Facial Recognition of Identical Twins

Transcription

1 Faal Recgnitin f Identical Twins Matthew T. Pruitt, Jasn M. Grant, Jeffrey R. Pane, Patrick J. Flynn University f tre Dae tre Dae, I {pruitt, jgrant3, jpane, flynn}@nd.edu Richard W. Vrder Bruegge Digital Evidence Labratry, Federal Bureau f Investigatin Quantic, VA Richard.VrderBruegge@ic.fbi.gv Abstract Bietric identificatin systes ust be able t distinguish between individuals even in situatins where the bietric signature ay be siilar, such as in the case f identical twins. This paper presents experients dne in faal recgnitin using data fr a set f iages f twins. This wrk establishes the current state f faal recgnitin in regards t twins and the accuracy f current state-f-theart prgras in distinguishing between identical twins using three ceral face atchers, gnitec 3.2., VeriLk 4., and PittPatt and a baseline atcher eplying Lcal Regin PA. Overall, it was bserved that gnitec had the best perfrance. All atchers, hwever, saw degradatin in perfrance cpared t an experient where the ability t distinguish unrelated persns was assessed. In particular, lighting and expressin seeed t have affected perfrance the st. 1. Intrductin Bietric signatures are dependent n the assuptin that each individual is unique. Identical twins can have bietric signatures that are very siilar, espeally when the signature is derived fr a face iage. While face recgnitin sftware exhibited pr perfrance in this cntext, there are ther bietric dalities that can ffer a perfrance increase at the cst f increased invasiveness. Jain and Prabhakar [6] shwed that the false accept rate fr an identical twin recgnitin task was fund t be 2-6% higher than fr a dataset with a nral twin-nntwin distributin. The priary tivatin f this wrk is t assess the accuracy f current generatin faal recgnitin systes n a particularly challenging data set, cntaining twins. In particular, we eply PittPatt [11], gnitec 3.2. [3], and VeriLk 4. [7] as representative f the state f the art faal recgnitin systes. We als cpare these ceral atchers against the LRPA atcher [1] as a baseline ethd. This paper is rganized as fllws. Sectin 2 discusses related wrk in the perfrance f face recgnitin sftware n twins. Sectin 3 utlines the data set used and the paraeters f the experients perfred in this paper. In Sectin 4 reviews the results f the experients perfred. Lastly, Sectins 5 and 6 discuss the results and future wrk. 2. Backgrund The prble with discriinatin f identical twins using faal recgnitin has been nted. Sun et al. [12] used the ASIA Multidal Bietrics Database f Twins cntaining the 134 subjects, including bth identical and nnidentical twins, t perfr recgnitin experients. Using the gnitec FaceVAS syste, a true accept rate f apprxiately 9% at a false accept rate greater than 1% was btained. In cntrast t the wrk f Sun et al. [2], this study takes ultiple experiental variables int cnsideratin when characterizing perfrance. The reslutin f the iages used here is uch higher, but the changes t values f faal pse variables (pitch, rll, yaw) are nt cnsidered. This prble has als been explred by Phillips et al. in [9]. Spefically, sae day and crss year perfrance were explred using the tp three subissins fr the Multiple Bietric Evaluatin (MBE). The naes f the algriths, hwever, were nt given in the paper. The ain cnclusin f this paper was that the best perfrance cae when all iages were taken n the sae day with a neutral expressin and cntrlled lighting Matchers LRPA Prinpal pnent Analysis (PA) fr face recgnitin, as first described by Turk and Pentland in [l3] represents face iages as pints in a high diensinal space deterined via training. T perfr a atch, ne can prject a prbe face int the space and take the label f the clsest /11/$ IEEE

