MediaEval 2012 Visual Privacy Task: Privacy and Intelligibility through Pixellation and Edge...

16
Privacy and Intelligibility through Pixellation and Edge Detection Prof. Atta Badii, Mathieu Einig School of Systems Engineering University of Reading, UK WWW: http://www.isr.reading.ac.uk eMAIL: [email protected]

description

 

Transcript of MediaEval 2012 Visual Privacy Task: Privacy and Intelligibility through Pixellation and Edge...

Page 1: MediaEval 2012 Visual Privacy Task: Privacy and Intelligibility through Pixellation and Edge Detection

Privacy and Intelligibility through Pixellation and Edge Detection Prof. Atta Badii, Mathieu Einig

School of Systems Engineering University of Reading, UKWWW: http://www.isr.reading.ac.ukeMAIL: [email protected]

Page 2: MediaEval 2012 Visual Privacy Task: Privacy and Intelligibility through Pixellation and Edge Detection

2

Introduction

• Privacy protection by visual anonymisation

• Two main challenges:

- Detecting faces

- Filtering faces

Page 3: MediaEval 2012 Visual Privacy Task: Privacy and Intelligibility through Pixellation and Edge Detection

3

Face Detection

• LBP Face Detector from OpenCV

- Extremely fast

- Good results for close-up frontal faces

• Histogram of Oriented Gradients

- Trained for detecting upper bodies

Page 4: MediaEval 2012 Visual Privacy Task: Privacy and Intelligibility through Pixellation and Edge Detection

4

Face Detection

Page 5: MediaEval 2012 Visual Privacy Task: Privacy and Intelligibility through Pixellation and Edge Detection

5

Face Detection

• Algorithms comparison:

 LBP Cascade

Histogram of Oriented Gradient

Speed + -Long distance - +

Medium distance + +Short distance + =

Light Invariance - +Occlusion Invariance

- +

Front/back discrimination

+ -

Page 6: MediaEval 2012 Visual Privacy Task: Privacy and Intelligibility through Pixellation and Edge Detection

6

Face Detection

• Combination

- Good in most situations

- Cannot differentiate between front and back in some cases

• Tracking

- Hungarian algorithm• Matching made on position and size of the face

- Faces kept even when lost• Face position extrapolated for a few frames

• Duration depends on the number of previous detections

Page 7: MediaEval 2012 Visual Privacy Task: Privacy and Intelligibility through Pixellation and Edge Detection

7

Face Detection

• Front/back discrimination:

- If LBP detector triggered, it is a frontal face

- If not• Assume that people looking at the camera are moving

towards it

• Use tracker to analyse the position and size of the faces

- HMM trained for 3 scenarios:

» Moving towards the camera

» Standing still

» Moving away from the camera

- Anonymisation is required only for the 2 first cases

Page 8: MediaEval 2012 Visual Privacy Task: Privacy and Intelligibility through Pixellation and Edge Detection

8

Face Filtering

• Privacy through pixellation

- Faces reduced to 12x12 pixels

- Additional scrambling with median blur

Page 9: MediaEval 2012 Visual Privacy Task: Privacy and Intelligibility through Pixellation and Edge Detection

9

Face Filtering

• Intelligibility through edge detection

- Sobel filter on the saturation component of the image

- Saturation component is the most ‘robust’ in different lighting conditions

Page 10: MediaEval 2012 Visual Privacy Task: Privacy and Intelligibility through Pixellation and Edge Detection

10

Face Filtering

• Merging of the two filters

Page 11: MediaEval 2012 Visual Privacy Task: Privacy and Intelligibility through Pixellation and Edge Detection

11

Results: Objective Evaluation

• Accuracy

- Overlap between the detected faces and the manual annotation

• Anonymity

- Ratio of faces that could no longer be detected after filtering

• Intelligibility

- Number of people detected even after filtering

• Similarity

- SSIM and PSNR scores

Page 12: MediaEval 2012 Visual Privacy Task: Privacy and Intelligibility through Pixellation and Edge Detection

12

Results: Objective Evaluation

• Results

Criteria Score

Accuracy 0.50 ± 0.19

Anonymity 1.00 ± 0.00

Intelligibility 0.93 ± 0.06

SSIM 0.96 ± 0.02

PSNR 35.80 ± 1.07

Page 13: MediaEval 2012 Visual Privacy Task: Privacy and Intelligibility through Pixellation and Edge Detection

13

Results: Subjective Evaluation

• Questionnaire

- Subjects’ accessories

- Subjects’ gender

- Subjects’ ethnicity

- Rating the perceived effectiveness of privacy protection

- Rating the level of perceived irritation/distraction from the filter

- Recognising filtered faces from a list of clear faces

Page 14: MediaEval 2012 Visual Privacy Task: Privacy and Intelligibility through Pixellation and Edge Detection

14

Results: Subjective Evaluation

• Results:

Page 15: MediaEval 2012 Visual Privacy Task: Privacy and Intelligibility through Pixellation and Edge Detection

15

Conclusion

• Privacy protected to some extent

- One misdetection gives away too much information on the person

- Better face detection is crucial

• Irritation/distraction need to be addressed

Page 16: MediaEval 2012 Visual Privacy Task: Privacy and Intelligibility through Pixellation and Edge Detection

16

Thank you

Atta BadiiIntelligent Systems Research Lab (ISR)

School of Systems EngineeringUniversity of Reading

Whiteknights RG6 6AY UKPhone: 00 44 118 378 7842

Fax: 00 44 118 975 [email protected], www.ISR.reading.ac.uk