CfAR Local Seminars

Date	Speaker	Title
09/10/04	Xuejun Hao	Efficient Geometry and Illumination Representations for Interactive Protein Visualization
09/17/04	Larry Davis, UMIACS	Monitoring Human Activity using Computer Vision
09/24/04	Hanan Samet	SIMILARITY SEARCHING AND K-NEAREST NEIGHBOR FINDING USING THE MAXNEARESTDIST ESTIMATE
10/01/04	Chang Ha Lee	Light Collages: Lighting Design for Effective Visualization(CSIC1121)
10/08/04	Ravi Ramamoorthi, Columbia	Signal-Theoretic Representations of Appearance(CSIC1121)
10/15/04	ĄĄ	ĄĄ
10/22/04	Daniel F. DeMenthon	Mousing Around with a TV Remote that Sees(AVW2120)
10/29/04	ĄĄ	ĄĄ
11/05/04	ĄĄ	ĄĄ
11/12/04	Rene Vidal, JHU	ĄĄ
11/19/04	Patrick Baker, NRL	ĄĄ
12/03/04	Stefan Jaeger	Classifier Combination
12/10/04	Changjiang Yang	Efficient Kernel Machines via the Fast Gauss Transform
01/28/05	Haibin Ling	Using the Inner-Distance for Classification of Articulated Shapes
02/18/05	ĄĄ	NO "cfar weekly seminar" today

