Browse Prior Art Database

System and Method for Analyzing Multimedia Modality Results

IP.com Disclosure Number: IPCOM000020195D
Original Publication Date: 2003-Oct-31
Included in the Prior Art Database: 2003-Oct-31
Document File: 2 page(s) / 175K

Publishing Venue

IBM

Abstract

A program is disclosed that provides multiple graphical views to do an in depth comparison between automatic multi-modal recognition results.

This text was extracted from a PDF file.
At least one non-text object (such as an image or picture) has been suppressed.
This is the abbreviated version, containing approximately 52% of the total text.

Page 1 of 2

System and Method for Analyzing Multimedia Modality Results

  Manual labeling of events, scenes, and objects in multimedia content is an arduous task. Automatic recognition and indexing of multimedia using semantic concept models is becoming a reality. The eventual labeling of multimedia content consists of applying multiple semantic concept models to evaluate the confidence that a semantic concept is present in the content. The various model results are then combined via a weighting function, to give a confidence that the concept is present in the content. Prior to this solution, model results were displayed as confidence ordered thumbnails which could be played back to observe the actual content. In addition, as a separate process, precision recall curves were generated in a graphing program.

The program inputs the results from several models for a variety of semantic concepts. The user may then select a particular semantic concept to analyze. The program will display up to 5 models for a given concept. If more than 5 models exist for the concept, the user may select which 5 models to display. In addition, a list of manually annotated results and a truth table for the selected concept are also displayed. The figure below is a sample of the analysis screen.

1

[This page contains 1 picture or other non-text object]

Page 2 of 2

The current concept, in this case "face" is displayed in the upper left corner. The truth table appears immediately below it. The middle three list boxes display the results for the models detecting the concept "face". The title of each list box is the...