Jinseok Seo, Gerard Jounghyun Kim, Kyo Chul Kang
Department of Computer Science and Engineering
Pohang University of Science and Technology (POSTECH)
Korea
Despite the rapidly improving processing power and graphics capabilities of today's computer per unit cost, realizing realistic, yet fast and interactive VR system will continue to demand engineering of "resource vs. effects" trade-offs. One of the most popular method of such performance optimization is the use of levels of detail (LOD). The use of LOD in VR, in most cases, actually refer to the use of different levels of "geometric" detail (an idea of improving display performance by rendering objects with less detail when they are visually less important in a frame), while exploitation of the similar idea for behavior or simulation has been suggested as well [Carlson 97, Kim 98]. Techniques of generating different levels of geometric LOD are mostly "bottom-up" approaches [Hoppe 96], that is, the most detailed mesh model are "simplified" to several levels. No particular method of producing behavior levels of detail seem to exist. On the other hand, the idea of "top-down" design of software or geometric objects have been advocated for many years. In [Kim 98], we have introduced a CASE tool called ASADAL/PROTO that supports the application of software engineering principles (such as the concept of hierarchical and incremental modeling, and simultaneous consideration of form, function and behavior) for modeling VR objects [Kim 98]. Each modeling stage driven by such a philosophy produces step-by-step form and its fitting function and behavior. We can make good use of these by-products as LOD's for adaptive display and simulation by additionally specifying conditions for LOD switching. It means that the development proceeds from abstract, but affordable models to detailed, but bulky models, and that abstract models need not be extracted from detailed ones. Our approach is demonstrated in a town navigation system in which a good many automobiles are cruising. At the beginning, automobiles are rendered as just one box and moving along a simple path in a constant velocity. Then, the geometric models and behavior of automobiles are refined in a stepwise manner, using ASADAL/PROTO, until an acceptably detailed model (geometric or behavioral) is obtained. At each stage, the intermediate models are simulated and provide valuable performance data such as approximate latency, frame rate, and the bottleneck processes. Before proceeding to the next modeling stage, we can decide whether certain part of the system could be refined or not on the basis of the simulation results. Hopefully, the resulting virtual world is a conglomeration of geometric objects and their behaviors that are appropriately refined at different levels, and a switching policies that are tailored for the user needs. We believe our approach combines the idea of structural and hierarchical refinement of virtual objects and the use of LOD in a very natural and intuitive manner.