Generating 2.5D Photorealistic Synthetic Datasets for Training Machine Vision Algorithms

摘要

The continued success of deep convolution neural networks (CNN) in computer vision can be directly linked to vast amounts of data and tremendous processing resources for training such non-linear models. However, depending on the task, the available amount of data varies significantly. Particularly robotic systems usually rely on small amounts of data, as producing and annotating them is extremely robot and task specific (e.g. grasping) and therefore prohibitive. Recently, in order to address the aforementioned problem of small datasets in robotic vision, a common practice is to reuse features that are already learned by a CNN within a large-scale task and apply them to different small scale ones. This transfer of learning shows some promising results as an alternative, but nevertheless it can not be compared with the performance of a CNN that is specifically trained from the beginning for that specific task. Thus, many researchers turned to synthetic datasets for training, since they can be produced easily and cost effectively. The main issue of such datasets that already exist, is the lack of photorealism both in terms of background and lighting. Herein, we are proposing a framework for the generation of completely synthetic datasets that includes all types of data that state-of-the-art algorithms in object recognition, and tracking need for their training. Thus, we can improve robotic perception without deploying the robot in time-consuming real-world scenarios.