To do this go to Window > Package Manager then click the + button at the top left of the Package Manager and type and click Add. Next we need to install the Perception Package which is what we will use for generating synthetic training data. You can call your project Roboflow Synthetic Data Tutorial and choose a location on your hard drive to store the project. Once it's done, open your new Unity installation and create a new project starting from the Universal Render Pipeline template. Add Visual Studio and Linux Build Support (Mono) to the installation.Īfter you've accepted the license, Unity will take a few minutes to download and install. In the installation options, be sure to select Visual Studio and Linux Build Support (Mono) which we will need later. Install Unity to your computer via Unity Hub Once you have Unity Hub installed and activated, click 'Installs' on the left and follow the process to install Unity 2019.4.18f1 (LTS) – this is the current 'Long Term Support' version of Unity and is known to work with Perception and the rest of this tutorial. Sign in to Unity Hub and activate your Unity license (this is in Preferences > License Management > Activate New License). Since Perception is in active development we'll want flexibility to upgrade or downgrade our Unity version easily as the package matures.

You can check the full criteria here then create an account here.Īfter signing up we need to install Unity Hub, a management tool that will allow us to switch between versions of Unity. You'll need a Unity account and license to get started this is free for hobbyists or those working at companies that had less than $100k revenue last year. The more closely your generated images match the scenes your model will see in the wild the better.

You can then add in real world data to improve your model over time. Synthetically generated data can be a great way to bootstrap a project and get a prototype running.

In this post, we'll show another technique: using Unity's 3D engine (and their Perception package) to generate perfectly annotated synthetic images using only software and then using that generated data to train a computer vision model.

In the past we have covered techniques such as augmentation, context synthesis, and model assisted labeling to improve the process. The promise of generating synthetic data to reduce the burden is alluring. Collecting images and annotating them with high-quality labels can be an expensive and time-consuming process.

Live visuals unity tutorial