Towards Visual Training Set Generation Framework

Jan Hula, Irina Perfilieva, Ali Ahsan Muhummad Muzaheed

Abstract: Performance of trained computer vision algorithms is largely dependent on amounts of data, on which it is trained. Creating large labeled datasets is very expensive, and therefore many researchers use synthetically generated images with automatic annotations. To this purpose we have created a general framework, which allows researchers to generate practically infinite amount of images from a set of 3D models, textures and material settings. We leverage Voxel Cone Tracing technology implemented by NVIDIA to render photorealistic images in realtime without any kind of precomputation. We have build this framework with two use cases in mind: (i) for real world applications, where a database with synthetically generated images could compensate for small or non existent datasets, and (ii) for empirical testing of theoretical ideas by creating training sets with known inner structure.