=Paper=
{{Paper
|id=Vol-1263/paper87
|storemode=property
|title=MediaEval 2014 Visual Privacy Task: Geometrical Privacy Protection Tool
|pdfUrl=https://ceur-ws.org/Vol-1263/mediaeval2014_submission_87.pdf
|volume=Vol-1263
|dblpUrl=https://dblp.org/rec/conf/mediaeval/KorshunovE14
}}
==MediaEval 2014 Visual Privacy Task: Geometrical Privacy Protection Tool==
https://ceur-ws.org/Vol-1263/mediaeval2014_submission_87.pdf
MediaEval 2014 Visual Privacy Task: Geometrical Privacy
Protection Tool
Pavel Korshunov Touradj Ebrahimi
MMSPG, EPFL MMSPG, EPFL
pavel.korshunov@epfl.ch touradj.ebrahimi@epfl.ch
ABSTRACT other explores the privacy-pleasantness tradeoff, which is about
This paper describes EPFL privacy protection tool for the MediaE- how socially acceptable is a given privacy protection tool for a hu-
val 2014 Visual Privacy task. The goal of the task is to obscure man observer. This year, the task is also separates visual privacy
faces, body silhouettes, and personal items of people in the pro- features into two types: low detailed features, such as body sil-
vided surveillance clips to preserve their personal privacy. The houettes, and features with high details, such as faces or personal
EPFL privacy protection tool mainly relies on two privacy protec- items [1].
tion filters: a warping-based reversible filter to obscure features In the submission to MediaEval 2014 Privacy task, EPFL aimed
with low visual details (body silhouettes) by distorting them with to address both tradeoffs and separately obscure two types of vi-
randomized warping and morphing-based reversible filter to ob- sual features. Therefore, the proposed privacy protection tool con-
scure features with high visual details (faces and personal items) by sists of two privacy protection filters: a warping-based filter [6] that
‘replacing’ them with a graphical representation. The aim of this obscures features with low visual details by distorting them with
tool is to achieve an acceptable balance between privacy (how well randomized warping and morphing-based filter [5] to obscure fea-
the privacy is protected) and intelligibility (how well the surveil- tures with high visual details by ‘replacing’ them with a graphical
lance task can still be performed), as well as, privacy and pleasant- representation. The privacy protection tool is implemented using
ness (how pleasant is the protection). The results of three types of Python, OpenCV1 , and Matlab.
subjective evaluations, via crowdsourcing, practitioners, and stake- Organizers of the task provided video dataset [4] with annota-
holders, provided by the organizers of the task demonstrates that tions of privacy sensitive regions including faces, hair, skin, acces-
EPFL privacy protection tool achieves a great overall balance be- sories, and body regions, as well as classification of these regions
tween privacy, intelligibility, and pleasantness, while being secure into low, medium, oh high detailed features. The tool, therefore,
and reversible in the same time. assumed the privacy regions known (in a practical scenario, they
can be detected by video analytics) and focused on developing the
privacy protection tool that achieves an acceptable balance between
1. INTRODUCTION privacy (how well the privacy is protected) and intelligibility (how
Recent adoption of digital video surveillance systems, especially well the surveillance task can still be performed), as well as, pri-
in public spaces and communities, has significantly increased the vacy and pleasantness (how pleasant is the protection).
concern for intrusion into individual privacy. New sensing tech-
nologies, such as ultra high definition, high dynamic range, or video
capturing with mini-drones, threaten to eradicate boundaries of pri- 2. KEY DECISIONS AND CHALLENGES
vate space even more. As a possible solution, many privacy protec- The best privacy preserving filter would be a blacked out camera
tion tools have been proposed for preserving privacy, ranging from with no video feed, but, in such case, there would be no surveillance
simple methods such as masking blurring, pixelization, or mask- possible and intelligibility would be zero. Therefore, a usable pri-
ing to more advanced methods satisfying the following desirable vacy protection filter should have a balance between privacy and
practical properties: reversibility, robustness, and security. The ad- intelligibility. Similarly, an encryption or scrambling based privacy
vanced methods can be divided into several categories: encryption- filters could lead to high privacy but can be annoying or even scary,
based [7], scrambling-based [2], and geometrical-based [6, 5] meth- resulting in very low pleasantness. Another important practical re-
ods. quirement is the secure reversibility of the privacy protection tool,
Despite wide availability of visual privacy protection tools, with so that the protection can be undone in secure way (only if one has
an exception of some work [3], little is known about which tools a secret key) to restore the original video in case police or court
are suitable for practical applications. To close this gap, MediaE- would require it.
val 2014 Visual Privacy task was designed to facilitate submissions To achieve the balance between privacy, intelligibility, and pleas-
of different protection tools and to benchmark them on practical antness, as well as to provide reversible protection, the proposed
privacy video dataset [4] via several types of subjective evalua- privacy protection tool adopted a two-stage approach: (i) reversible
tions. Moreover, the focus of this task is twofold: one explores the warping filter [6] is applied on body silhouettes as a low detailed
privacy-intelligibility tradeoff, which is between how well surveil- visual feature to distort general personal appearance and (ii) re-
lance can be performed while privacy is being preserved, and an- versible morphing filter [5] on faces and personal items as high de-
tailed visual features to remove all the identifiable details. Figure 1
illustrates how the proposed tool protects privacy of people.
