Overview of th e MediaEval 2013 Visu al Privacy T ask

Abstract

AB STRACT T his paper describes t he Visual Privacy T ask (VPT ) 2013, it s scope and object ives, relat ed dataset and evaluat ion approach.

2. THE VPT 2013 DATASET

T he dat a set consist s of videos collect ed from a range of st andard and high resolut ion cameras and cont ains clips of different scenarios showing one or several persons walking or int eract ing in front of t he cameras. People may also carry specific it ems which could pot ent ially reveal t heir ident it y and may t herefore need t o be filt ered appropriat ely. For t his year, people can carry backpacks, umbrellas, wear scarves, and can be seen fight ing, pickpocket ing or simply walking around. People may be at a dist ance from t he camera or near t he camera, making t heir faces vary considerably in pixel size and qualit y. T he videos have variable ambient light ing wit h half of t he clips recorded at night . T he dat aset cont ains 22 video clips and associat ed annot at ions in xml form. T he videos include indoor, out door, day-t ime and night -t ime environment s, showing people int eract ing or performing various act ions. T he clips are in t he mpeg format wit h a resolut ion of 1920x1080 pixels at 25 frames per second. Publicat ions arising from experiment s performed using PEViD must acknowledge it s publishers [ 1 ]. Copyright is held by t he aut hor/owner(s).

3. VISUAL PRIVACY TASK

T his t ask explores how image processing, comput er vision and scrambling t echniques can deliver t echnological solut ions t o some visual privacy problems [2] [ 3 ] [4]. T he goal of privacy prot ect ion is t o prevent pot ent ial access t o informat ion, t he divulgement of which can amount t o a (perceived) int rusion of an individual’s privacy. T he ext ent of such a (perceived) loss of privacy depends on t he individual as well as t he cont ext and as such can only be det ermined by reference t o t he user (“ dat asubject ”) in each case. Cont ext -specific privacy prot ect ion const it ut es an int erest ing ext ension of t his VPT t ask which is planned t o be included in fut ure challenges. T he goal of t his VPT is t o propose met hods whereby persons feat ured in digit al imagery can be obscured so as t o render t hem unrecognisable. Privacy level variat ions may also be t riggered by det ect ed anomalies, crit ical event s, and alert s et c. or be based on prior official permission grant ed by higher aut horit ies t o suspend t he masking of t he ident it y of an individual in specific cases. Since t he result ing part ially obscured videos would st ill have t o convey some video informat ion t o be wort h viewing, an opt imal balance should be st ruck so t hat despit e t he ext ent of such masking of t he facial ident it y as may be necessary, t he cat egorical ident it y of any masked act ors e.g. humans can st ill be recognisable t o t he viewer. T hus ident it y obscuring t echniques should not result in art efact s t hat are ‘socially inappropriat e/offensive’ and unaccept able t o t he human users. T he part icipant s should also demonst rat e t hat t heir choice of obscuring t echnique is such t hat t he result ing obscured (e.g. pixelat ed) faces do not t end t o fixat e a viewers’ at t ent ion t hus dist ract ing t he viewer and/or adversely impact ing t he accept abilit y-usabilit y of any obscured/scrambled images, from t he perspect ive of bot h t he dat a-subject as well as ot her viewers. Part icipant s are provided wit h videos cont aining faces from different camera angles. T he ground t rut h consist s of annot at ions of persons’ images, including face, hair, visible skin regions, as well as t heir personal accessories.

3.1 Obje ctive me trics

T he object ive met rics are comput ed aut omat ically wit h a mixt ure of object det ect ion and mat ching in order t o evaluat e t he impact of t he filt ering on t he privacy and int elligibilit y. Some addit ional image qualit y measures will be t aken int o account in order t o give credit t o filt ers result ing in visuallypleasant masking.

3.1.1 Face Detection

A face det ect ion algorit hm will be run on t he obscured videos submit t ed for t he evaluat ion using t he Viola-Jones face det ect ion from OpenCV library. Ideally, no faces should be found, since t hey all should be obscured. T he faces found by t he face det ect ion algorit hm are mat ched against t he ground t rut h t o avoid including t he false posit ives of t he det ect ion algorit hm.

3.1.2 Object Tracking

T he int elligibilit y is measured by applying t he Hist ogram of Orient ed Gradient as a human det ect or t aking t he video images as input . Successful det ections of a human means t hat even alt hough t he sensit ive areas may have been obscured, t he result ing video could st ill carry sufficient visible clues for Video Analyt ics including t racking. T hese det ect ions are compared against t he det ect ions from t he raw video.