2 prjectin f an existing face as the identity f the prjected prbe. The Lcal Regin Prinpal pnent Analysis (LR PA) is a faal recgnitin algrith develped by lrad State University which initially was develped as a baseline algrith fr the Gd, Bad, and the Ugly Face Recgnitin hallenge Prble. This PA apprach reduces the face t 128 by 128 pixels. ext, it segents the face regin int 13 subregins including an eye and tw eyebrw regins (fr bth the left and right eye); an upper, iddle, lwer left, and lwer right regins f the nse; and the left, iddle, and right regins f the uth. These regins are nralized t attenuate variatin in illuinatin. PA is perfred n these 13 subregins and the entire face blck and scre fusin is used t btain an identity label PittPatt Recgnitin experients were perfred using the Pittsburgh Pattern Recgnitin syste knwn as PittPatt [11]. PittPatt is ptiized t wrk n lw reslutin iage data. It perfrs recgnitin in three stages at varying sizes (12, 2, and 25 pixels between the eyes). Texture infratin is nt typically available at these reslutins. In versin 2 f the Multiple Bietrics Grand hallenge (MBG), PiuPau was fund t have a true accept rate f apprxiately 9% at false accept rates as lw as.1 % using cntrlled lighting and iages as sall as 9 by 9 pixels [8]. On a large dataset f 1,6, iages in the MBE [4], PittPatt yielded a rank ne identificatin rate f nly 62%, which was lwer than st ther perfrers Verilk 4. The VeriLk Sftware Develper's Kit is a set f faal recgnitin tls develped by eurtechnlgy. eurtechnlgy clais that their sftware assures syste perfrance and reliability by ipleenting live face detectin, siultaneus ultiple face recgnitin, and fast face atching in ne-t-ne and ne-t-any des [7]. The algrith is able t perfr siultaneus ultiple face prcessing in a single frae and uses live face detectin. At its wrst, Veri Lk btained a False Reject Rate f 3.21 % at a False Accept Rate f.1 % in the Face Recgnitin Grand hallenge (FRG) evaluatin FaceVAS 3.2. gnitec is a cerally available face recgnitin syste aking use f a Face Visual Access ntrl Syste (FaceVAS). gnitec has cpeted in several face recgnitin challenges. The general ai f gnitec is t perfr effiently n large-scale databases with high perfrance. gnitec sftware achieved a rank ne identificatin rate f 83% n a dataset f 1,6, iages in the MBE and n an even larger dataset f 1,8, iages btained a rank ne identificatin rate f 88%. Outside I "bi.al l "bi a I I D91 1 D9 I Inside Figure 1. The 29 setup. Frntal iages were taken bth indrs and utdrs using all fur caeras. Inside!$> Subject D91 ",.. Figure 2. The 21 setup. Iages were taken indrs fr a single caera. 3. Experients 3.1. Data Set A dataset cntaining iages f twins was used. Iages f subjects were acquired at the Twins Day festival [1] in 29 using a three caera set up as seen in Fig. 1. The sae iages were acquired bth inside and utside. Tw frntal face phts were acquired fr tw separate caeras. The subject wuld lk at caera three, then rtate t caera ne. This wuld cause differences in the sensr and lighting n the face. In 21, a slightly different setup was used at the sae festival. A five caera setup as seen in Fig. 2 was used with data taken indrs nly. There were tw frntal iages taken in 21, hwever, they were taken fr the sae sensr within secnds f each ther. Subjects were instructed t stand up, turn arund, and sit back dwn after the first pht was taken in rder t take a secnd pht. Fr this data set, nly the frntal iages, crrespnding t a yaw f zer degrees were used. Due t the way the data was cllected, galleries taken fr different caeras culd be used. Fr 29 a set f 5,8 iages was used fr this experient. 1,635 iages fr 21 were used as well. The iages were cnverted t PPM frat fr the riginal ikn RAW iage frat (EF) taken using ikn D9 caeras and had a reslutin f 2848 cluns and 4288 rws Using the atchers PittPatt PittPatt first creates sets f prbe and gallery teplates. This step takes the lngest t cplete. Iages are laded and PittPatt teplates are generated, then added t a single gallery r prbe file. After gallery and prbe file creatin is