Top

Abstracts

09/10/04	Efficient Geometry and Illumination Representations for Interactive Protein Visualization
Speaker	Dr. Xuejun Hao
Abstract	We explore techniques of designing efficient geometric and illumination data representations for developement of algorithms that achieve better interactivity for visual and computational proteomics, as well as graphics rendering. In particular, we will talk about efficient computation and visualization of molecular electrostatics, together with interactive rendering of translucent materials. Molecular electrostatics is important for studying the structures and interactions of proteins, and is vital in many computational biology applications, such as protein folding and rational drug design. We have developed a system to efficiently solve the non-linear Poisson-Boltzmann equation governing molecular electrostatics. Our system simultaneously improves the accuracy and the efficiency of the solution by adaptively refining the computational grid near the solute-solvent interface. In addition, we have explored the possibility of mapping the PBE solution onto GPUs. We use pre-computed accumulation of transparency with spherical-harmonics-based compression to accelerate volume rendering of molecular electrostatics. In addition, we present a compact mathematical model to efficiently represent the six-dimensional integrals of bidirectional surface scattering reflectance distribution functions (BSSRDFs) to render scattering effects in translucent materials interactively. Our analysis first reduces the complexity and dimensionality of the problem by decomposing the reflectance field into non-scattered and subsurface-scattered reflectance fields. While the non-scattered reflectance field can be described by 4D bidirectional reflectance distribution functions (BRDFs), we show that the scattered reflectance field can also be represented by a 4D field through pre-processing the neighborhood scattering radiance transfer integrals. We use a novel reference-points scheme to compactly represent the pre-computed integrals using a hierarchical and progressive spherical harmonics representation. Our algorithm scales linearly with the number of mesh vertices. ĄĄ Top
09/17/04	Monitoring Human Activity using Computer Vision
Speaker	Dr. Larry Davis, UMIACS
Abstract	During the past decade, my students have been studying problems related to detection, tracking and behavioral analysis of humans in action. This talk will review that research, emphasizing work on real time vision algorithms for visual surveillance. After some amusing historical images and videos, I'll focus on visual surveillance, surveying the work done by my graduate students over the past several years, and ending with an overview of the new VACE video exploitation project, and its goals for the next few years. ĄĄ Top
09/24/04	SIMILARITY SEARCHING AND K-NEAREST NEIGHBOR FINDING USING THE MAXNEARESTDIST ESTIMATE
Speaker	Dr. Hanan Samet, Dept. of CS, Univ. of MD
Abstract	Similarity searching often reduces to finding the k nearest neighbors to a query object. Finding the k-nearest neighbors is achieved by applying either a depth-first or a best-first algorithm to the search hierarchy containing the data. These algorithms are generally applicable to any index based on hierarchical clustering. The idea is that the data is partitioned into clusters which are aggregated to form other clusters, with the total aggregation being represented as a tree. These algorithms have traditionally used the estimate of the minimum distance at which a nearest neighbor can be found (termed MinDist) to prune the search process by avoiding the processing of some of the clusters as well as individual objects when they can be shown to be farther from the query object q than any of the current k nearest neighbors of q. An additional pruning technique that uses an estimate of the maximum possible distance at which a nearest neighbor is guaranteed to be found (termed MaxNearestDist) is described. The MaxNearestDist estimate is adapted to enable its use for finding the k nearest neighbors instead of just the nearest neighbor (i.e., k=1) as in its previous uses. Both the depth-first and best-first k-nearest neighbor algorithms are modified to use the MaxNearestDist estimate, which is shown to enhance both algorithms by overcoming their shortcomings. In particular, for the depth-first algorithm, the number of clusters in the search hierarchy that must be examined is reduced thereby lowering its execution time, while for the best-first algorithm, the number of clusters in the search hierarchy that must be retained in the priority queue used to control the ordering of processing of the clusters is reduced, thereby lowering its storage requirements. In this talk we discuss issues in similarity searching and review the two common k nearest neighbor finding algorithms while also showing how to modify them to incorporate the MaxNearestDist estimate. Hanan Samet received the B.S. degree in engineering from the University of California, Los Angeles, and the M.S. Degree in operations research and the M.S. and Ph.D. degrees in computer science from Stanford University, Stanford, CA. He is a Fellow of the IEEE, ACM, and IAPR (International Association for Pattern Recognition). In 1975 he joined the Computer Science Department at the University of Maryland, College Park, where he is now a Professor. He is a member of the Computer Vision Laboratory of the Center for Automation Research and also has an appointment in the University of Maryland Institute for Advanced Computer Studies. At the Computer Vision Laboratory he leads a number of research projects on the use of hierarchical data structures for geographic information systems. His research group has developed the QUILT system which is a GIS based on hierarchical spatial data structures such as quadtrees and octrees, the SAND system which integrates spatial and non-spatial data, the SAND Browser which enables browsing through a spatial database using a graphical user interface, the VASCO spatial indexing applet (found at http://www.cs.umd.edu/~hjs/quadtree/index.html), and a symbolic image database system. ĄĄ Top
10/01/04	Light Collages: Lighting Design for Effective Visualization
Speaker	Chang Ha Lee, Xuejun Hao, and Amitabh Varshney
Abstract	We introduce Light Collages - a lighting design system for effective visualization based on principles of human perception. Artists and illustrators enhance perception of features with lighting that is locally consistent and globally inconsistent. Inspired by these techniques, we design the placement of light sources to convey a greater sense of realism and better perception of shape with globally inconsistent lighting. Our algorithm segments the objects into local surface patches and uses a number of perceptual heuristics, such as highlights, shadows, and silhouettes, to enhance the perception of shape. We show our results on scientific and sculptured datasets. ĄĄ Top
