The VISIONS system has undergone a continuous evolution over the years, driven primarily by a deeper appreciation of the nature of image understanding as both our research and that of the field have progressed. The current version of our system has produced reasonable interpretations of fairly complex outdoor scenes (e.g., house scenes and road scenes), and variations of the system have been applied to aerial images with some success. We view the system as an experimental testbed within which further evolution may be accomplished; we expect that there will be long term but slow improvement in the power and capabilities of the system as particular areas/modules/system components are improved and generalized.
Several areas of research are being vigorously pursued. One of these is the integration of motion into the interpretation process and the concurrent investigation into dynamically changing egocentric-world representations. A second component of the system that needs significant development (and there is some effort underway) is the three-dimensional representation of shape and methods for using three-dimensional information during interpretation. Although the current system has some three-dimensional knowledge of objects and can infer some three- dimensional structure, the capabilities are not adequate for a system expected to exhibit a reasonable degree of generality. There is an effort underway that involves compilation of three-dimensional information into a network of two-dimensional characteristic views associated with objects and scenes (36). A third area of research involves the inferential/evidential reasoning capabilities of the system, particularly methods for inferencing over the knowledge base using partial information from many sources. We have explored Dempster-Shafer mechanisms for evidential reasoning over the past 3-4 years (85, 113, 116, 142, 143, 146), and we now have a version of an object-hypothesis system based on feature evidence that operates in a manner similar to the rule-based object hypotheses presented in this chapter; comparisons of effectiveness are planned. However, the use of Shafer-Dempster evidential approaches to carry out propagation of partial beliefs/evidence in a semantic network has not yet developed to a point where it is clear how and whether to integrate it into our system. Mechanisms for propagating confidences/ probabilities /beliefs will continue to be a primary research issue. Fourth, the knowledge base of the system needs to be expanded to more objects and more interpretation strategies, and the system needs to be exercised on a wider variety of images. We have a series of experiments planned for the near future which will exercise both our entire system and test our research group's ability to build a knowledge base for new task domains relatively quickly. Finally, in the UMass Image Under-