Acquisition of view-based 3D object models using supervised, unstructured data

Existing techniques for view-based 3D object recognition using computer vision rely on training the system on a particular object before it is introduced into an environment. This training often consists of taking over 100 images at predetermined points around the viewing sphere in an attempt to account for most angles for viewing the object. However, in many circumstances, the environment is well known and we only expect to see a small subset of all possible appearances. In this paper, we test the idea that under these conditions, it is possible to train an object recognition system on-the-fly using images of an object as it appears in its environment, with supervision from the user. Furthermore, because some views of an object are much more likely than others, the number of training images required can be greatly reduced.