Dr. JunChul Chun:
A CORBA-based Medical Image Analysis and Modeling System
I'll introduce a CORBA-based medical image analysis and modeling system, which is platform independent and provides high accessibility and usability of the system for the users at remotely located sites. The system allows us to manage datasets and manipulates medical images such as segmentation and volume visualization of computed geometry from biomedical images in distributed environments. Using an active contour model and additive re-projection, the system provides boundary information of specific tissue and real time rendering of the 3D volume. In addition, it supports collaborative work among multiple users using broadcasting and synchronization mechanisms. For the development of the system we adopted the DICOM(Digital Image and Communications in Medicine) standard for a method of transferring images and associated information device.
Since the system is developed using Java and CORBA, which provide distributed programming, the remote clients can access server objects via method invocation, without knowing where the distributed objects reside or what operating system it executes on.
Dr. Keechul Jung:
Text In Images
This presentation shows an approach for attacking several difficulties for scene text detection, and shows its affirmative results. Basically, we use the texture-based approach using a multi-layer perceptron (MLP). Using the learning capability of the MLP, we can reduce the difficulty in manually generating texture filters in the case of large variations in size and shape of texts. To reduce the processing times in texture analysis stage, we use a multiple continuously adaptive mean shift (MultiCAMShift) algorithm on the text probability image, produced by a texture classifier using a MLP. We aim at the planarity and rectangularity of the planes including texts. On this assumption, we can get accurate correction results of perspective distortion using warping-parameter-calculation without relatively inaccurate and time-consuming motion estimation. The proposed method has a little processing time and shows satisfactory correction results.
Based on the fact that texts within images are very useful for describing the contents of the image and can be easily extracted comparing with other semantic contents, such as faces, motions, vehicles, etc.
Umut Uludag:
Securing Biometric Data
Biometrics-based personal identification techniques are becoming increasingly popular compared to traditional token-based or knowledge-based techniques such as identification card (ID), passwords, etc. One of the main reasons for this popularity is the ability of the biometrics technology to differentiate between an authorized person and an impostor who fraudulently acquires the access privilege of an authorized person.
On the other hand, the problem of security and integrity of biometric data poses new issues. If a person's biometric data is stolen, it is not possible to replace it unlike replacing a stolen credit card, ID or password. Furthermore, while biometric data provide uniqueness, they do not provide secrecy. For example, a person leaves fingerprints on every surface he/she touches. Encryption, watermarking and steganography are possible techniques to solve this problem.
In this seminar, our research on securing biometric data will be presented. The details of our watermarking/data hiding method will be highlighted. Application scenarios with different characteristics (in terms of host data, communication medium, etc.) will be analyzed. Experimental results will be provided in terms of data decoding accuracy, matching performance and robustness.
Yilu Zhang:
Chained Action Learning through Real-time Interactions
A developmental cognitive learning architecture is proposed for an artificial agent to learn composite behaviors upon the acquisition of basic ones, observed as higher order classical conditioning in animal learning areas. The problem here is to handle missing contexts in developing new stimuli-response associations.
Compared to former works, the proposed architecture enables an agent to conduct learning in unknown environments through online real-time experiences instead of in synthesized or symbolic domains. All possible perceptions and actions, including even the actual number of classes, are not available until the programming is finished and the robot starts to learn in the real world. The learning problem is extremely challenging because of the need for generating internal representation autonomously, the richness of real sensory inputs, and the environmental uncertainty. The careful architecture design and the power of incremental hierarchical discriminant regression (IHDR) technique played an important role in the success.
In the experiments reported, upon learning the basic gripper tip movements, the SAIL robot learned to combine individually instructed movements to be a composite one invoked by a single verbal command without any reprogramming. To solve the problem of missing context in action chaining, we modeled a primed context as the follow-up sensation and action of a real context. By backpropagating the primed context, a real context was able to predict future contexts, which enabled the agent to react correctly even with some missing contexts.
Dr. Ronald Arkin:
Tasking and Execution for Multiagent Robotic Teams
Research conducted over the last decade within the Mobile Robot Laboratory at Georgia Tech has centered on many important issues involving multi-robot teams. In this talk, we broadly review a range of research results: from the role of communication in multiagent robotic systems; to multiagent mission specification tools for building complex robot missions using a graphical user interface validated by usability studies; to formation control for small teams of robots including results demonstrated on two HMMWVs; to team teleautonomy where an operator can interface smoothly at varying levels of autonomy with a large number of robotic agents. These results are currently being applied within three different ongoing DARPA programs that serve as feeder programs for the DARPA/Army Future Combat System (FCS) effort.
