An approach to three-dimensional scene databases

Current image database research is concerned for the most part with the encoding and processing of two-dimensional images. However, the most successful approach to computer vision is based on 3-dimensional information, organized as either stacks of 2-D images (e.g., the intrinsic images of Barrow and Tennenbaum, or the 2 1/2 dimensional sketch of Marr) or as actual 3-dimensional data (e.g., the Multisensor Kernel System of Henderson). Efficient techniques for two-dimensional image processing have been well-developed over the last few decades and special purpose architectures are now available. However, the study of the organization, processing and analysis of three-dimensional scene data is only just beginning. We describe one approach to the representation of three-dimensional image data and evaluate several computer vision algorithms performed on the data. Finally, we describe how three-dimensional feature operations can be performed as relational database operations.