10/08/04	Signal-Theoretic Representations of Appearance
Speaker	Ravi Ramamoorthi, Columbia
Abstract	Many problems in computer graphics require compact and accurate representations of the appearance of objects, and the mathematical algorithms to manipulate them. For instance, high quality real-time rendering needs models for appearance effects like natural illumination from wide-area light sources such as skylight, realistic material properties like velvet, satin, paints, or wood, and shading effects like soft shadows. These effects are also important in many computer vision problems like recognition and surface reconstruction. In these problems, we must often deal with complex high-dimensional spaces. For instance, for real-time relighting in computer graphics, or for lighting-insensitive recognition in computer vision, we must consider the space of images of an object under all possible lighting conditions. Since the illumination can in principle come from anywhere, the appearance manifold would seem to be infinite-dimensional. However, one can find lower-dimensional and more compact structures that lead to efficient algorithms. In this talk, we discuss a signal-theoretic approach to representing appearance, where the illumination and reflection function are signals and filters, and we apply many signal-processing tools such as convolution, wavelet-based representation and non-linear approximation. These representations and tools are applicable to a variety of problems in computer graphics and vision, and we will present examples in real-time rendering in computer graphics, as well as image-based and inverse problems, multiple scattering for volumetric effects, and efficient sampling for image synthesis. ĄĄ Ravi Ramamoorthi is currently an assistant professor of Computer Science at Columbia University, since August 2002, when he received his PhD in computer science from Stanford University. He is interested in many aspects of computer graphics and vision, including mathematical foundations, real-time photorealistic rendering, image-based and inverse rendering, and lighting and appearance in computer vision. ĄĄ Top
10/22/04	Mousing Around with a TV Remote that Sees
Speaker	Daniel DeMenthon, LAMP, UMD
Abstract	I will describe our research for the development of a TV remote that can be used as a mouse device. A miniature camera in the TV remote can be used to compute the remote's motion. The user handles the remote as if it was a laser pointer to control a cursor on a TV screen and select DVD chapters or menu items. This talk will focus on computer vision solutions rather than hardware issues. Two processes are combined in our approach, a fast feature tracking process that computes and cumulates camera motions, and a slower screen recognition process that attempts to detect and recognize the screen in order to cancel the drift of the first process when screen recognition is successful. With this combination, cursor positions are being computed even when the TV screen is not recognized or is out of the field of view. Screen recognition includes two steps, detection and verification. The screen detection uses an iterative method to find the pose and correspondences for a screen model consisting of a list of line segments. The screen model combines lines from the frame around the TV screen, and lines found by image processing of what is displayed on the screen (available from video RAM). A verification of the detected screen is performed by correlating the pixel map contained inside the area detected as belonging to the screen with the pixel array displayed on the screen. I will give some details about how we used OpenCV in this project. This is work with Phil David, Ankur Mohan and Larry Davis. Top
12/03/04	Classifier Combination
Speaker	Dr. Stefan Jaeger
Abstract	Classifier Combination has turned out to be a powerful tool for achieving higher recognition rates in fields where single classifier systems require tedious optimization efforts. Despite the intensive investigation of multiple classifier systems in the recent past, however, a convincing theoretical foundation of multiple classifier systems is still missing. Lacking proper mathematical concepts, many systems use empirical heuristics and ad hoc combination schemes. In my talk, I will present an information-theoretical framework for combining confidence values generated by different classifiers. My main idea is to normalize each confidence value in such a way that it equals its informational content. Based on Shannon's notion of information, I measure information by means of a performance function that estimates the classification performance for each confidence value on an evaluation set. Having equalized each confidence value with its information actually conveyed, I apply the elementary sum-rule to combine confidence values of different classifiers. At the end of my talk I will present experiments for combined on-line/off-line Japanese character recognition, showing clear improvements over the best single recognition rates. Top
12/10/04	Efficient Kernel Machines for Learning via the Fast Gauss Transform
Speaker	Changjiang Yang
Abstract	In this talk, we will review the kernel methods and a recently "rediscovered" Support Vector Machine: Regularized Least-Squares Classification (RLSC). The amount of computation required for such kernel machines with $N$ training samples is $O(N^3)$. This computational complexity is significant even for moderately sized problems and is prohibitive for large datasets. We present an approximation technique based on the improved fast Gauss transform to reduce the computation into $O(N)$. We also give error bound for the approximation, and show experimental results on the UCI datasets without significant decrease in the accuracy. (joint work with Ramani Duraiswami, Larry Davis and Nail Gumerov) ĄĄ Top
01/28/05	Using the Inner-Distance for Classification of Articulated Shapes
Speaker	Haibin Ling, UMD
Abstract	We propose using the inner-distance between landmark points to build shape descriptors. The inner-distance is defined as the length of the shortest path between landmark points within the shape silhouette. We show that the inner-distance is articulation insensitive and more effective at capturing complex shapes with part structures than Euclidean distance. To demonstrate this idea, it is used to build a new shape descriptor based on shape contexts. After that, we design a dynamic programming based method for shape matching and comparison. We have tested our approach on a variety of shape databases including an articulated shape dataset, MPEG7 CE-Shape-1, Kimia silhouettes, a Swedish leaf database and a human motion silhouette dataset. In all the experiments, our method demonstrates effective performance compared with other algorithms. PS: the TR version is available at http://www.cs.umd.edu/~hbling/Research/Inner-Distance/inner-dist-tr.pdf Top

CfAR WEEKLY SEMINAR

Abstracts