Abstract:

There are many different ways to create synthesized data for use in any machine learning domain, harnessing the power of automatically generated annotations. When augmenting image datasets, generated images should be similar to the real images, aiming for the photorealism that will make sure that synthesized images contain adequate features for classifier training. Using such datasets for pose and activity detection requires sufficient pose space coverage and texture diversity. While these conditions can be met by traditional computer-generated imagery (CGI) approach, constructing 3D scenes using already made 3D objects and animations, or those specifically modeled and animated for the task in hand, is economically inefficient - especially when many variants of each scene elements are required to ensure the diversity. With these challenges in mind, we are developing and enhancing a method to solve these problems by using rule-based procedural approach paired with augmented reality.