Copyright is held by the author/owner(s). 1
MediaEval 2014 Workshop, October 16-17, 2014, Barcelona, Spain http://opencv.org/
Table 1: Results of three different subjective evaluations for EPFL privacy protection tool compared to average.
Crowdsourcing Stakeholders Practitioners
EPFL Average EPFL Average EPFL Average
Intelligibility 73.2 74.8 67.7 79.3 59.5 69.5
Privacy 51.0 50.2 61.6 46.5 57.3 41.6
Pleasantness 23.6 24.8 40.7 69.5 46.3 59.7
more suitable for the scenarios where the observers are naïve sub-
jects as it is in the case of crowdsourcing. The low intelligibility
score can be compensated by the fact that the tool is reversible and
original video can be securely restored, which would allow the de-
tailed examination of the video data if necessary. Low pleasantness
value is probably due to the choice of graphical representations for
faces and personal items (see Figure 1), which subjects did not like.
A more appropriate and use case oriented choice of such represen-
tation may improve the pleasantness of the visual protection.
4. CONCLUSION
EPFL privacy protection tool combines warping and morphing
privacy protection filters and achieves an acceptable balance be-
tween privacy, intelligibility, and pleasantness, providing, in the
same time, ability to securely restore the original content if neces-
sary. In a practical scenario, a better fitting graphical representa-
tions of the faces and personal items can be selected.
Acknowledgments
This work was conducted in the framework of the EC funded Net-
work of Excellence VideoSense.
5. REFERENCES
[1] A. Badii, T. Ebrahimi, C. Fedorczak, P. Korshunov, T. Piatrik,
Figure 1: Original (above) and privacy protected (below) snapshots V. Eiselein, and A. Al-Obaidi. Overview of the MediaEval
of fighting scene video. 2014 visual privacy task. In MediaEval 2014 Workshop,
Barcelona, Spain, October 16-17 2014.
[2] F. Dufaux and T. Ebrahimi. Scrambling for privacy protection
Warping filter makes the details of the visible object unrecogniz- in video surveillance systems. IEEE Trans. on Circuits and
able (i.e., privacy is increased), but, by controlling the strength of Systems for Video Technology, 18(8):1168–1174, Aug. 2008.
the filter, the overall general shape of the object can be preserved, [3] P. Korshunov, C. Araimo, F. De Simone, C. Velardo,
so it would still be possible to understand what is going on in the J. Dugelay, and T. Ebrahimi. Evaluation of visual privacy
surveillance scene (i.e., intelligibility is not decreased). Morphing filters impact on video surveillance intelligibility. In
filter replaces faces and personal items with the graphical represen- International Workshop on Quality of Multimedia Experience
tation, e.g., ‘smiley face’ instead of the original face, which effec- (QoMEX), pages 150–151, July 2012.
tively removes all personal details, i.e., privacy is increased, with [4] P. Korshunov and T. Ebrahimi. PEViD: privacy evaluation
the aim to keep both intelligibility and pleasantness/appropriateness video dataset. In SPIE Applications of Digital Image
high. Processing XXXVI, volume 8856, San Diego, California,
USA, Aug. 2013.
3. EVALUATION RESULTS [5] P. Korshunov and T. Ebrahimi. Using face morphing to protect
The organizers provided the results from three subjective eval- privacy. In IEEE International Conference on Advanced Video
uations: crowdsourcing-based evaluation in Stream 1, evaluation and Signal-Based Surveillance (AVSS), pages 208–213,
by stakeholders and surveillance experts in Stream 2, and evalua- Krakow, Poland, Aug. 2013.
tion by practitioners and data protection experts in Stream 3. The [6] P. Korshunov and T. Ebrahimi. Using warping for privacy
corresponding results of the EPFL privacy protection tool are sum- protection in video surveillance. In 18th International
marized in the Table 1 and compared against the average of the total Conference on Digital Signal Processing (DSP), pages 1–6,
8 submissions to the Privacy Task of MediaEval 2014. Santorini, Greece, July 2013.
From the table, it can be noted that across all evaluations, the [7] T. Winkler and B. Rinner. TrustCAM: Security and
tool demonstrates higher than average level of privacy but under- privacy-protection for an embedded smart camera based on
performs in terms of intelligibility and pleasantness. In crowd- trusted computing. In IEEE International Conference on
sourcing evaluation, the performance of the tool is nearer to average Advanced Video and Signal Based Surveillance (AVSS), pages
compared to other two evaluations. It means that the tool would be 593–600, Sept. 2010.