David S. Ebert, Purdue University:
Most scientists, doctors, analysts, and even computer animators are faced with a data deluge. They must analyze, extract meaningful information, and create solutions from huge quantities of data and the size of these data sources is increasing at an enourmous rate. However, screen resolution, image generation techniques, and visualization techniques are only making modest improvements. This has led us to to take a different approach to image generation and visualization that we call perceptualization. In this talk, I'll describe our initial work in perceptualization and our current research endeavors.
Our perceptualization approach will create information that is meaningful to the human perception system and uses as many perceptual channels as feasible, including visual, haptic, and proprioceptive channels. We are concentrating on techniques that make it easier for the user to understand the information in their data using both perceptual cues and techniques from illustration and art. We are also developing visualization techniques that are at a higher and more effective level than traditional techniques that require the user to examine gigabytes to terabytes of data. Through the use of procedural abstraction, visual simulation, and inclusion of domain knowledge, these techniques will enable the user to work with relationships among feature and high-level properties so that they can more effectively perform their task.
Many mesh processing algorithms assume the actual geometry of a triangle mesh to be characterized by the vertex positions only. From the manifold point of view however, triangle meshes have to be considered as continuous piecewise linear surfaces. In sufficiently smooth and flat regions of the surface this observation doesn't really matter since any triangulation will yield a decent approximation to the underlying geometry. In the presence of sharply curved features however, this is not true. Here, severe alias artifacts can affect the perceived surface quality and can lead to quite bad approximation behavior. In my talk I will discuss several consequences of this observation and present recently developed algorithms for feature sensitive mesh generation and re-meshing techniques. I will report recent results in feature sensitive surface extraction from volume data, surface anti-aliasing by remeshing of blend regions in technical data sets, and diffusion based remeshing of triangle meshes.
Reinhard Koch, University of Kiel:
In this presentation we will explore the possibilities of visual-geometric scene reconstruction from uncalibrated image streams, taken by handheld and freely moving cameras. Visual-geometric scene reconstructions try to capture the visual appearance and geometry of real 3D scenes, depending on the type of camera motion and type of 3D scene structure. From the representations one can render new views of the scene with high degree of realism.
Certain constraints on camera motion and scene structure allow to select specific representation models. In case of a purely rotating camera or with 2D background scenes only one may obtain a 2D mosaic scene representation. This representation type is mopdelled by global scene-to-image homographies that can be measured robustly from the image sequence. When full 3D camera motion and full 3D structure is present one must resort to a 3D structure from motion approach that is able to completely model the surfaces of 3D scene objects from a projective reconstruction based on local image correspondences. Between these two extreme representations one may choose from a variety of 2D-3D representations that model the scene with a varying amount of depth information. Examples are image based rendering methods like image interpolation or the lumigraph.
We will discuss these representations and their application to an uncalibrated recording situation, meaning that neither the camera motion nor the intrinsic camera parameters are controlled or calibrated beforehand. We will devellop a unified framework that handles the continuum of 2D- to 3D-representations, based on a selective switching of the underlying representation model. Self-calibration of the intrinsic camera parameters are included in the approach.
The paper is concerned with the handling of situations where multiple visual information occurs and thus the fusion of visual information is required. This is a very common task found in the processing of multisource / multitemporal datasets, in sensor fusion, and in all kinds of active vision systems. A general approach to this problem is presented which goes beyond previous information theoretic investigations. Starting from the paradigm of "Active Fusion", where entropy is used as a measure to evaluate the expected gain in information from a potential data source, we develop the concept of data "consistency". In multisource visual information processing, consistency can be expressed by vicinity in space, by similarity of visual landmarks or by higher level constraints like smoothness of motion trajectories, rigid body, or articulated motion constraints. We find that consistency evaluation is a very powerful method to reduce complexity and to resolve otherwise ill-posed problems like ambiguity. Several sample applications are presented, including an active object recognition system, the definition of salient landmarks, and an optical tracking system.
Ross T. Whitaker, University of Utah:
A confluence of technologies in 3D sensing, high-performance computing, and interactive graphics is providing new opportunities for generating accurate, 3D models of complex scenes. However, as the variety of applications for this capabilities grows, so do the demands on the quality of the models. In some applications, especially those that place time and space restrictions on the data acquisition, the requirements for model fidelity can exceed the raw capabilities of the sensor. Furthermore, as real-life applications become more prevalent, it becomes infeasible to devote large amounts of time to manually modifying the scene to suit the sensing technology or massaging and aligning data sets so that they can be properly processed.
These developments suggest the need for new methods of processing range data that will automatically combine range images in a robust way while making the best use of all of the available information. This talk describes such a framework. The proposed framework relies on estimation theory and incorporates the statistics of the sensor to produce a best estimate of the surface shape and associated parameters. The framework uses a maximum likelihood formulation, which is known to be unbiased and efficient, and can combine measured data with prior information about the scene or the application. This formulation gives rise to a family of algorithms for registration, calibration, and reconstruction, which are shown to be robust and accurate.