How to Generate Synthetic Data and Train a Model with it?
In this presentation, we will discuss how to run experiments using the Unity platform to generate synthetic data to train models for tasks such as object detection and recognition or pose estimation.
Please, check the slides here.
What is the cost of building a synthetic dataset? What tools can be used for this process? How complex is it to integrate synthetic data into model training?
These are just some of the questions you might be asking yourself if you want to know more about generating synthetic data for machine learning. When it comes to visual computing, media platforms like Unreal and Unity engines and more recently NVidia Omniverse have been used to support application research and development. In this type of environment, we can automate the construction of scenes with a high degree of control over properties such as the appearance and the pose of objects of interest, camera parameters or variation of background objects. We can also control the type of data generated – RGB images, depth maps, segmentation masks, point clouds, etc – and the amount of data generated, allowing a balance of examples between categories.
In this presentation, we’ll see how you can experiment with the Unity platform to generate synthetic data that can be used to train models for tasks like object detection and recognition or pose estimation. From the discussion started in first seminar, let’s check how we can get automatically labeled images and also perform domain randomization operations.