MediaEval 2014 Visual Privacy Task: Context-Aware Visual Privacy Protection Atta Badii Ahmed Al-Obaidi ISR Laboratory ISR Laboratory University of Reading, UK University of Reading, UK atta.badii@reading.ac.uk a.al-obaidi@reading.ac.uk ABSTRACT In this paper, we describe a privacy filter proposed for VPT 2014 in an attempt to provide a context-aware solution. The proposed solution comprises three different techniques applied to the face, skin, and body regions separately. The proposed combination of filtering techniques aimed to produce an adaptive solution and provide an example of a context-aware-like privacy filtering capability. The results demonstrated the effectiveness of the proposed techniques in maintaining a high level of Intelligibility and retaining the appeal of the video i.e. Pleasantness. However the still identifiable gender and race of certain individuals Figure 1: Outputs of the proposed filter contributed to the perception of lower levels of privacy of the person in the case of some of the video frames. Face filter (H): The face being a highly identity revealing region has been ranked high (H) for privacy protection compared to other Categories and Subject Descriptors body parts. Accordingly, we sought to ensure the anonymity of I.2.10 [Artificial Intelligence]: Vision and Scene Understanding - the visible face. First an empirical threshold was set to examine video analysis, representations, data structure, and transforms the minimum size of the face in which it could be identifiable (i.e. 50 pixels in our case, rightmost snapshot in Figure 1). Below the 1. INTRODUCTION said threshold, a simple median blur filter proved sufficient to The Visual Privacy Task at MediaEval 2014 [1] acknowledged protect the person’s identity. Once the face size exceeded the that the perceived privacy of a citizen cannot be divorced from the threshold, a key point detector was applied on the face region contexts in which the privacy is valued by the citizen and thus followed by adaptive colour quantisation and circle texturing to worth protecting. A citizen may have a variety of roles, produce the effect as shown in Figure 1. responsibilities and relationships; each associated with a particular persona which may be activated in a given context Skin filter (M): The exposed skin regions could provide sufficient within their everyday life-style e.g. husband, father, employee, information to enable the detection of the ethnicity of the subject. boss etc. Each such persona of the citizen is commensurate with a Skin provides a focus of attention. Exposed skin regions, e.g. particular privacy boundary linked to a certain context. This could hands, are also important in activity recognition and detection of guide the levels and scopes of privacy filtering according to the weapons. Therefore, morphologic changes or non-homogenous situated (context-dependent) scenario. The VPT task evaluation colour changes are not suitable in this case. To manipulate the methodology has responded to the need for a more inclusive, skin region in a unified fashion, we reduced the colour saturation holistic and high resolution assessment of privacy filtering and luminance values in the R, G, and B colour channels requirements as well as the evaluation of the efficacy and impacts separately within the oriented boxes enclosing the skin regions. of the resulting privacy filtering solutions based on the UI-REF The filtered skin regions were still recognisable. There are only methodology [2]. The PEViD dataset [3] was updated with two identifiable skin colours: light-like and dark-like. The skin privacy ranking system for the subject body parts for context- texture which is responsible for the skin attractiveness was also aware impact assessment of privacy protection solutions. eliminated. 2. THE PROPOSED FILTER Person filter (L): The last stage of privacy protection is applied to We proposed a privacy filter which primarily aims to achieve a the bounding box enclosing the subject region. This region is balance in the well-addressed Privacy-Intelligibility trade-off. In ranked low in the provided annotation. An edge-based analysis is addition, the Pleasantness and the appropriate filter application implemented on the foreground region(s) within the subject criteria are also considered alongside the real-time applicability. bounding box. The procedure begins with morphological Accordingly our three different filtering techniques for face, skin, operations to enhance the foreground mask to reduce the and body regions of the subject featured in the video were applied background noise and minimise the holes in the foreground as follows: region. Canny edge detector is subsequently applied and further refined to eventually draw the subject’s contour. The final effect Copyright is held by the author/owner(s). produced as depicted in Figure 1 is the result of a distance MediaEval 2014 Workshop, October16-17, 2014, Barcelona, Spain transformation which calculates the distance of each pixel of the resulting binary contour map with the closest zero pixel in the appearance. On the other hand, the proposed solution would image. OpenCV implementation based on [4] was used which clearly prevent the viewer from being able to identify the featured calculate the Euclidean distance to the nearest zero pixel subject in the normal cases by successfully hide most of the face consisting of basic shifts: horizontal, vertical, diagonal, or details. Regarding the Pleasantness criterion the obtained scores knight’s move. A mask of size (5 X 5) was used for the best were comparable to the median values for the three streams which results. fell within the range of 50-70% for stream 2 and 3 and noticeably lower for stream 1 for slightly above 20%. Table 1 summarises the 3. EVALUATION RESULTS numerical values of the obtained scores for the evaluated streams. A subjective evaluation consist of three (3) streams were conducted and the performance of the proposed privacy protection Intelligibility Privacy Pleasantness solution is examined in terms of the defined criteria namely, Privacy, Intelligibility, and Pleasantness as described in [1]. Stream_1 75.10% 42.80% 23.50% Figure 2-5 illustrate the performance of the proposed solution in Stream_2 84.07% 35.55% 69.27% the three evaluation streams respectively. A noticeable trend can Stream_3 73.27% 39.01% 56.61% be generalised from the three sets of results with only marginal Table 1: Scores for evaluated streams variations. Figure 2: Scores from Stream 1 Figure 5: Integrating the scores for all three streams 4. CONCLUSION In this paper we have proposed a video privacy filter using a combination of filtering techniques to simulate a context-aware solution. The filter aims to achieve the highest privacy with minimum content distortion and viewer distraction. The obtained scores were comparable to the average values of the scores for all the privacy filtering solutions as proposed for the Visual Privacy Task 2014. One possible future work would be the addition of Figure 3: Scores from Stream 2 person re-identification test to be included in evaluating the Privacy criterion as an important aspect. 5. ACKNOWLEDGMENTS We would like to thank Lucas Teixeira and Kevin Lelu for their valuable inputs. This work was supported by the European Commission under contracts FP7-261743 VideoSense project. 6. REFERENCES [1] Badii, Atta, Ebrahimi, Touradj, Fedorczak, Christian, Korshunov, Pavel, Piatrik, Tomas, Eiselein, Volker, Al-Obaidi, Ahmed “Overview of MediaEval 2014 Visual Privacy Protection Task”, Figure 4: Scores from Stream 3 Proceedings of the MediaEval 2014 Workshop, Barcelona, Spain, 16-17 October 2014 The proposed solution scored above the average for Intelligibility [2] Badii, Atta, “User-Intimate Requirements Hierarchy Resolution criterion and well above the region of 70% in the three streams Framework (UI-REF): Methodology for Capturing Ambient Assisted which ensures that the processed video will still serve the main Living Needs”, Proceedings of the Research Workshop, Int. Ambient Intelligence Systems Conference (AmI’08), Nuremberg, purpose of CCTV security objective. Germany, November 2008 However, the Privacy scores were slightly below the median [3] Korshunov, Pavel, and Ebrahimi, Touradj. “PEViD: privacy which is in general below the value of 50% in this competition. evaluation video dataset”, Applications of Digital Image Processing One possible explanation of the overall low score of Privacy is the XXXVI, San Diego, California, USA, August 25-29, 2013. fact that this criterion has been measured based on the ability to [4] Borgefors, Gunilla. “Distance transformations in digital images” identify the gender and the ethnicity of the person which could be Computer vision, graphics, and image processing 34.3 (1986): 344- hard to conceal without significantly manipulating the person 371.