3 cplete, the prbe and gallery teplates are cpared t each ther in an "all versus all" fashin t prduce a siilarity atrix, which takes less than a inute VeriLk T perfr atching experients, a eurtechnlgy prprietary teplate was created fr each enrlled iage. After creating a teplate fr each iage, a single teplate was atched against all teplates in the gallery. There was n distinctin between teplates used as prbes r teplates used in the gallery. The atching experient returns a atrix f siilarity scres which ranged fr t 18, with higher values crrespnding t higher prbable atches gnitec gnitec first detects the eyes and then face lcatins in the iage. The iage with eye psitins akes up the Priary Faal Data. Once a face is fund, the iage is nralized and standard preprcessing techniques are applied. Se f the preprcessing techniques include histgra equalizatin and intensity nralizatin. Features f the face are then extracted and stred as a vectr in high-diensinal space. This vectr represents the Secndary Faal Data. The siilarity between tw faces is then the distance between the tw crrespnding vectrs. A crrect verificatin can be ade if the distance scre exceeds a given threshld LRPA THe LRPA algrith was executed in the partially autated de; therefre, we preprcessed the set f iages t deterine the hrizntal and vertical psitins f each pairs f eyes in the set. Signature sets were generated, which cntained the list f iages, their lcatins, and filetypes. Tgether, the signature sets and the table f eye lcatins, were passed t LRPA which perfred the rest f the atching experient internally. Perfect atches received a scre f 1 and the lwest atch scres averaged apprxiately Variables In these experients, there are three variables. Fr lighting, cntrlled lighting iages were acquired inside a tent using studi laps. Fr uncntrlled lighting, iages were taken utside. Eyeglasses were als used in the data acquisitin phase. These were a set f zer-prescriptin glasses used t siulate nral eyewear. Subjects were prpted t express either a neutral expressin r a sile. Exaples f the different iages used in this experient can be seen in Fig. 3. Fr the graphs in this paper, the legends refer t the systes by their syste ID. Table 1 shws the translatin between syste id and syste nae. Figure 3. Iages representing different cvariate cnditins acquired at Twins Days fr a subject. (a) Inside with eutral Expressin (b) Inside with Sile (c) Inside with eutral Expressin and n-prescriptin Glasses (d) Outside with eutral Expressin (e) Outside with Sile (f) Outside with Sile and n-prescriptin Glasses 4. Results Syste ID Syste ae LRPA VeriLk 4. gnitec 3.2. PittPatt Table 1. Matchers and Syste IDs In these experients, three cvariates are exained, naely expressin, lighting, and glasses. Fr the baseline experients, the expressin is neutral, the lighting is cntrlled, and the subject is nt wearing any kind f glasses. These crrespnd t ideal cnditins fr atching. Btstrapping was perfred by by randly sapling the atch and nn-atch scres separately t get a set 8% f the riginal size with the sae ix f atch and nn-atch scres. After enrllent, an all versus all experient was perfred n all fur atching systes. t all atchers were able t enrll every iage leading in a different nuber f entries in a siilarity atrix fr each atcher. All atchers were syetric, s nly the upper right half f the siilarity atrix was stred Baseline Perfrance Perfrance results are given using standard RO plts with bx and whisker plts staged at even intervals alng the curve. Here, the whiskers represent a 95% cnfidence interval fr the True Accept Rate at that False Accept Rate and

