Acquisition of a dense 3D model database for robotic vision

Service robots in real world environments need to have computer vision capability for detecting a large class of objects. We discuss how freely available 3D model databases can be used to enable robots to know the appearance of a wide variety of objects in human environments with special application to our assistive kitchen. However, the open and free nature of such databases pose problems for example the presence of incorrectly annotated 3D models, or objects for which very few models exist online. We have previously proposed techniques to automatically select the useful models from the search result, and utilizing such models to perform simple manipulation tasks. Here, we build upon that work, to describe a technique based on morphing to form new 3D models if we only have a few models corresponding to a label. However, morphing in computer graphics requires a human operator and is computationally burdensome, due to which we present our own automatic morphing technique. We also present a simple technique to speed the matching process of 3D models against real scenes using visibility culling. This technique can potentially speed-up the matching process by 2-3 times while using less memory, if we have some prior information model and world pose.