About Research Publications Curriculum Vitae
ExSciTecH: Experience Science, Technology and Health
Volunteer Computing (VC) uses idle cycles of personal computers (PC's) connected through the Internet to address fundamental problems in science. VC projects use simple screensavers to attract and retain ordinary people, also called volunteers. At this time, a major problem in VC systems is the lack of user inputs and interactions. For example, in many VC projects targeting molecular simulations, volunteers have played a limited role. Recent research shows that human-centered computation can significantly benefit such VC projects, e.g., by selecting molecule shapes, velocities, and docking sites. The project presented in this poster radically transforms traditional VC systems by involving volunteers as central participants in VC projects supported by powerful graphical interfaces and immersive technologies (i.e., the Nintendo Wiimote). Our final goal is to encourage volunteers to overcome existing barriers and preconceptions towards computers and use Volunteer Computing to both learn and discover science, fostered by immersive gaming environments. We use Docking@Home, a VC project that targets protein-ligand docking simulations, to provide us with the thematic framework for our VC system.


Real-time Projector Depixelation for Videos
The screen-door effect or projector pixelation is a visual artifact produced by many digital projectors. In this paper, we present a real-time projector depixelation framework for displaying high resolution videos. To attenuate pixelation, we use the common defocusing approach of setting the projection a little out-of-focus. Using a camera-projector pair, our system efficiently measures the spatially-varying defocus kernel and stores it as a texture on the graphics hardware. We explore two novel techniques to compensate for the defocusing blur in real-time. First, we develop a steepest descent algorithm on the GPU for estimating the optimally deblurred video frames based on the measured defocus kernel. Second, we present a novel optical flow algorithm that uses an illumination ratio map (IRM) to model illumination transformations between consecutive frames. We store both the IRM and optical flow as textures. To process a frame at runtime, we use the textures to warp the optimized previous frame into an initialization for the iterative GPU optimization. We show that our GPU optimization achieves one magnitude acceleration using this warping. Experiments on various video clips show that our framework can realize real-time video playback with significantly reduced pixelation and defocus while preserving fine details and temporal coherence.


Wiimote for Human Computer Interaction
Computers of the future will count on innovative new technologies to revolutionize the way humans and interact with them. One such technology is the Wiimote from Nintendo. This simplistic and intuitive device can be tailored to a wide variety of interactive tasks including wireless data manipulation and presentation, as well as motion sensing and recognition. A classroom, for example, is an environment which could greatly benefit in employing these tasks. Teachers would no longer sit at a computer showing slides but could move freely about and remain an active presence in their classrooms. Teaching physics or visualizing physical simulations involving acceleration, motion, and trajectory need not be represented only in the form of pictures and diagrams; instead they can be simulated using the Wiimote's motion sensing capabilities. We develop a small application based on the Wiimote, and demonstrate several applications which showcase its features.


Machine Translation and Applications of NLP
As I am interested learning foreign languages and linguistics itself, one of my hobbies has been studying NLP and its application in Machine Translation. Specifically I've focused on several subproblems with respect to generating a machine translated output text from a given input text. The first problem I considered tackeling is that of word order or syntax. In order to generate a correct translation, one must know the grammar of the both the source and target language. Often times manually capturing all the rules and exceptions of a natural language is intractable. Instead I propose a statistical approach which queries the internet to gather evidence to assign probabilities to permuted orderings of words. Given a set of words, this application will return the ordering of the set of words which is most likely to appear in text on the internet. Another sub-problem I have worked on is that of Word Sense Disambiguiation. WSD is important to Machine Translation in that it can be thought of as the intrinsic meaning of the word. For example the word "bark." A translator must know which sense of the word is being used, to correctly translate the word. Is it bark on a tree, or is it a dog's bark. I've developed an application which takes as input, a sentence containing a word to be disambiguated surrounded by a local context (the other words in the sentence). Based on this context we can use the internet to find collocations and assign probabilities of importance between the word to be disambiguated and each word in the local context. The application can then make a prediction as to which sense (of the those provided in word.net) is most likely given the local context.