Stephan Mantler

Extreme model simplification for forest rendering

Models of large forest scenes are of a geometric complexity that surpasses even the capabilities of current high end graphics hardware. We propose an extreme simplification method which allows us to render such scenes in realtime. Our work is an extension of the image based-simplification method of Billboard Clouds. We automatically generate tree model representations of 15-50 textured polygons. In this paper, we focus on the algorithmic details to improve the simplification process for foliage. We use the simplified models as static levels-of-detail in the medium to far field and demonstrate how our approach yields real-time rendering of dense forest scenes for walkthroughs and flyovers.

Time-critical rendering of discrete and continuous levels of detail

We present a novel level of detail selection method for real-time rendering, that works on hierarchies of discrete and continuous representations. We integrate point rendered objects with polygonal geometry and demonstrate our approach in a terrain flyover application, where the digital elevation model is augmented with forests. The vegetation is rendered as continuous sequence of splats, which are organized in a hierarchy. Further we discuss enhancements to our basic method to improve its scalability.

Interactive smooth and curved shell mapping

Shell mapping is a technique to represent three-dimensional surface details. This is achieved by extruding the triangles of an existing mesh along their normals, and mapping a 3D function (e.g., a 3D texture) into the resulting prisms. Unfortunately, such a mapping is nonlinear. Previous approaches perform a piece-wise linear approximation by subdividing the prisms into tetrahedrons. However, such an approximation often leads to severe artifacts. In this paper we present a correct (i.e., smooth) mapping that does not rely on a decomposition into tetrahedrons. We present an efficient GPU ray casting algorithm which provides correct parallax, self-occlusion, and silhouettes, at the cost of longer rendering times. The new formulation also allows modeling shells with smooth curvatures using Coons patches within the prisms. Tangent continuity between adjacent prisms is guaranteed, while the mapping itself remains local, i.e. every curved prism content is modeled at runtime in the GPU without the need for any precomputation. This allows instantly replacing animated triangular meshes with prism-based shells.

Interactive landscape visualization using GPU ray casting

This paper demonstrates the simple yet effective usage of height fields for interactive landscape visualizations using a ray casting approach implemented in the pixel shader of modern graphics cards. The rendering performance is output sensitive, i.e., it scales with the number of pixels rather than the complexity of the landscape. Given a height field of a terrain and a topographic map or similar data as input, the vegetation cover is extracted and stored on top of the height field in a preprocess, enhancing the terrain with forest canopies or other mesostructure. In addition, enhanced illumination models like shadowing and ambient occlusion can be calculated at runtime with reasonable computational cost, which greatly enhances the scene realism. Finally, including the presented technique into existing rendering systems is relatively simple, mainly consisting of data preparation and pixel shader programming.

Evaluating realistic visualizations for safety-related in-car information systems

This paper reflects on the currently observable evolution of in-vehicle information systems towards realistic visualization. As compared to common schematic maps, hi-fidelity visualizations might support an easier recognition of the outside world and therefore better contribute to driving safety. On the other hand, too much visual detail might distract from the primary driving task. We present an experimental car-simulator study with 28 users, in which the in-car HMI was systematically manipulated with regard to representation of the outside world. The results show that perceived safety is significantly higher with 1:1 realistic views than with conventional schematic styles, despite higher visual complexity. Furthermore, we found that the more demanding the safety recommendation on the HMI, the more realistic visualization are perceived as a valuable support.

Fast approximate visible set determination for point sample clouds

We present a fast, efficient method to determine approximate visible sets for vegetation rendered as point sample clouds. A hardware accelerated preprocessing step is used to determine exact visibility for a selected set of views; at runtime the current view is rendered using an approximate visible set constructed from the three closest precalculated views. We will further demonstrate how this method leads to a significant per-frame reduction of the original data size.

Augmenting the driver's view with realtime safety-related information

In the last couple of years, in-vehicle information systems have advanced in terms of design and technical sophistication. This trend manifests itself in the current evolution of navigation devices towards advanced 3D visualizations as well as real-time telematics services. We present important constituents for the design space of realistic visualizations in the car and introduce realization potentials in advanced vehicle-to-infrastructure application scenarios. To evaluate this design space, we conducted a driving simulator study, in which the in-car HMI was systematically manipulated with regard to its representation of the outside world. The results show that in the context of safety-related applications, realistic views provide higher perceived safety than with traditional visualization styles, despite their higher visual complexity. We also found that the more complex the safety recommendation the HMI has to communicate, the more drivers perceive a realistic visualization as a valuable support. In a comparative inquiry after the experiment, we found that egocentric and birds eye perspectives are preferred to top-down perspectives for safety-related in-car safety information systems.

Interactive person path analysis in reconstructed public buildings

The behavior of a large number of persons in big buildings is pretty difficult to predict. Long time observations have to be made and many things have to be tested in the real environment, for example how to route people around locked escalators, how to place vending or ticket machines, or how to place various kinds of advertisements. A lot of experience is needed to make the right decisions and mistakes and omissions can easily happen. Models of buildings and simulations can improve the analysis of such environments where many people move about. A good visualization of human behavior in these models can be a great help in planning and improve the quality of necessary changes. It is also possible to analyze the focus of persons who are wandering around in the building, before performing actual changes and thus judge the impact of various alternatives. A major problem is that for most buildings no adequate 3D models to perform such visualizations exist. Current building models need modeling experts for manipulation, and current simulation software needs simulation experts for analyzing movement of people within a building. Ideally, simulated movement of people should be presented in a recognizable 3D model in such a way that it can be easily identified and interpreted without further explanation. It should also be possible to manipulate the building and automatically trigger the changed simulation in the background, without being an expert. This should make it possible to explore alternative planning scenarios and improve the applicability of an envisioned planning tool. This paper presents a way to create an indoor person behavior visualization shown in its application to railway stations. The models are based on a combination of image based reconstruction and existing models of public buildings. The possibility to easily change the environment on the fly creates many possibilities to analyze alternative places for advertisement, vending and ticket machines. Therefore experiments on the real environment can be avoided to save a lot of time and money and increase customer satisfaction.

Saving the z-cull optimization

Modern graphics hardware employs z-culling, or early-z culling, as a conservative z-test before the actual per-fragment z-test [Kilgariff and Fernando 2005]. Z-culling is used to avoid executing expensive fragment shaders for invisible fragments, and minimize reading, testing, and updating z-buffer entries. However, these optimizations need to be disabled for shaders that modify the z value of a fragment (depth replace). Such shaders are rapidly gaining in importance, however, especially in the context of per-pixel displacement mapping such as relief mapping [Policarpo et al. 2005]. We propose a slight modification to graphics APIs and hardware drivers that would allow to retain many of the z-cull optimizations for a large class of depth-modifying shaders, including displacement shaders. The basic idea is for shaders to specify a non-negative depth offset that is added to the z value of the fragment. While the exact z value is not known before the shader is executed, the lower bound is sufficient to perform all z-culling tests. Updates to the z-cull buffer are limited, but not worse than with texkill shaders.

Time-Critical Rendering of Huge Ecosystems Using Discrete and Continuous Levels of Detail

We present a novel level-of-detail selection method for real-time rendering, which works on hierarchies of discrete and continuous representations. We integrate smoothly represented, point-rendered objects with discrete polygonal geometry and demonstrate our approach in a terrain-flyover application. In this testing application the digital elevation model is augmented with forests in accordance with satellite data. The vegetation is rendered as a continuous sequence of splats generated from a procedural description. Further, we discuss enhancements to our basic method to improve its scalability.