4 29 Perfrance Prbe: Baseline Gallery: Baseline 21 Perfrance Prbe: Baseline Gallery: Baseline r Ei.." < "! :!a.. 8.." < " jl.... : Figure 4. Perfrance f all atchers under the baseline cnditins. Syste Prbe Gallery Rate gnitec Baseline Baseline.3 LRPA Baseline Baseline.43 VeriLk Baseline Baseline.5 PiuPaU Baseline Baseline.24 Table Baseline EER the tp and btt f the bx represent the upper and lwer quartiles repectively. Fr the RO curve in Fig. 4, we can iediately see that bth gnitec and VeriLk dinate PittPatt and LRPA by a significant argin with bth VeriLk and gnitec being very clse in perfrance. The whiskers f the bx plts verlap between the gnitec and VeriLk curves, hwever, indicating that there is n statistically significant difference between the perfrance f these atchers. The cnfidence intervals fr PittPatt and LRPA are significantly different; their cnfidence intervals are wider, thugh, indicating re variability and subsequently less reliability in the perfrance. T get a better estiate f f hw well separated these curves are, it is benefial t lk at the equal errr rates as well. Lking at Table 2, it can be seen that gnitec has a saller equal errr rate by a sall argin. netheless, bth gnitec and VeriLk are still very clse t perfect. On the ther hand, PittPatt and LRPA bth fall shrt here. A scre f.43 fr LRPA indicates that LRPA is perfring just slightly better than rand chance. The cnfidence bands fr LRPA actually intersect t the rand line in se places. One reasn PittPatt ay be perfring s prly when cpared t the ther atchers is that it is ptiized fr saller faces. Using such large iages ay play a part Figure 5. This RO shw the perfrance f all atchers under the baseline cnditins. As can be seen in the graph, all atchers saw a perfrance bst ver the 29 data. This ay be due t the face that the prbe and gallery iages fr 2 1 were taken within secnds f each ther with the sae caera and in the sae psitin. Syste Prbe Gallery Rate gnitec Baseline Baseline.1 LRPA Baseline Baseline.31 VeriLk Baseline Baseline.2 PittPatt Baseline Baseline.5 Table Baseline EER in the pr perfrance n identical twins. gnitec and VeriLk ay be able t use the extra texture in the large nuber f pixels acrss the face. We assued that 21 baseline data wuld fllw the sae trend as was seen in the 29 baseline data. This is nt the case. As can be seen in Fig. 5, all atchers saw a significant perfrance bst ver the 29 results. The cnfidence bands indicate n statistically significant difference between all three ceral atchers, perfect perfrance being within the cnfidence intervals. Even LRPA sees a significant iprveent in perfrance. VeriLk, gnitec, and PittPatt perfr nearly perfect in this case. A reasn fr the iprveent ay ce fr the way the caeras were set up at Twins Days. The tw frntal iages that are being cpared in 21 are taken fr the sae caera at the sae psitin within secnds f each ther and are likely t be re siilar than iages in the 29 data where subjects had their secnd frntal iage taken with a different sensr at a different relative psitin. This phenenn can be seen in all RO curves generated fr 29 data.

5 . ''; r < " 29 Perfrance Prbe: eutral Expressin Gallery: Sile -.;*;+ - - _ ,. _.r-., Z ---- _.; ""'T" _ -....L. ; -....L { i. l I. i f Syste Prbe Gallery Rate gnitec eutral Expressin Sile.5 LRPA eutral Expressin Sile.39 VeriLk eutral Expressin Sile.16 PittPatt eutral Expressin Sile.19 Table eutral Expressin vs. Sile EER 29 Perfrance Prbe: Sile Gallery: Sile r..2 Figure 6. These RO curves shw the perfrance f the fur atchers fr 29 data where the prbe has a neutral expressin while the gallery has a sile. There is degradatin in perfrance acrss all fur atchers fr the 29 baseline.. 1 < 6 I..., : ;EB '., 4.2. Expressin The expressin cvariate variable's perfrance ipact was easured under cntrlled indr lighting where n glasses are wrn. A change in expressin can cause different wrinkles and textures t appear in the face that ight have nt been apparent in the baseline iages. The expressin experients take tw factrs int accunt. The first is where the prbe iage has a neutral expressin and the gallery iage has a siling face. This cnditin set represents the st difference between prbe and gallery. The RO curve in Fig. 6 shws that the perfrance f gnitec is statistically significantly better than all ther atchers. VeriLk perfrs