Dr. Rama Chellappa:
Monte Carlo Markov Chain Methods for Video Understanding
Recent advances in Monte Carlo Markov Chain (MCMC) methods have enabled the design of elegant algorithms for many problems related to the understanding of video images. In this talk, I will summarize some of our recent work in this area. Specifically, I will present MCMC techniques for tracking humans using their motion, shape and identification. Algorithms for recovering the structure of multiple moving objects will also be presented. Some open issues on critical motions sequences that arise in self-calibration will be discussed and a Bayesian approach for self-calibration is presented. Many illustrative video examples will be shown.
Micky Badgero:
Developmental Motor Map Learning
Developmental Motor Map Learning uses the developmental approach for an artificial agent to learn motor movements. Incremental Pricipal Component Analysis reduces the dimensionality of learned motor responses to reduce complexity and smooth actions.
Chandan Reddy:
Prediction of Spatial Location based on fMRI Signals
Functional Magnetic Resonance Imaging (fMRI) is a powerful technique for studying the working of the human brain. This project explores a novel method for the analysis of fMRI data in order to discover the activation of a network of regions involving some parts of the human brain while a person is navigating in a virtual environment. Spatially sensitive voxels are extracted by selecting those that have high mutual information. Each of these extracted voxels is then used to create a response curve for the stimulus of interest, in this case spatial location. Following the voxel extraction stage, the set of extracted voxel timeseries is treated as a population and used to predict the location of the subject at a random time in the experiment by essential "voting" with their current activity. The approach used for prediction is the Bayesian reconstruction method. The ability to predict the location of a subject in the virtual environment based on brain signals will be useful in developing a physiological understanding of many applications involving virtual environment.
In this talk, I will present an overview of our research in the use of computer vision for wearable systems. We strive to develop robust techniques in computer vision and augmented reality to apply such techniques to wearable interfaces that enable human-centered interaction. I will begin by describing our wearable vision systems, VizWear. Then I will briefly cover the following topics:
For more details, please visit the VizWear website
Rein-Lien (Vincent) Hsu:
Face Detection and Modeling for Recognition
Over the past ten years, face recognition has received substantial attention from researchers in biometrics, pattern recognition, and computer vision communities. This common interest in facial recognition technology among researchers working in diverse fields is motivated both by our remarkable ability to recognize people and by the increased attention being devoted to security. My presentation will start with the applications and challenges of face recognition. I will present my approach to two important techniques to fulfill face recognition: (1) detecting human faces and (2) modeling/representing them. Human faces are detected based on a parametric skin-tone classifier and feature detectors for facial components such as eyes, mouth, and the face boundary. This feature-based face detector can find faces in images with variations in head orientation, pose, and facial expression. The detected faces are then aligned with a 3D generic face model and are represented as a 3D triangular mesh. This object-centered representation allows machines to recognize faces under different illuminations, poses, and expressions. Furthermore, from a 3D face model, we can derive 2D semantic face graphs for identifying faces at a high/semantic level. I will demonstrate the effective classification and visualization of human faces using the derived semantic face graphs.
Arun Ross:
A Survey on Multimodal Biometrics
User verification systems that use a single biometric indicator often have to contend with noisy sensor data, restricted degrees of freedom, non-universality of the biometric trait and unacceptable error rates. Attempting to improve the performance of individual matchers in such situations may not prove to be effective because of these inherent problems. Multimodal biometric systems seek to alleviate some of these drawbacks by providing multiple evidences of the same identity. These systems help achieve an increase in performance that may not be possible using a single biometric indicator. Further, multimodal biometric systems provide anti-spoofing measures by making it difficult for an intruder to spoof multiple biometric traits simultaneously. However, an effective fusion scheme is necessary to combine the information presented by multiple domain experts. In this talk we present a survey of multimodal biometrics.
Miguel Figueroa-Villanue:
Local Tomography Reconstruction using Wavelets
In this talk, I will present a brief review on the techniques available for localized reconstruction of tomography (CT, MRI, etc.) data.