Visual Privacy Task Overview Tomas Piatrik, Queen Mary University of London Atta Badii, University of Reading
Visual Privacy Task Overview
Tomas Piatrik, Queen Mary University of London
Atta Badii, University of Reading
Task Description
• Visual privacy protection of people occurred in a video taking into consideration the acceptability and attractiveness of the resulting obscured or scrambled region.
Data
200 videos • 30s-90s• 25fps, 30fps• 640x480, 704x576• Daytime• Evening • People wearing accessories (hat, scarf, sunglasses)
Evaluation Set – 12 videos (4 per each category)
Evaluation Methodology
1. Objective Evaluation• Accuracy of the bounding boxes• Anonymity – evaluated by automatic face detection• Intelligibility – evaluated by object tracking• Quality: SSIM & PSNR – similarity between the
original and obscured images
2. User Study (subjective evaluation)• 20 subjects • 2-range score (0,1; 0-0,5-1)
User Study
Questionnaires • General Evaluation
• Clip Evaluation - cropped obscured face
• Clip Evaluation – whole frames
User Study
Questionnaire: • Is the person of interest wearing sunglasses? … • What is the person’s gender?• What is the person’s ethnicity?• Is it possible to guess the person’s facial expression?• Is the masking hiding sufficient features? • Do you feel that this masking technique could protect
your identity if used on CCTV systems?• Do you find the masking effect distracting?• Do you find the masking effect irritating?• …
Results – objective evaluation
Participants:
Objective evaluation:
co-organising team
Accuracy Anonymity Intelligibility SSIM PSNR
EPFL 0.24 1.00 0.85 0.92 28.26
Eurecom 0.37 1.00 0.82 0.93 32.89
UoR 0.50 1.00 0.93 0.96 35.80
Results – User Study
Results – User Study
Thank You