Konversatorium im SS 2013

In this talk we will give an overview of the recently started FWF Project 'Modern Functional Analysis in Computer Graphics' (MOFA). The main goal of this project is to explore the potential of recently developed functional spaces for their usage in the field of computer graphics and, more specifically, in the field of rendering. We will consider generalizations of wavelets, so-called anisotropic wavelets, such as curvelets and contourlets both on the sphere and hemisphere. As these kinds of functions more readily adapt to sharp features in light transport representations (i.e. radiance and visibility signals), they should allow for increased performance in global illumination computations both in terms of memory and time consumption.

Our new rendering group member, Károly Zsolnai, will also outline this ideas on how to generalize his recent results on automatizing the parameter selection of a Metropolis Light Transport renderer to the more general case of lighting setups for ray tracing systems.

We propose a new method for the fast computation of light maps using a many-light global-illumination solution. A complete scene can be light mapped on the order of seconds to minutes, allowing fast and consistent previews for editing or even generation at loading time. In our method, virtual point lights are clustered into a set of virtual polygon lights, which represent a compact description of the illumination in the scene. The actual light-map generation is performed directly on the GPU. Our approach degrades gracefully, avoiding objectionable artifacts even for very short computation times.

We propose a novel way to efficiently calculate soft shadows in real-time applications by overcoming the high computational effort involved with the complex corresponding visibility estimation each frame: We exploit the temporal coherence prevalent in typical scene movement, making the estimation of a new shadow value only necessary whenever regions are newly disoccluded due to camera adjustment, or the shadow situation changes due to object movement. By extending the typical shadow mapping algorithm by an additional light-weight buffer for the tracking of dynamic scene objects, we can robustly and efficiently detect all screen space fragments that need to be updated, including not only the moving objects themselves, but also the soft shadows they cast.

By applying this strategy to the popular Percentage Closer Soft Shadow algorithm (PCSS), we double rendering performance in scenes with both static and dynamic objects - as prevalent in various 3D game levels - while maintaining the visual quality of the original approach.

In this paper, we present a novel rendering method which integrates reflective or refractive objects into a differential instant radiosity (DIR) framework usable for mixed-reality (MR) applications. This kind of objects are very special from the light interaction point of view, as they reflect and refract incident rays. Therefore they may cause high-frequency lighting effects known as caustics. Using instant-radiosity (IR) methods to approximate these high-frequency lighting effects would require a large amount of virtual point lights (VPLs) and is therefore not desirable due to real-time constraints. Instead, our approach combines differential instant radiosity with three other methods. One method handles more accurate reflections compared to simple cubemaps by using impostors. Another method is able to calculate two refractions in real-time, and the third method uses small quads to create caustic effects. Our proposed method replaces parts in light paths that belong to reflective or refractive objects using these three methods and thus tightly integrates into DIR. In contrast to previous methods which introduce reflective or refractive objects into MR scenarios, our method produces caustics that also emit additional indirect light. The method runs at real-time frame rates, and the results show that reflective and refractive objects with caustics improve the overall impression for MR scenarios.

Remote sensing of land cover has, not at least because of the rapidly growing world population, gained scientific and economic importance. Several international projects have aimed to determine and map out the global land cover. Unfortunately the results of remote sensing projects such as GLC-2000, MODIS or GlobCover are often ambiguous or have significant differences. The project Geo-Wiki.org examines the results of that remote sensing projects and tries to validate them with the help of volunteers. This task requires a large amount of participants. Among other, competitions animate volunteers to contribute to the crowdsourcing project. To motivate volunteers to participate in the remote sensing of land cover is one of the key aspects within the Geo-Wiki.org project. One approach to reach new users for the project is to develop computer games that implement the classification of land cover areas as part of the game.

The project Landspotting is a platform that aims to develop Serious Games for land cover classification. This thesis focuses on the development of a computer game for mobile devices that implements the classification of land cover as an integral part. Previous projects such as the game Landspotting for the platform Facebook are presented in the thesis. This games inspired the design process. Mobile devices with built-in, touch-sensitive displays offer the possibility to classify the land cover by painting appropriate sections of satellite images on the screen. The results of this categorizations are compared with data provided by the Geo-Wiki.org project. The result of this comparison is the basis for the achieved progress in the game.

The as a part of this thesis developed computer game Landspotting was published on the 4th of January 2013. The game was downloaded and installed from a lot of players. The huge amount of land cover data obtained by players illustrates the great potential in the combination of crowdsourcing systems and serious games. The comparison of the data obtained permits both, conclusions about the quality of the results of the developed computer game, as well as on the quality of the results of the Geo-Wiki.org project.

