ARDA VACE

1
ARDA VACE

• Advanced Research and Development Activity (ARDA)
  Video Analysis and Content Extraction (VACE)

2
Video Analysis and Content Extraction (VACE)

• Project Description
• Goal Develop revolutionary advances in automatic
  video extraction, understanding, and
  multi-modal-fusion.
• Funded in 3, 2-year phases. (we are currently in
  the middle of phase 2.
• Multi-domain meetings, Broadcast News, UAV
  (Unmanned Air Vehicle), Surveillance/Ground recon
• NIST is supporting VACE meeting video needs and
  evaluation oversight
• Sponsor
• ARDA
• Participants

• Progress/Insights
• Building corpus and evaluation infrastructure for
  video extraction community by applying knowledge
  from speech recognition/understanding evaluation.
• Developing diagnostic metrics with the mind of
  moving toward a single-number metric.

• Future Plans
• 2005 Core evaluations
• face detect/tracking
• Text detect/track/recognition
• person detect/tracking
• hand detect/tracking
• 2006
• event detection

• BAE
• BBN
• USC
• UMD
• TASC
• USF
• UCF

• Virginia Tech
• U Illinois Urbana Champaign
• Salient Stills
• IBM/Columbia
• Carnegie Mellon U
• Pittsburgh Pattern Recognition
• Advanced Interfaces

3
Issues

• Having enough data of the types being
  evaluated.
• Ground Truth (Currently we use ViPER developed
  at Univ. Maryland)
• Temporal/Spatial
• bounding box or point reference
• object definition
• Metrics
• Funding
• Definition of Purpose

4
Evaluation Domain Sub-Forums

• Broadcast News (in coordination with TRECVID)
• Meetings
• Surveillance
• UAV
• Tasks
• Develop evaluation recommendations for 2006
  event detection tasks
• These include task definitions, metrics,
  schedule, source data, and necessary annotations
• Chairs work with both research community and
  interested partners to satisfy Sponsor and
  Research needs
• Identify existing research sour data to be used
  or list realistic requirements, ensuring that
  corpora meet above needs
• produce task list and required data resources
• Provide recommendations to Core Evaluation Forum
• NIST will create mail lists and can set up
  conference calls
• Forums should hold biweekly teleconferences as
  necessary to maintain momentum
• Forum chairs will give report at Fall Workshop

5
Planned Core Eval Tasks/Domains (34 evaluations)
Domain Task Meetings Broadcast News UAV Surveillance
Text Detection (English) - Y - -
Text Tracking (English) - Y - -
Text Recognition (English) - Y - -
Text Detection (Chinese) - - - -
Text Tracking (Chinese) - - - -
Text Detection (Arabic) - Y - -
Text Tracking (Arabic) - Y - -
Face Detection Y Y - Y
Face Tracking Y Y - Y
Hand Detection Y Y - Y
Hand Tracking Y Y - Y
People Detection Y - Y Y
People Tracking Y - Y Y
Vehicle Detection - - Y Y
Vehicle Tracking - - Y Y
6
http//www.nist.gov Information Technology
Laboratory http//www.itl.nist.gov itlab_at_nist.gov
Information Access Division http//www.itl.nist.g
ov/iaui/ john.garofolo_at_nist.gov rachel.bowers_at_nist
.gov ARDA http//www.ic-arda.org