CS 545I - Advanced Image Databases

Winter quarter, 1997
Fridays, 3:15 - 4:30
Gates building, Room 100

Oscar Firschein (oscar@infolab.stanford.edu) use for queries,
Dragutin Petkovic (petkovic@almaden.ibm.com).
Ramesh Jain (jain@ece.uscd.edu);
sponsored by Gio Wiederhold (gio@cs.stanford.edu).

An image matching system by James Ze Wang is based on a wavelet approach

Planned schedule for CS545I (click for class summaries)

Seminar abstract

CS545I is a reading and projects seminar devoted to advanced image and video databases, a topic of increasing interest due to large image and video databases soon to be available on the Web. The emphasis will be on combining image-derived descriptors and text descriptors to retrieve local and online images. There are many challenges in storing high-dimensional feature vectors for fast retrieval, and developing metrics of closeness between query and stored vectors. Applications including art collections, sales catalogs, aerial photographs, and medical images will be surveyed to illustrate the strengths or weaknesses of specific techniques.

Retrieval from video databases is also an active area of research. We will be interested in how to automatically construct an inverse story-board (a visual summary) from video, and how to use the audio track to extract frames of interest.

Credit for the seminar only and a presentation will be 1 unit. Additional units (3 normally) under CS-395 will be available for students who wish to execute a project of their own choosing this or successor quarters. Some ideas for projects (contact Dr. Petkovic):

Presentations can be based on the following papers or one of your own choosing. (contact Oscar Firschein, oscar@infolab.stanford.edu, for a copy of the paper and to set up a date for the presentation.)

  1. "A Pattern Thesaurus for Browsing Large Aerial Photographs" UC Santa Barbara
  2. "Video and Image Processing in Multimedia Systems," Furht, Smoliar, and Zhang, Kluwer Academic Press 1995. (pick out a chapter on image compression, indexing schemes for images, video indexing).
  3. "Photobook: Tools for Content-Based Manipulation of Image Databases," MIT Media Lab.
  4. Carnegie-Mellon research for Digital Library on video search using audio track
  5. "Finding pictures of objects in Large Collections of Images," UC Berkeley Digial Library reseach

Seminar format

A typical seminar meeting will consist of an instructor or student presentation of papers from the literature and a discussion of the material. Credit for the seminar only and a presentation will be 1 unit. Additional units (3 normally) under CS-395 will be available for students who wish to execute a project of their own choosing. We will use IBM WWW QBIC for doing research in image databases.


Students should have a background in database concepts of at least CS145.


Oscar Firschein recently retired from the Artificial Intelligence Center, SRI International and a four-year assignment to ARPA, where he was Program Manager for Image Understanding (IU). He has performed research and supervised many advanced projects in image processing. He is, with Marty Fischler (SRI), co-author of a survey book on artificial intelligence and a collection of readings on IU.

Dragutin Petkovic is the Manager, Visual Media Management, at IBM Resesarch in the Alamaden Laboratory. He has led the development of the QBIC project. QBIC received the Seybold award for the most innovative software product in 1995.

Ramesh Jain is Prof. of Electrical & Computer Engineering and Computer Science & Engineering, University of California, San Diego. He is Director of the Visual Computing Laboratory at UCSD, and founder and chairman of Virage, Inc, a developer of systems for visual information retrieval.

Course outline:

1 Introduction (2 lectures)

The users/customers of retrieval from image databases
Problems in describing image content.
Image indexing for content-based retrieval.
Automated and semi-automated approaches.
Image capture, annotation and database population
Integrating image-derived descriptors and conventional SQL-like databases.

<[AH2>2 Commercial Image and Video databases (4 lectures)

IBM (Jan 24,31) and Virage (Feb 7,14)
Overview of current work in image content based retrieval and current applications

WWW - on-line image databases
Overview of some specific issues like low bandwidth and how to overcome them, browsers, copyright issues etc. We will have WWW access in the class and look into several systems on the WWW, including QBIC.

Introduction to QBIC with demo
Details on the technology behind QBIC (texture, shape characterization) Review of the work of others like MIT Media Lab PhotoBook.

3 University Research (3 lectures)

Naval Postgraduate School, Prof. Neil C. Rowe, The MARIE system (Feb 21)
Stanford University, Prof. Gio Wiederhold, Database Indexing Techniques (Feb 28), Prof. Carlo Tomasi, Navigating in Color Images (Mar 7)

4 Student Presentations, Summary, and discussion (2 lectures)

Future work and opportunities
- Image indexing
- Architectures of an extensible system where new features can be added dynamically
- Efficient integration within traditional databases
- User interface issues

Student projects Using WWW QBIC

We will use IBM WWW QBIC, a set of tools that helps the user organize a visual database and find its cataloged images. WWW QBIC uses both QBIC technology and traditional SQL.
You can obtain software that enables you to create a WWW server using QBIC. Mouse the Friday Jan 10 item on the Class schedule, and look at Item 10. Or you can look at wwwqbic.almaden.ibm.com.