UCLA Statistics Department       Center for Image and Vision Science

Past Projects

A Web Calendar Integrating More Sharing Functionalities (Nov 2006 - Dec 2006)

Supports finding "common free time" among a group of people, etc. This is a course project of Stat202A.

Multiple Instance Learning for High-level Feature Extraction (March 2006 -- June 2006)

Using High-level Features for Video Retrieval (July 2005 -- Oct 2005)

This is a following-up research project after Trecvid 2005. The main innovation point is in introducing Mutual Information measure for weighting high-level features.

High-level (or concept) features are utilized as a main engine for finding target video/image. The idea of this approach, in brief, lies in representing both query and samples(dataset) in Concept Space, and evaluating similarity between query and samples. In order to fuse different high-level features, we can examine the relavance between each high-level feature and ideal result set, and this is where Mutual Information comes in. Morever, Text-based video retrieval model(which often works better because of its technical maturity) can also be integrated into this weighting scheme of features. An abstract of our submitted work is available here. The diagram below also illustrate the framework:

Parallelized Texture Synthesis on CMP model

I am working on parallelizing image "quilting"[reference 2] algorithm on CMP (chip multi-processor) model to accelerate texture synthesis. Basic parallelization methods are analized and compared for several CMP system models and various image sizes, with speedup being the main criterion. Among all factors which impact performance, cache miss is the most complex one to estimate, and my result is mainly about this. As a course project it's done. The report is in Chinese; English version will be available later.

Image quilting and its parallelization

References:

1. Lance Hammond Basem A. Nayfeh Kunle Olukotun, "A Single-Chip Multiprocessor", IEEE Computer, Sep 1997.
2. EFROS, A., AND FREEMAN, W. 2001. "Image quilting for texture synthesis and transfer", ACM SIGGRAPH 2001, 341-346.
3. Sylvain Lefebvre and Hugues Hoppe, "Parallel Controllable Texture Synthesis", ACM SIGGRAPH 2005, 777-786.

Content-based Video Retrieval

Following Trecvid2005 Video Retrieval Search Task (see below), we are going to further explore content-based video retrieval, aiming to put more features to use. Concept (high-level) features (e.g. person, meeting) have been "extracted" from video data (video shots) for use of search task, in which we understand the query topic by representing it with concepts, and then locate the specific video data we want. Query understanding and concept feature extraction are two main topics to explore in this project. Motion feature, although relatively immature, will also be experimented on and integrated.

- Datamining: Improving performance of NaiveBayes in WEKA

- Virtual Reality System Design: Endoscopic Surgery Training System

- Trecvid 2005 Video Retrieval Search Task

This content-based video retrieval task aims at locating video shots relevant to the search query, as many and accurate as possible, from a large database containing numerous videos and subtitle scripts. It takes advantage of different features extracted from database, including: low-level image feature (e.g. location and orientation of edges), high-level/concept feature (e.g. person) and text features (e.g. word frequency), and fusion on all of them. For each approach, we built an independent retrieval system, and then had them integrated. Our technical report elaborates the approaches used in this task.

This project is supported by National Key Basic Research Project of China (2004CB318108) and National Natural Science Foundation of China (60135010).

Please see the official website of Trecvid for reference: http://www-nlpir.nist.gov/projects/trecvid/

Singular Point Based Fingerprint Classification

My work in this project mainly focused on fingerprint classification into several types (such as Arch, Left Loop, Whorl, etc.) by the number and geometrical information of singular points, computed via Poincare index (a quantity related to orientation spin around the point of interest). To guarantee robustness in locating singular points, image refining techniques such as Gabor filtering were used. Machine learning approaches like SVM and HMM were also tried to utilize global orientation tendency of the fingerprint image, which provide additional information about the type of fingerprint.