TUB-IRML at MediaEval 2014 Visual Privacy Task: Privacy Filtering through Blurring and Color Remapping Dominique Maniry, Esra Acar, Sahin Albayrak DAI Laboratory, Technische Universität Berlin Ernst-Reuter-Platz 7, TEL 14, 10587 Berlin, Germany dmaniry@cs.tu-berlin.de, esra.acar@tu-berlin.de, sahin.albayrak@dai-labor.de ABSTRACT This paper describes the participation of the TUB-IRML group to the MediaEval 2014 Visual Privacy task. We present a method for privacy protection of individuals in surveillance videos. In order to achieve this, our method obscures both shape and appearance of identity-related regions through blurring and color remapping. The intelligibility is preserved by displaying edges and anomalous events are hinted at by special colors. The experimental results obtained on surveil- lance videos show that our method considerably outperforms other participating teams in terms of privacy score. How- ever, the drawback is that the results in terms of intelligi- bility are below average. 1. INTRODUCTION The MediaEval 2014 Visual Privacy Task addresses the problem of privacy protection in video surveillance, which is Figure 1: A walking person is shown as a green sil- gaining more and more importance due to concerns raised houette. about the privacy of monitored individuals. Detailed de- scription of the task, the dataset and the evaluation method- ologies are given in the paper by Badii et al. [1]. As part of the MediaEval 2014 Visual Privacy Task, our privacy fil- event (e.g., fighting, stealing or dropping a bag) happens. In ter is evaluated using the Privacy Evaluation Video Dataset other cases (i.e., non-anomalous), the individuals are shown (PEViD) [2]. in a green color. The aim of this red-green coloring is to en- In the context of this task, we propose a simple but ef- able human operators to focus on any event which requires fective privacy filter which aims not only at obscuring facial particular attention. The second step removes, depending identity, but also at protecting other identity revealing fea- on the blur level and number of colors, most of the shape tures such as accessories and clothing. This is achieved by and appearance information that could potentially reveal a obscuring both shape and appearance of identity-revealing person’s identity, gender or ethnicity, while preserving their regions in videos. movements and actions. In the third step, the obscured image I(x,ˆ y) is blended back into the original frame I(x, y) to create a smooth tran- 2. THE PROPOSED METHOD sition between obscured regions and the background. The The application of our privacy filter is a four-step pro- blending mask mask(x, y) is a binary image where anno- cess. First, we convert each frame into grayscale and apply tated regions have a value of 1 and remaining regions have a a Gaussian blur to all privacy-related regions of the frame. value 0. The smoothing is achieved by applying a Gaussian The intensity of the blurring can be controlled using three blur to the blending mask. The result is: different blur levels (obtained by varying the standard devi- ation of the Gaussian kernel) for regions labeled with low, ˆ y) + (1 − mask(x, y)) · I(x, y) result(x, y) = mask(x, y) · I(x, medium and high privacy requirements. As a second step, the pixel values are quantized to a given number of values (e.g., 8). These values are remapped to In the final step, we target a better intelligibility by in- either a green or red color with the corresponding pixel in- cluding some shape information in the image. The obscured tensity, so that the relation between light and dark regions regions are overlaid with edges obtained with Canny Edge remains same. The red color is used whenever an anomalous detection. Edges in regions with a high privacy require- ment (i.e., faces) are discarded in order not to reveal identity through the edges of facial features. The remaining edges Copyright is held by the author/owner(s). are emphasized using morphological dilation with a 3x3 cir- MediaEval 2014 Workshop, October 16-17, 2014, Barcelona, Spain cle as structuring element. our method and the other participating teams are summa- rized in Figure 3, Figure 4 and Figure 5, respectively. 80,00% 70,00% 60,00% 50,00% intelligibility score 40,00% privacy score 30,00% pleasantness score 20,00% 10,00% 0,00% TUB-IRML Median Figure 3: Stream 1 results. 90,00% 80,00% Figure 2: The silhouettes of two people fighting. 70,00% 60,00% The red color indicates that an anomalous event is 50,00% intelligibility score privacy score happening. 40,00% pleasantness score 30,00% 20,00% 10,00% 0,00% Table 1: User study results for Stream 1, 2 and 3. TUB-IRML Median Stream 1 Stream 2 Stream 3 Our Intelligibility 73.4% 67.7% 60.7% Figure 4: Stream 2 results. Median 74.9% 79.3% 69.6% Our Privacy 69.4% 80.0% 78.7% Median 50.2% 46.5% 41.7% 90,00% 80,00% Our Pleasantness 22.7% 46.0% 52.6% 70,00% Median 24.8% 69.6% 59.7% 60,00% 50,00% intelligibility score 40,00% privacy score 30,00% pleasantness score 20,00% 3. RESULTS AND DISCUSSION 10,00% 0,00% Our submitted run was created using a constant blur level TUB-IRML Median of 14 for all three privacy levels. The number of colors is 8. This choice of parameters favors privacy over intelligi- Figure 5: Stream 3 results. bility. The submissions of eight teams have been evaluated in a user study. The user study has been conducted with three different groups (i.e., streams). Stream 1 represents 230 crowd-sourcing workers, Stream 2 is 65 people working 4. CONCLUSIONS at Thales (mainly in Research&Development – R&D) and In this paper, we proposed a privacy filter that obscures Stream 3 has 59 participants from sectors including R&D, both shape and appearance of privacy-related regions. The data protection and law enforcement from all around the user study has shown that our method is very effective at world. The results for our method and the median across protecting privacy. As a future work, we plan to evaluate dif- all 8 submissions can be seen in Table 3. ferent parameters to find a suitable balance between privacy Among the participants, our proposed method achieved and intelligibility for different contexts. Another interesting the highest privacy score. The privacy protection of our future work would be to improve the appropriateness by re- method still comes with a trade-off in intelligibility, as seen ducing the obscured regions using a pixel-wise segmentation. by the consistent below-average scores. We think that this could be improved by adding additional hints during and Acknowledgments after anomalous events. The research leading to these results has received funding The appropriateness/pleasantness score is also consistently from the European Community FP7 under grant agreement below average. One possible cause for this is that the pri- number 261743 (NoE VideoSense). vacy filter obscures the whole rectangular region around a person including a significant portion of the background. This could be improved with a pixel-wise foreground seg- 5. REFERENCES mentation. However, this requires the foreground segmen- [1] A. Badii, E. Touradj, C. Fedorczak, P. Korshunov, tation to be very accurate, since every false positive could T. Piatrik, V. Eiselein, and A. Al-Obaidi. Overview of potentially reveal identity-related information. Another un- the mediaeval 2014 visual privacy task. In MediaEval pleasant artifact is the blinking of the overlaid edges. When 2014 Workshop, Barcelona, Spain, October 16-17 2014. edge values oscillate around threshold values, the edges can [2] P. Korshunov and T. Ebrahimi. PEViD: privacy become distracting. We think that adaptive thresholds or evaluation video dataset. In Applications of Digital temporal smoothing should be explored as a future work. Image Processing XXXVI, San Diego, CA, August The evaluations of Stream 1, Stream 2 and Stream 3 for 25-29 2013.