BlendGen: A Blender Add-on for General RGB-D Databases Generation

Abstract

With the growing usage of convolutional neural networks in image classification[1], [2], detection and segmentation tasks, the need for image databases has also increased. Image database creation from real scenes is a time-consuming process. It requires capturing several thousand images of objects that are not necessarily easy to access in various contexts. In addition, the labeling process also takes a lot of time and human labeling precision is far from perfect. With time, numerous image databases have been made available. However, these databases rarely match new application requirements in terms of image resolution, labeling quality, object specificity and quantity. They can also lack information such as depth images. Creating synthetic image databases using database generation tools is a solution that makes image database creation easier and faster. In this paper, a new database generation tool named BlendGen is proposed. Database generation tools enable to create image databases with various image resolutions, perfect labeling quality, highly specific objects and high context variety. However, the existing database generation tools require users to use APIs to set up scenes and database parameters. The APIs usage can become heavy and less convenient for scene setups when the camera and numerous objects are following complex trajectories. BlendGen on the other hand is a Blender[3] add-on that enables RGB-D image database generation via a complete and intuitive graphical user interface (GUI). BlendGen GUI enables the creation of whole synthetic RGB-D databases including complex scenes, camera movements and lighting environments.