Function-Based Generic Recognition for Multiple Object Categories

Abstract

Abstract In function-based object recognition an object category is represented by knowledge about object function. Function-based approaches are important because they provide a principled means of constructing generic recognition systems. Our work concentrates specifically on the relation between shape and function of rigid 3D objects. Recognition of an observed object shape is performed by reasoning about the function that the shape might serve. Recent research has demonstrated the feasibility of function-based shape recognition. However, previous efforts have dealt with only a single basic level object category. A number of important issues arise in extending a function-based approach to handle multiple basic level categories. One issue is whether the knowledge about function can be organized into general primitive chunks that are reusable across different categories. Another issue is how to efficiently index the knowledge base so as to avoid exhaustive testing of an object shape against each known category. In order to explore these issues, we have implemented a second-generation function-based recognition system that handles a collection of basic level object categories within the superordinate category furniture. The recognition capabilities and indexing performance of this system have been evaluated on a database of over 250 shapes. We also show recognition results from some 3D shape descriptions acquired from laser range finder data. To our knowledge, this is the first (only) work in function-based recognition to address the recognition of multiple object categories.