Object recognition using large modelbases
Kim L. Boyer, Kuntal Sengupta · 1996
We present a model library (or modelbase) organization strategy for fast object recognition in computer vision. Scene objects and model objects are stored as graph theoretic abstractions because object parts and their spatial relations can be efficiently embedded in a graph. Object recognition essentially involves the search for the object model that best describes the scene object. To avoid the brute force linear search over the model library during recognition, we structure the model library in the following manner. We first partition the library into submodelbases based on the spectral features of the model objects. The spectral features essentially capture the overall structural information in an object. Each partition is next organized hierarchically in the form of a tree using an information theoretic inter-model dissimilarity measure. During recognition, the correct partition is first identified. This is done by computing the spectral features from the hypothesized object, and using a perturbation model of these features. This is followed by one matching process between the scene object and the partition representative. Here, we introduce a fast matching technique using the geometric hashing approach, which is used in conjunction with the organized hierarchy. Next, we perform an inexpensive tree search and identify the correct model. We also present a method of incrementally updating the library with a new object model, a feature essential for any practical model based object recognition system. We present extensive experimental results on a CAD modelbase of more than 100 objects and numerous range images to verify the efficiency of our organization strategy during recognition.