better than PittPatt when FAR is less than.2, well within a nral perating range. paring this infratin with the baseline fr 29, the edians f all fur curves has been shifted dwn indicating wrse perfrance. VeriLk saw a statistically significant decrease in perfrance fr the 29 baseline data. This is again supprted by by the EER as seen in Table 4. The EER f Veri Lk is significantly higher than that f the EER fr VeriLk with the 29 data while the ther atchers are relatively unaffected. When bth the prbe and gallery faces are siling, it can be seen that bth VeriLk and gnitec perfr statistically significantly better than PittPatt and LRPA, even thugh there is n statistically significant difference between the perfrance f VeriLk and PittPatt. The statistically significant difference in perfrance fr Veri Lk during crss-expressin atching indicates that expressin..2 Figure 7. These RO curves shw the perfrance f the fur atchers fr 29 data where the prbe hand gallery bth are siling. This shws better perfrance fr cparisns where bth prbe and gallery are f the sae expressin rather than different expressin. Syste Prbe Gallery Rate gnitec Sile Sile.3 LRPA Sile Sile.42 VeriLk Sile Sile.6 PittPatt Sile Sile.18 Table Sile VS. Sile EER change can ptentially degrade perfrance. The sae cannt be said when bth prbe and gallery have the sae expressin, regardless f whether r nt that expressin is neutral r a sile. This wuld suggest that fr best atching results, the gallery shuld cntain saples f ultiple expressins, r at least trained with data f the expected expressin fr atching sessins. The perfrance n the 29 data where bth prbe and gallery were siling did nt shw as uch degradatin as the perfrance f crss-expressin atching. The EER, seen in Table 5 shws a siilarity t the baseline data fr 29. ne f the RO curves, seen in Fig. 7, have any statistically significant difference fr the RO curves fr baseline 29 experients.

6 While the 21 data, where prbe and gallery had different expressins, shws the sae trend as seen in the 21 baseline data, it can be seen that there is still a statistically significant difference in the perfrance f all three atchers, seen in Fig. There was n statistically significant difference fr the 21 data where bth prbe and gallery were siling fr all ceral atchers fr the 21 baseline. This is particularly interesting cnsidering the bservatin that this data shuld be inherently siilar due t the cnditins under which the data was acquired. 21 Perfrance Prbe: eutral Expressin Gallery: Sile '. ' ';!? r v l, ; -:- : ka,., I"! ;-'- 29 Perfrance Prbe: ntrlled Lighting Gallery: Uncntrlled Lighting r.&!? v..2 Figure 9. These RO curves shw the perfrance n 29 data where the prbe was taken under cntrlled lighting and the gallery was taken under uncntrlled lighting. There is nly slight degradatin fr the baseline fr the sae year...2 Syste Prbe Gallery Rate gnitec ntrlled Uncntrlled.3 LRPA ntrlled Uncntrlled.43 VeriLk ntrlled Uncntrlled.5 PittPatt ntrlled Uncntrlled.2 Figure These RO curves shw the perfrance f the fur atchers fr 21 data where the prbe has a neutral expressin while the gallery has a sile Lighting Fr the lighting cvariate analysis, subjects had iages taken bth indrs under studi lighting and utdrs with uncntrlled lighting. In this instance, uncntrlled lighting can range anywhere fr direct sunlight t the aunt f light visible thrugh the cluds n a rainy day. As the setup between 29 and 21 changed, utdr iages were nly acquired in 29. This eans, unfrtunately, that nly an analysis f the 29 data was pssible fr the lighting cvariate. Just like the expressin cvariate, thugh, we have iages where prbe and gallery were taken under different lighting schees and iages where bth prbe and gallery were taken under uncntrlled lighting schees. Lking at the RO curves fr 29 data, fr Fig. 9, it can be seen that gnitec and VeriLk perfr well as cpared t LRPA and PittPatt. In the case f lighting, gnitec actually perfrs statistically