Synthetic-to-real domain adaptation – issues and challenges
In this section, you will explore the main issues and challenges of synthetic-to-real domain adaptation. This will help you to understand the limitations of this approach. Additionally, it will give you a better insight into how to overcome these issues in your own problem. Therefore, we will focus on the following issues:
- Unseen domain
- Limited real data
- Computational complexity
- Synthetic data limitations
- Multimodal data complexity
Let’s discuss them in detail in the following subsections.
Unseen domain
In many cases, the aim is to make sure that your ML model will generalize well to new domains. If we know the domain, domain adaptation methods may work. However, sometimes it is not possible to predict the properties of this new domain. For example, assume you have a computer vision model that works well in Europe but you also want this algorithm to work well in China, Africa...
