Image and Video Databases:
Who Cares?

 VIPER Research Projects

Outline

Penetration of TV/Video

Maybe TVs Should Just Stay Dumb?

What Do Users Want?

Video Database Problem

Content-Based Access Applications

Goals

Application Models

Consumer Model

Consumer Model

Consumer Model

Video-on-Demand

Video-on-Demand

Video-on-Demand

Digital Library Model

Digital Library Model

Digital Library Model

More Comments

Goals

Further Motivation

MPEG-7

What is “Content”?

MPEG-7 Framework

 ViBE: A Video Database Structured for Browsing and Search

ViBE Research Team

The Problem

Video Analysis: Overview

Slide 30

Temporal Segmentation of
Video Sequences

The Temporal Segmentation Problem

Hierarchical Structure of Video

Examples of Some Shot Transitions

Slide 35

Previous Work

Common Approach to
Temporal Segmentation

Problems With This Approach

Working in the Compressed Domain

The DC Sequence

An Example of an Extracted DC Frame

The Generalized Trace (GT) / Regression Tree Methodology

Slide 43

Decision Trees

Feature Windowing

Advantages of the GT/Regression
 Tree Methodology

Detecting Cuts

Detecting Gradual Transitions

Postprocessing of Results

Slide 50

Slide 51

Slide 52

Slide 53

Current Work

HIERARCHICAL SHOT REPRESENTATION

Tree Representation of Shots

PSEUDO-SEMANTIC
SHOT LABELING

Pseudo-Semantic Labeling

Pseudo-Semantic Labeling Problem

Pseudo-Semantic Label

Slide 61

Skin Detection

Skin Detection Examples

Unsupervised Segmentation

Unsupervised Segmentation
Using Chrominance

Example of Unsupervised
Segmentation Using Chrominance

Slide 67

Slide 68

Slide 69

Slide 70

Face Extraction Results

Face Recognition Results

“Indoor/Outdoor” Feature Label

Shot Length Feature

Current Work

BROWSING AND
SEARCHING ENVIRONMENT

Browsing with a Similarity Pyramid

Navigation via the Similarity Pyramid

Slide 79

Slide 80

Database Reorganization Based on the Relevance Set

Video Genre Classification
Using the Pseudo-Semantic Trace

Using Hidden Markov Models to Analyze Video Sequences

The Pseudo-Semantic Trace

HMM Training Procedure for Genres

Slide 86

Slide 87

Slide 88

Slide 89

Slide 90

Slide 91

Future Research