The PAMIS Project: Parallel Multimedia Index Server

Faculty: Tzi-cker Chiueh

Group Members:

• Peter Deng (pdeng@ecsl.cs.sunysb.edu)
• Kartik Gopalan (kartik@ecsl.cs.sunysb.edu)
• Srinidhi Varadarajan (srinidhi@cs.sunysb.edu)
• Lan Huang (lanhuang@cs.sunysb.edu)
• Chuan-Kai Yang (ckyang@cs.sunysb.edu)

Research Projects

• Content-Based Image Indexing: With the advent of multimedia databases, it becomes crucial to develop efficient mechanisms to access image-oriented data sets other than based on textual keywords. An important access method in such systems is to provide "like-this" image retrieval. We have developed a universal indexing system based on a search data structure called Vantage-Point Tree , which identifies the closest match in a database to a given query while accessing only a small portion of the database elements. The current PAMIS prototype allows indexed retrieval of multiple nearest neighbors of a given query from a 50K-element polygonal shape image database, and is built on top of 10 Pentium machines connected by a 100-MHz Fast Ethernet. Each shape is represented as a turning function. To demonstrate this technology, we have also built a document recognition system, which first extracts the contour of the first page of every document as its signature, and represents it as a polygonal shape. Then the document signatures are stored in PAMIS, which is searched using the Vantage-Point Tree index at run time to match the signature of a submitted document image query. We have demonstrates this system in AIPA '97. We are currently developing a multi-resolution indexing scheme based on wavelet representations, and refining the scanned document recognition system to accommodate noises due to contaminations, different scanning resolutions, misalignment, tilts, and different scales (shrink or expand).
  
  
• Indexed Retrieval of Compressed Text: One of the major problems with existing text retrieval systems is the size of the index files. We have developed an innovative scheme that integrates text compression and text indexing in a unified framework. In particular, the proposed scheme allows direct searching through compressed text without decompressing the text first. Moreover, the framework allows a flexible tradeoff between index storage requirement and retrieval speed so that a text retrieval system can dynamically re-configure itself to optimize the access performance based on observed access patterns. Currently, we are enhancing this technology to include more sophisticated term weighting schemes, as well as to include an automatic document classification scheme for Internet documents.
  
  
• Stony Brook Video Server: As a part of the second phase of the PAMIS project, we have developed a LAN-based distributed video server called Stony Brook Video Server (SBVS). This is the first video server that provides end-to-end performance guarantee from the server's disk subsystem, through the LAN, and to the client machines' display. The current prototype, SBVS-II, runs over a multi-segment Ethernet LAN. That is, the video server and its clients can sit on different Ethernet segments and the video playback is still guaranteed to be smooth. SBVS-II is built on top of an off-the-shelf Pentium 100-MHz machine, and is capable of delivering up to 45 MPEG-1 streams (1.5 Mbits/sec) over a 100-Mbps Fast Ethernet. It currently supports automatic acquisition and indexing of TV programs, and provides keyword-based accesses to video archives. Currently we are building the next-generation of SBVS that features scalable PC clusters and automatic fault tolerance across disk and node failures.
  
  
• Terabyte Storage System: With the goal of storing high-resolution digital video such as surgical procedures and 2D and 3D biomedical image data, we are working on a terabyte storage system called Stonehenge , which features a hierarchical storage architecture consisting of a a large-scale software-driven disk array and a parallel array of automated tape libraries. Our focus is on building innovative storage management software on top of off-the-shelf commercial storage hardare that minimizes the data retrieval response time while cost-effectively scaling the total storage capacity to the terabyte and petabyte range. The techniques we are investigating include multi-resolution data compression schemes and dynamic load balancing of parallel disk arrays.

Publications

• Tzi-cker Chiueh, Dimitris Margaritis, Srinidhi Varadarajan, "Design, Implementation, and Evaluation of a Parallel Image Indexer," in Proceedings of Visual '97, San Diego, CA, December 1997.
• Tzi-cker Chiueh, Srinidhi Varadarajan, "Sase: Implementation of A Compressed Text Search Engine, in Proceedings of USENIX Internet Technologies and Systems, Monterey, CA, December 1997.
• Tzi-cker Chiueh, Chitra Venkatramani, Michael Vernick, "Design and Implementation of the Stony Brook Video Server," in Software -- Practice and Experience, February, 1997.
• Michael Vernick, Chitra Venkatramani, Tzi-cker Chiueh, "Performance Evaluation of Stony Brook Video Server," ECSL-TR-24, February 1997.
• Michael Vernick, Chitra Venkatramani, Tzi-cker Chiueh, "Adventures in Building The Stony Brook Video Server," in Proceedings of ACM Multimedia '96, Boston, MA., October 1996.
• Tzi-cker Chiueh, Allen Ballman, Kevin Kreeger, "Multi-Resolution Indexing of Shape Images," ECSL-TR-41, October 1997.
• Tzi-cker Chiueh, Srinidhi Varadarajan, "Compression-Domain Text Indexing and Retrieval," ECSL-TR-25 (revised), October 1997.

Acknowledgement