This dissertation describes the real-time rendering techniques to efficiently deal with the large scale of terrain data, which make an orientation-invariant and a view frustum scale-invariant smooth rendering possible while maintaining the image quality.
For an out-of-core terrain visualization on a resource-limited computing environment like a low-end graphics system or even a portable system, the real-time continuous rendering must be crucial to give people a higher level of visual fidelity. Such bulky and high resolution terrain data often arises from the progressive development of satellite technologies. And the need for visual realism is getting higher and higher. Terrain-based flight simulation tends to be essential for a large class of game and simulation applications including flight training simulation, mission planning and rehearsal, 3D situation assessment, and GIS(Geographic Information System). Most of the existing techniques often suffer from overloading massive terrain data on the-fly, especially texture data, by an abrupt rotational movement or a geo-reference to an arbitrary direction to the course of action around a viewpoint and a reasonably large viewing volume, which results in severe fluctuation in rendering performance and therefore can not guarantee the conservative frame rate.
Since most researches seem to have focused on developing and enhancing parts of either the data paging/caching scheme from a hard disk or a network server or the scene processing of a polygon mesh construction(CPU-bound work) and texture management (GPU-bound work), we learned that the local benefits from those methods can not be fully utilized because they are not feasibly correlated to the other modules in the overall rendering processes.
In this dissertation, we pursue the efficiency of the whole rendering processes by a consistent and coherent design of the rendering modules which appears the consistent data flow control from hard disk to the main memory and subsequently to the texture memory and the correlation of it with per-cell or per-polygon manipulation based on the common data structure. We present three core methods which can greatly impact scene management and data flow control for an efficient real-time continuous rendering.
First, we propose the adaptive texture management using multiple VFmaps (View Frustum maps) which are concentric VF footprints swept on viewpoint, to feasibly work out performance inconsistency by adaptively dealing with large scale of terrain texture. The resolution range and the size of each VFmap are determined by texture mipmap levels contributed to the viewport, viewing parameters, and texture memory capacity. The reason that we have multiple classes of VFmap is to adaptively fit the amount of texture needed to load to the texture memory, because the VFmap may cause the excessive data loading in order to cover all area around a viewpoint. We also carefully select the position of concentric center of multiple VFmaps which plays an important role in controlling the rendered image quality and the performance.
Second, We propose the continuous Levels of Detail(LoD) to manage out-of-core height fields(which, in turn, are to construct polygonal mesh) on-the-fly. Our policy is to improve the existing memory-efficient and easy-to-handle in-core algorithm ((Mat, Quadtree Matrix) to be able to apply for out-of-core data without any visual artifacts. Our method is based on top-down refinement of regular-grid, screen-space error metric considering a distance from the viewpoint and a surface roughness, neighboring nodes with no difference greater than one level. In order to treat the cracks we apply edge deletion for an infra-cell and edge stitching for inter-cells with little overhead and for the popping we adapt the geomorphing, which pictured no cracks and no waving popping within some appropriate number of triangles.
Finally, We propose the fast view frustum culling based on the quad-cell bounding box and the dynamic allocation. This approach, what we call a IP:Intersection Polygon, utilizes terrain data property which is a skewed 3D, and significantly reduces the computation time for identifying the terrain cells which are likely to contribute to the rendered scene, while maintaining the accuracy similar to the existing methods.
Besides the above three techniques, we also present some implementation issues which might contribute to enhance the rendering perfoi'Hance and the image quality. Those are a frame regulation technique to control both target number of triangles and the screen-space pixel distortions(often called tau), a multi-threading of data caching and rendering, and texture templates for seamless texturing along terrain cell boundaries. The foundation of all the approaches we have developed is the quadtree of terrain cells which are the minimal units to per-cell operation and make up the large terrain.
Applying to the terrain-based simulation consolidates the efficiency of the proposed methods for orientation-invariant and view frustum scale-invariant out-of-core visualization compared to the existing out-of-core terrain handling techniques, while maintaining a rendered image quality. As such, the time saved in the rendering process can be utilized for giving more visual fidelity.
When all the methods we have developed are integrated into the real-time terrain rendering processes, we believe this kind of approach might be the first which is totally software-wise and can be implemented and operated on an existing low-end PC.
