1. TUB-IRML at MediaEval 2014 Visual Privacy Task:Privacy
Filtering through Blurring and Color RemappingDominique Maniry,
Esra Acar, Sahin AlbayrakCompetence Center Information Retrieval
& Machine Learning
2. OutlineThe Privacy FilterSample Outputs of the
FilterDiscussion on the FilterPerformance EvaluationConclusions
& Future Work17 October 2014 TUB-IRML at MediaEval 2014 Visual
Privacy Task 2
3. The Privacy Filter (1)Main idea: To obscure both shape and
appearance ofidentity-related regions through blurring and
colorremapping.Preserve the intelligibility bydisplaying edges,
andhinting anomalous events through special colors.17 October 2014
TUB-IRML at MediaEval 2014 Visual Privacy Task 3
4. The Privacy Filter (2)The filter contains four steps:Step 1:
Blur all privacy-related regionsStep 2: Reduce number of colors
& remap colorsStep 3: Apply a blending maskStep 4: Include
shape information by incorporatingedges17 October 2014 TUB-IRML at
MediaEval 2014 Visual Privacy Task 4
5. Step 1: Blur17 October 2014 TUB-IRML at MediaEval 2014
Visual Privacy Task 5
6. Step 2: Reduce Colors17 October 2014 TUB-IRML at MediaEval
2014 Visual Privacy Task 6
7. Step 2: Remap Colors17 October 2014 TUB-IRML at MediaEval
2014 Visual Privacy Task 7
8. Step 3: Apply a Blending MaskThe blending mask mask(x, y) is
a binary image whereannotated regions have a value of 1 and
remainingregions have a value 0.The smoothing is achieved by
applying a Gaussian blurto the blending mask.17 October 2014
TUB-IRML at MediaEval 2014 Visual Privacy Task 8
9. Step 4: Include Shape Information (1)The obscured regions
are overlaid with edgesobtained with Canny Edge detection.Edges in
regions with a high privacy requirement (i.e.,faces) are
discarded.The remaining edges are emphasized usingmorphological
dilation with a 3x3 circle as structuringelement.17 October 2014
TUB-IRML at MediaEval 2014 Visual Privacy Task 9
10. Step 4: Include Shape Information (2)A walking person Two
people fighting17 October 2014 TUB-IRML at MediaEval 2014 Visual
Privacy Task 10
11. Sample Outputs of the Filter (1)17 October 2014 TUB-IRML at
MediaEval 2014 Visual Privacy Task 11
12. Sample Outputs of the Filter (2)17 October 2014 TUB-IRML at
MediaEval 2014 Visual Privacy Task 12
13. Discussion on the FilterPros Cons Parameters to tune
trade-offbetween privacy and intelligibility(blur intensity and
number ofcolors). Remapped colors can conveyadditional information.
Different regions can havedifferent privacy levels by
usingdifferent blur intensities (e.g.,face more blurred than
fullbody).17 October 2014 TUB-IRML at MediaEval 2014 Visual Privacy
Task 13 Simple. Identity related details can leakthrough
shape.
14. Performance Evaluation (1)Stream 1: 230 crowd-sourcing
workers.Stream 2: 65 people working at Thales (mainly in
R&D).Stream 3: 59 participants from sectors including R&D,
data protection andlaw enforcement from all around the world.17
October 2014 TUB-IRML at MediaEval 2014 Visual Privacy Task 14
16. Conclusions & Future WorkThe user study has shown that
our method is veryeffective at protecting privacy.Future
workEvaluating different parameters to balance privacy
andintelligibility, and Improving the appropriateness by reducing
the obscuredregions using a pixel-wise segmentation.17 October 2014
TUB-IRML at MediaEval 2014 Visual Privacy Task 16
17. M.Sc.Competence Center Information Retrieval &Machine
Learningwww.dai-labor.deFonFax+49 (0) 30 / 314 74+49 (0) 30 / 314
74 003DAI-LaborTechnische Universitt BerlinFakultt IV
Elektrontechnik & InformatikSekretariat TEL
14Ernst-Reuter-Platz 710587 Berlin, Deutschland17Esra
[email protected]!013TUB-IRML at 17
October 2014 MediaEval 2014 Visual Privacy Task