The talk is based on a literature review, which investigates the combination of visualization, interaction, and computational analysis for different types of scientific data. In many disciplines, data and model scenarios are becoming multi-faceted: data are often spatio-temporal and multi-variate; they stem from different data sources (multi-modal data), from multiple simulation runs (multi-run/ensemble data), or from multi-physics simulations of interacting phenomena (multi-model data resulting from coupled simulation models).

This heterogeneity of data characteristics presents new opportunities as well as technical challenges for visualization research. Visualization and interaction techniques are thus often combined with computational analysis. In our survey, we study existing methods for the visualization and visual analysis of multi-faceted scientific data. Based on a thorough literature review, a categorization of approaches is proposed. We cover a wide range of fields and discuss to which degree the different challenges are matched with existing solutions for visualization and visual analysis. This leads to conclusions with respect to promising research directions, for instance, to pursue new solutions for multi-run and multi-model data as well as techniques that support a multitude of facets.

Ultrasound is one of the most frequently used imaging modalities in modern medicine. The high versatility and availability of ultrasound workstations is applied in various medical scenarios, such as diagnosis, treatment planning, intra-operative imaging, and more. Modern ultrasound workstations provide live imaging of anatomical structures, as well as physiological processes, such as blood flow. However, the imaging technique have a high presence of noise, a small scan sector, and are much affected by attenuation artefacts. Thus, traditional techniques for segmentation and visualization are not applicable to ultrasound data.

In this talk, we present our latest advancements in segmentation and visualization techniques, tailored specifically for the characteristics of ultrasound data. We present new methods for interactive vessel segmentation for both 3D freehand and 4D ultra-sound. By directly involving the examiner in the segmentation approach as well as combining data from different probe viewpoints, we are able to obtain 3D models of blood vessels rapidly and robustly.

With the ability of robust vessel extraction, we introduce novel visualization techniques which utilize the previously acquired 3D vessel models. For anatomical imaging, we present a new physics-based approach for volume clipping, enhanced slice rendering and even defining curved Couinaud-surfaces. The technique creates a deformable membrane to adapt to structures in the underlying data, defined either by predefined segmentation, iso-values, or other data attributes.

For functional imaging, medical ultrasound can use the Doppler principle to image blood flow. However, Doppler ultrasound only measures a projected velocity magnitude of the data. In this talk, we present a technique that uses the direction of the blood vessels in order to reconstruct 3D blood flow from Doppler ultrasound. By extending Doppler ultrasound with this directional information, we are able to apply traditional flow visualization techniques for displaying the blood flow. Finally, we investigated the usage of moving particles as a means to depict velocity in flow visualization. Based on a series of studies targeted for motion perception, we present a new compensation model to correct for distortions in the human visual system. This model can help users to make a more consistent estimation of velocities from evaluating the motion of particles.

The increasingly computational power and programmability of modern graphics hardware provides developers of real-time rendering applications with the resources needed to realize more and more complex graphical effects. Some of these effects, like Depth-of-Field, require an efficient image blurring technique to achieve real-time frame rates of 30 frames per second or above. This work presents a comparison of various blurring techniques in terms of their performance on modern graphics hardware. Whereas most of the chosen methods are exclusively used to blur an image, some of them are capable of applying an arbitrary filter. Furthermore, the quality of the different methods has been determined using an automatic process which utilizes a calibrated visual metric. Another aspect when using modern graphics hardware is the increasing scope of operations, especially in the domain of image processing, that can be carried out by using general-purpose computing on graphics processing units (GPGPU). In the recent years, utilizing GPGPU has become increasingly popular inside real-time rendering applications for special tasks like physics simulations. Therefore, all chosen algorithms have been implemented using shaders (GLSL) and GPGPU (CUDA), to answer the question whether or not the usage of a general purpose computing language is applicable for image blurring in real-time rendering and how it compares to using a shading language.

This thesis presents a method to compute soft shadows from environment maps and local light sources on dynamic height fields, extending previous work on height field illumination. While direct illumination in static scenes is very common in video games and 3D applications, real-time global illumination methods supporting dynamic scenes and lights are still an active field of research. In our method, visibility at each receiver point of a height field is determined by the visible horizon, which can be approximated efficiently using a multi-resolution sampling approach. Local light sources are represented by spherical lights and the incident radiance at receiver points is projected into the spherical harmonic basis. Hence, this method produces convincing shadows on dynamic height fields more efficiently than global illumination methods for general geometry.