3.1.3 Person Re-identif ication

A visual model of t he un-filt ered images of persons as feat ured in t he video set will be developed and mat ched against t he privacyfilt ered versions of t he images of t he same persons as select ed from t he submission set. T he mat ching process will be implement ed in t wo ways so as t o provide t he basis for a Merit Crit erion Framework for Privacy Impact Assessment based on Efficacy, Consist ency, Disambiguit y and Int elligibilit y PIAF[5]; as follows: i ) by building a visual model from t he original unfilt ered image in each case and t hen at t empt ing t o mat ch t his against t he respect ive filt ered image, and, i i ) by building t he model from t he filt ered image and at t empt ing t o mat ch it against t he respect ive unfilt ered original set . A low reident ificat ion score arising from t he above mat ching cycles would indicat e a higher Efficacy privacy prot ect ion afforded by t he privacy filt ering t echniques as deployed in each case. Consistency, Disam biguity and Intelligibility propert ies of t he deployed Privacy Filt ering approach will also be assessed by comparing t he filt ered visual model t o t he filt ered inst ances of t he t arget person(s) in t he image set . A high score would indicat e t hat t he filt ered video st ill carries sufficient informat ion t o enable an observer t o perform t asks such as person t racking across images from t he CCT V net work wit hout finding out t he person’s ident it y. T he framework has been ext ended t o enable t he video-cont ext -sensit ive t hresholding of t he Merit Crit eria. T his provides a powerful benchmarking mechanism for t he spect rum of possible privacy filt ering t echniques, in t erms of t heir opt imisat ion of t he t rade-offs (ident it y maskabilit y /t rackabilit y) across t he specific crit eria t o suit t he object ives of t he video processing wit h privacy prot ect ion and surveillance by best balancing t he result ing Efficacy, Consist ency, Disambiguit y and Int elligebilit y impact s of part icular privacy filt ering t echniques as deployed in arbit rary situated video-cont ext s and UI-REF based privacy requirement s.

3.1.4 Metric f or Visual Appropriateness

Obscuring of t he image of persons and t heir accessories will be evaluat ed using SSIM and PSNR met rics for image qualit y based on t he human eye percept ion of salience in t he image. A successful privacy filt ering syst em should have a minimal impact on t he global qualit y of t he image wit h modificat ions occurring only on t he sensit ive areas which should be t hus anonymised.

3.2 Us e r Study for As s e s s me nt of

Appropriate ne s s of Vis ual Privacy Filte ring

A random subset of videos from t he submit t ed runs will also be evaluat ed t hrough a user st udy aimed at developing a deeper underst anding of user percept ions of appropriat eness in t erms of UI-REF based privacy prot ect ion requirement s. T his subject ive evaluat ion will t ake int o account t hree main aspect s of any obscured (element of) image, namely int elligibilit y, privacy, and appropriat eness. In t he cont ext of surveillance scenarios, quest ions relat ed t o whet her a person wears personal it ems t hat can be used for ident ificat ion e.g. (branded) backpack, scarf, et c, will be considered as relevant t o privacy and int elligibilit y.

Fi gure 1. Sampl e frame from the VPT Data Se t [ 1 ] T he visual appropriat eness of t he obscured images will be evaluat ed based on t he various aspect s such as pleasant ness, dist ract ion, and user accept ance for video surveillance, et c. T he visual appropriat eness crit erion will essent ially follow a UI-REF based evaluat ion met hodology [6]. T his met ricat es: i ) t he cat egoric “ recognisabilit y” of an obscured image as a member of a part icular species, and, i i ) t he obscuring Effects, and SideEffects on t he percept ion of t he image by a viewer, and, i i i ) t he ext ent of any result ing negat ive or posit ive emot ions or Affects or dist ract ion in t he mind of t he viewer of an image t hat has been subject ed t o such obscuring (indignit y/st igma). Insight s from t his user st udy will serve as a baseline for refining t he met rics and shall inform t he design of t he fut ure privacy t asks.

4. ACKNOWLEDGMENTS

T his Visual Privacy MediaEval t ask was support ed by t he European Commission under cont ract s FP7-261743 VideoSense. [2] Dufaux, F. & Ebrahimi, T ., “ Scrambling for Privacy Prot ect ion in Video Surveillance Syst ems,” IEEE T ransact ion on Circuit s and Syst ems for Video T echnology, Vol. 18, Nr. 8 (2008), p. 1168-1174 [4] Senior, A., “ Privacy Prot ect ion in a Video Surveillance Syst em,” Privacy Prot ect ion in Video Surveillance, Springer, 2009

References

[1] Korshunov P. & Ebrahimi T . , “ PEViD: privacy evaluat ion video dat aset ” . Applicat ions of Digit al Image Processing XXXVI , San Diego, California, USA, August 25- 29 , 2013 .

[3] Dufaux , F. & Ebrahimi , T . , “ A framework for t he validat ion of privacy prot ect ion solut ions in video surveillance ,” 2010 IEEE Int ernat ional Conference on Mult imedia and Expo (ICME) , pp. 66 - 71 , 19 - 23 July 2010 .