significantly better than VeriLk with lw variability in bth f the curves. paring this data t the baseline, VeriLk did nt perfr statistically significantly wrse than VeriLk fr Table ntrlled Lighting vs. Uncntrlled Lighting EER the 29 baseline, even thugh it was statistically significantly wrse that gnitec fr this experient. In fact, n curve perfred statistically significantly wrse than its crrespnding curve fr the 29 baseline, althugh there was re variability. In Table 6, it can be seen that gnitec, VeriLk, and LRPA all have the sae EER as their 29 baseline cunterpart. Lking back at the perfrance fr the crssexpressin experient with 29 data, crss lighting des nt see t effect atcher perfrance as uch as crssexpressin atching. VeriLk is statistically significantly better fr crss-lighting. The rest f the curves are nt statistically significantly different. paring the perfrance f crss-lighting cnditins t sae lighting cnditins where bth prbe and gallery were taken under uncntrlled lighting, we can see that it is better t have at least ne iage, prbe r gallery, be taken under cntrlled lighting. Perfrance fr all ceral atchers saw a statistically significant drp in perfrance fr bth the crss-lighting experient and the baseline perfrance fr 29 as can be seen in Fig. 1. The EER fr all ceral atchers was greater acrss the bard as well. The verlap f scres distributins was

7 29 Perfrance Prbe: Uncntrlled Lighting Gallery: Uncntrlled Lighting 29 Perfrance Prbe: Glasses Gallery: n-prescriptin Glasses B 8 : Figure 1. These RO curves shw the perfrance n 29 data where bth prbe and gallery were taken under uncntrlled lighting. A degradatin in perfrance can be seen fr bth the baseline fr 29 and the situatin where at least ne iage was taken under cntrlled lighting. Syste Prbe Gallery Rate gnitec Uncntrlled Uncntrlled.12 LRPA Uncntrlled Uncntrlled.49 VeriLk Uncntrlled Uncntrlled.14 PittPatt Uncntrlled Uncntrlled.33 Table Uncntrlled VS. Uncntrlled EER greater as well, espeally fr LRPA and PittPatt Eyewear Zer-Prescriptin Glasses These results are presented under cntrlled indr lighting cnditins where the subject is wearing a neutral expressin. While the effect f eyewear has been studied in ther papers, it has nt been deterined whether it is the fraes f the glasses that cause a change in perfrance r the change in intercular distance assated with prescriptin lenses. The glasses used here were designed t reduce the visibility f the fraes t eludate the effect f intercular distance. As can be seen in Fig. 11, we see gnitec and VeriLk utperfring PittPatt and LRPA. While gnitec and Veri Lk are nt statistically significantly different, PittPatt's perfrance is statistically significantly wrse than the ther tw. The interesting fact abut this graph ces when we cpare the perfrance with the 29 baseline perfrance. While gnitec and VeriLk are nt statistically significantly different that the perfrance in the 29 baseline, Figure 11. These RO curves shw perfrance n 29 data fr a prbe that had n glasses and a gallery wearing nn-prescriptin glasses. There is inial perfrance degradatin fr the baseline. Syste Prbe Gallery Rate gnitec Glasses Glasses.2 LRPA Glasses Glasses.4 VeriLk Glasses Glasses.4 PittPatt Glasses Glasses.l7 Table 29 Glasses vs. n Prescriptin-Glasses EER PittPatt actually perfrs better here by a statistically significant argin. As was seen with the ther cvariates, such as expressin r lighting, intuitin wuld see t suggest deviating fr the baseline wuld cause perfrance t be less than r equal t the baseline perfrance. This result, hwever, shws the cntrary. When ne iage, prbe r gallery, has an iage with glasses, the perfrance sees t be better. While this ay see better, there culd be ther explanatins fr this result. Glasses can cause an increase in failures t enrll, but we have nt been able t cnfir that in this experient. If this effect did ccur, then there wuld be less data in this experient than there was in the baseline, which wuld cause a statistical irregularity. As a result, it reains t be deterined if glasses can iprve the perfrance f the PittPatt algrith. When we lk at the perfrance where bth prbe and gallery are wearing nn-prescriptin glasses, seen in Fig. 12, the scres are nt statistically significantly different fr the 29 baseline nr significantly different fr the crss-glasses experient. The results reprted abve which shw little r n degradatin in perfrance with eyeglasses cntradict earlier

