusion Models for Synthetic Data

13. Part 4:Case Studies and Best Practices

14. Chapter 10: Case Study 1 – Computer Vision

15. Chapter 11: Case Study 2 – Natural Language Processing

16. Chapter 12: Case Study 3 – Predictive Analytics

17. Chapter 13: Best Practices for Applying Synthetic Data

18. Part 5:Current Challenges and Future Perspectives

19. Chapter 14: Synthetic-to-Real Domain Adaptation

20. Chapter 15: Diversity Issues in Synthetic Data

21. Chapter 16: Photorealism in Computer Vision

22. Chapter 17: Conclusion

23. Index

Why subscribe?

24. Other Books You May Enjoy

Photorealism approaches

In this section, you will learn about and explore the main approaches usually deployed to generate photorealistic synthetic data. We will learn about the following:

Physically based rendering
Neural style transfer

Physically Based Rendering (PBR)

The Physically Based Rendering (PBR) approach is widely used in game engines such as Unity and Unreal to accurately simulate how materials in the 3D virtual world interact with light. In the real world, this is a complex process, thus it requires a significant understanding of optics and many simplifications to make these processes applicable to game engines. Physically based materials are essential to this approach. They resemble how similar materials in the real world interact with light. These materials usually have properties and parameters that are calculated based on real measurements from real-world materials. The properties may include absorption, scattering, and refraction coefficients and parameters. It should be noted that the main principle behind PBR is energy conservation, which means that light energy reflected and scattered by a material should not exceed the total incoming or received light energy by this material.

As expected, deploying a photorealistic rendering pipeline will help us to simulate and render more accurate and realistic light behaviors and materials. Thus, we can generate more photorealistic synthetic data.

Neural style transfer

Neural style transfer is a well-known technique that transfers an artistic style from one image to another while preserving the content of the latter. This method can be applied to synthetic datasets to improve their photorealism and thus mitigate the domain gap between synthetic and real data. For example, the Sim2Real-style transfer model can be deployed to bridge the gap between synthetic and real data for the task of pose estimation. For more information, please refer to Sim2Real Instance-Level Style Transfer for 6D Pose Estimation (https://arxiv.org/abs/2203.02069). Additionally, there are many interesting works that explore how to adapt the GTA5 synthetic dataset (https://www.v7labs.com/open-datasets/gta5), which was generated from the Grand Theft Auto V video game, to the real Cityscapes dataset (https://www.cityscapes-dataset.com).