Current satellite-derived land cover products, which are very important for answering many crucial questions, show huge disagreements. In this paper, we introduce four serious game prototypes { a Facebook strategy game played on Google Maps, a Facebook tagging game, a tower-defense game, and an aesthetic tile game for the iPad { with the purpose of improving global land cover data. We describe the games in detail and discuss the design decisions we made and challenges we faced while developing the games. We evaluate how much the players have already been able to improve global land cover data and provide evidence that games can be a useful way to increase the quality of this data. Finally, we discuss how the main game is being perceived by the players and what has to be further improved to attract a bigger audience.

This paper presents a parallel, implementation-friendly analytic visibility method for triangular meshes. Together with an analytic filter convolution, it allows for a fully analytic solution to anti-aliased 3D mesh rendering on parallel hardware. Building on recent works in computational geometry, we present a new edge-triangle intersection algorithm and a novel method to complete the boundaries of all visible triangle regions after a hidden line elimination step. All stages of the method are embarrassingly parallel and easily implementable on parallel hardware. A GPU implementation is discussed and performance characteristics of the method are shown and compared to traditional sampling-based rendering methods.

Statistics play an important role in sports since they record events like goals, shots, yards, speeds, points, etc. Recording all these events leads to a huge amount of data which needs to be analyzed. Information visualization and sports visualization deal with ways and means how this data can be visualized. The data is used by different experts (coaches, players, media and fans) who are interested in different aspects of the data. Therefore, a goal of this master's thesis is to find out which parts of the data are interesting for the stakeholders. Based on this survey a prototype for the visualization of ice-hockey data will be implemented. Furthermore, it should be possible to view the proposed visualization methods on mobile devices.

Non-photorealistic rendering through stylized lines can be a very useful way to show features and properties of 3D models. In some cases, it can be even more useful tan realistic rendering.

One of the major drawbacks to render stylized line drawings from 3D models is visibility computation. F. Cole and A. Finkelstein present in ?Two Fast Methods for High Quality Line Visibility? a couple of methods for speeding up this computation, obtaining robust results and providing new additional effects.

The goal of the master thesis is the research, comparison, optimization and feature improvement of these methods.

The reconstruction of 3D models out of image data plays an important role in many research areas. Previous work has brought fully automatic or semi-automatic algorithms for extracting 3D data out of multiple images or videos. The output is usually a 3D point cloud, which is a big challenge when further manipulation of the scanned object, such as modeling and texturing, is needed. The goal of this master thesis is the research on methods for creating simple geometry out of such initial data. In the process of the work a modeling tool will be implemented, which enables users to integrate polyhedral primitives in a 3D point dataset. Being assisted by optimization techniques, the user shall be able to simplify such a 3D point cloud to a combination of basic geometric objects, which allow an easy manipulation in further tasks.

In this thesis, an importance sampling algorithm based on a threshold matrix is discussed. The matrix originally designed for ordered dithering is created off-line by a repulsive force field, after which sample points can be obtained by simple thresholding. By the design of the matrix, these samples are distributed irregularly and roughly equidistantly, which is desirable for stochastic sampling. Non-uniform distributions can be achieved by thresholding the matrix with an importance function rather than a constant value. This allows the generation of multidimensional sampling patterns at very low on-line computation cost as compared to importance sampling strategies based on transforming initial (quasi-)random sample sets. The quality of the resulting sample point sets is quantified with the help of established metrics in order to compare the different approaches.

An approach called Inductive Rotation (IR), developed by artist Hofstetter Kurt can be used to create intricate patterns that fill the 2D plane from a single prototile by repeated translations and rotations. Those patterns are aperiodic and have interesting features, both from a mathematical and artistic viewpoint. During the course of my thesis I did some research on the patterns generated by Inductive Rotation and developed algorithms that allow for automatic generation of those patterns. The implementation is called the Irrational Image Generator, a tool that on the one hand is a reference implementation of the IR method, and on the other hand can be used by the artist for further experimentation to fully utilize the artistic possibilities of the IR approach. I have decided on generation algorithms that operate on geometry instead of directly manipulating bitmap data. The program makes use of the GPU through OpenGL to render the resulting patterns as textured polygons.

Volume segmentation is important in many applications, particularly in the medical domain. Most segmentation techniques, however, work fully automatically only in very restricted scenarios and cumbersome manual editing of the results is a common task. In this paper, we introduce a novel approach for the editing of segmentation results. Our method exploits structural features of the segmented object to enable intuitive and robust correction and verification. We demonstrate that our new approach can significantly increase the segmentation quality even in difficult cases such as in the presence of severe pathologies.