8 r 29 Perfrance Prbe: Glasses Gallery: Glasses still insuffient t perfr a statistically significant cparisn t see whether r nt aging can be a factr in the atching phase. Fr future papers, having re crss year data can be used t see the effect f aging n twins recgnitin which ay turn ut t be significant due t the fact that as twins age, their features are re influenced by natural prcesses leading t different features fr each twin... " < "..2 Figure 12. These RO curves shw perfrance n 29 data where bth prbe and gallery were wearing nn-prescriptin glasses. There is slightly re degradatin in perfrance fr where at least ne iage did nt wear glasses. Syste Prbe Gallery Rate gnitec Glasses Glasses.3 LRPA Glasses Glasses.39 VeriLk Glasses Glasses.5 PittPatt Glasses Glasses.2 Table n Prescriptin vs. n Prescriptin EER results. This ay indicate that the degradatin bserved in previus investigatins ay reflect the presence f the fraes, re than the eyeglasses theselves. 5. Discussin These experients shwed that there will be a need fr better techniques t differentiate between twins. While current technlgies can distinguish between twins st f the tie under near ideal cnditins, as the iaging variables between prbe and gallery vary, the accuracy f these systes can decrease, as seen here and as seen in [5] with the FaceVAS syste. In additin, the false accept rate under which ne btains these recgnitin rates is very high. The st significant variables that can affect recgnitin systes sees t be expressin and lighting. By using arks n the face as recgnitin features, hwever, these variables wuld be less nticeable. The glasses wuld nt be as uch f a prble either since the eye area is asked. 6. urrent and Future Wrk While there is bth 29 and 2lO data currently available, the nuber f twins captured in bth acquisitins is References [1] Twins days festival. [2] T. Bracc. Our 'nt s ipssible' issin. etwrk Wrld, 2. [3]. S. rpratin. gnitec brchure. Sep 2lO. [4] P. Grther, G. Quinn, and P. Phillips. Reprt n the evaluatin n 2d still-iage face recgnitin algriths. Technical reprt, IST IR 779, 2lO. [5] A. Jain and U. Park. Faal arks: Sft bietric fr face recgnitin. IEEE Internatinal nference n Iage Prcessing, pages 37-4, veber 29. [6] A. Jain and S. Prabhakar. an identical twins be discriinated based n fingerprints? Technical Reprt MSU-SE--23, Michigan State University, 22. [7] EUROtechnlgy. VeriLk SDK Brchure. [8] P. Phillips. Still face challenge prble: Multiple bietric grand challenge preliinary results f versin Y'IAL.pdf, Deceber 29. [9] P. Phillips, P. Flynn, K. Bwyer, R. V. Bruegge, P. Grther, G. Quinn, and M. Pruitt. Distinguishing identical twins by face recgnitin. In 211 IEEE nference n Autatic Face Gesture Recgnitin and Wrkshps (FG 211), March 211. [lo] P. J. Phillips, J. R. Beveridge, B. A. Draper, G. Givens, A. 1. O'Tle, D. S. Ble, J. Dunlp, Y. M. Lui, H. Sahibzada, and S. Weier. An intrductin t the gd, the bad, & the ugly face recgnitin challenge prble. In The 9th IEEE nference n Autatic Face and Gesture Recgnitin, Santa Barbara, A, March 211. [11] H. Schneideran,. echyba, and M. Sipe. PittPatt lO. [12] Z. Sun, A. A. Paulin, J. Feng, Z. hai, T. Tan, and A. K. Jain. A study f ulti bietric traits f identical twins. In Prceedings f the SPIE, Bietric Technlgy fr Huan Identificatin VII, vlue 7667, pages 7667T-12, 2lO. [13] M. Turk and A. Pentland. Face recgnitin using eigenfaces. IEEE n! n puter Visin and Pattern Recgnitin, pages , 1991.