ࡱ> ac`'`bjbj0   6@BBBBBB$h_ ff{@@ p= 4 0   ,ffD Dissertao de Mestrado Aluno: Giancarlo Vasconcelos Taveira do Nascimento Ttulo: Automatic Alignment and Reconstruction of Facial Depth Images Face, gender, ethnic and age group classification systems often work through an alignment, feature extraction, and identification pipeline. For that reason, the classification capacity depends on the quality of the source data. For instance, the classification can be affected by color or depth data with small flaws or damaged regions. Appropriate image reconstruction is therefore crucial for the correct operation of those systems. This work presents an elegant and effective method for aligning and reconstructing facial depth images from damaged depth data in a completely automatic way. The approach uses information extracted from valid pixels to adjust a smooth interpolating function that naturally reconstructs the depth information of missing pixels and computes smooth transitions among existing ones. The approach also explores facial landmarks in order to determine the actual position and orientation of the imaged face in the 3-dimensional space. The relation between the set of landmarks in the actual face and a set of canonical landmarks is used to map the shape of the imaged face to a standard space where the resulting aligned image is generated by ray casting the reconstructed surface. In contrast to existing solutions, the proposed approach is straightforward and easily fits into popular processing pipelines. It can also be extended to produce correct color images for resulting depth images. The experiments show that the approximation errors produced by the proposed method are up to two orders of magnitude smaller than those using 2-dimensional alignment with linear interpolation. This work also presents a comparative study among four distinct interpolation methods (i.e., nearest-neighbor, linear, natural-neighbor and Thin Plate Spline) using the proposed alignment method (in the 3-dimensional domain). Each interpolation method was applied as part of a gender classification process available in the literature. From the obtained results it is possible to state which interpolation approach is better to be applied with direct or iterative classifiers. Keywords: depth image, alignment, non-linear interpolation, resampling, face image. :?@hdMh>G5hdMhdM5 hdM6hdMhdM6hdM hrhdMhrhdMCJKE M $a$gddM $a$gddMgddM .:p{|. A!"#$% 666666666vvvvvvvvv666666>6666666666666666666666666666666666666666666666666hH666666666666666666666666666666666666666666666666666666666666666666H@H NormalCJ_HaJmH nHsH tH>A> Fonte parg. padroTiT 0 Tabela normal4 l4a ,k, 0 Sem lista   KEM l >.l l l l  l l l l l KEM 08000000000      rdM>G{So @XW1 P@UnknownGz Times New Roman5Symbol3& z ArialG  MS Mincho-3 fg7K@Cambria 1hQ[Q[kk!hh4  2HP$P'dM2Dissertao de MestradoGiancarlo TaveiravivianeOh+'0@ $0 P \ ht|Dissertao de MestradoGiancarlo Taveira Normal.dotviviane2Microsoft Office Word@@>dz@>dzkGQt3 AA&" WMFC :Zl^iQt EMFZ'   ^iRp@Cambria@06h8g / |(506h6hm%05@00;0ЊЊU5UU7K@Cambri@50@5Ln0X5dv%   TX@ \@L^i|Dissertao de MestradoXM!2290(949>A9_92'09A>TTX@ \@L^iP ERp@CambriaD5߈g /,߈(5}06h6hm%05@00;00phv0#h X3!0]|7K@Cambri@50@5Ln0X5dv%  Tx4 @ \@2L^iAluno: Giancarlo Vasconcelos Taveira do Nascimento>785=18,1)5<1+,68,15+;121)175D1+,S18"5TT 4 @ \@ L^iP ; Tq@ \@\EL^iTtulo: Automatic Alignment and Reconstruction of Facial Depth Images;"75>7"5S1",>18S18"187>1,58+")7,"58561,1B18"7 S111+TTq@ \@\L^iP ; TT(9@ \@$L^iP ; T,@ \@PL^iFace, gender, ethnic and age group classification systems often work through an 61,111871)1"78,1771111)578,1++,1"58+2+"1S+5"18M6)5"7)571817 Tw@ \@bL^ialignment, feature extraction,118S18"*11"7)1*10"(1,"58Tw@ \@b3L^i and identification pipeline. For that reason, the *087*718",1"58*87181*65)*"71"*(11+58*"60 T4x@ \@QL^iclassification capacity depends on the quality of the source data. For instance, ,1++,1"58$,181,"2$718187+$58$"71$771"3$5$"72$+57),2$71"1$65)$8+"18,1 T b@ \@MNL^ithe classification can be affected by color or depth data with small flaws or "71,,1++,1"58,,18+71,11,"17,72,,55),6),718"7,71"1,M"7,+S2,1M+,6* T`cF @ \@.L^idamaged regions. Appropriate image reconstruct71S1117!)1158+ >88*58)1"1 S111 )1,58+"*7,"TG c@ \@G !L^iion is therefore crucial for the 58 + "71)15)1!,)8,1 5* "70 T L@ \@7#L^icorrect operation of those systems.,5))1,"581)1"585"75+1+2+"1S+TTL@ \@7L^iP : TTM@ \@L^iP ,TM@ \@DL^iThis work presents an elegant and effective method for aligning and ;7+0M5*408)1+18"+018011118"0187011,"210S1"75705*0118820187 T6 @ \@! KL^ireconstructing facial depth images from damaged depth data in a completely )1,58+")7,"81*1,1*718"7*S111+*)5S*71S1117*718"7*71"1*7*1*+5S81"11 TP8 @ \@ +L^iautomatic way. The approach uses informatio17"5S1",)M12);71)178)51,7)7+1+)85)S1"5T 8  @ \@ !L^in extracted from valid pixels to 8)10")1,"17))5S)217)801+)!4 T ! @ \@ ML^iadjust a smooth interpolating function that naturally reconstructs the depth 177+"/1/+S55"7/8"1*851"81/78,"58/"71"/81"7)12/)1,58+")8,"+0"71/718"6 T"  @ \@ ML^iinformation of missi&" WMFC Zng pixels and computes smooth transitions among existing 85)S1"58!5!S++81!801+!187!,5S87"1+!*S55"7!")18+"58+!1S481!10+"81 T S @ \@ EL^iones. The approach also explores facial landmarks in order to determi581+*;71*088)51,7*1+5*1085)1+*1,1*186S1)4+*8*5)71)*"5*71"1)STxT  @ \@T L^i\ne the 81*"70 T  @ \@l ;L^iactual position and orientation of the imaged face in the 31,"71%85+"58%087%5)18"1"58%5%"71%S1116%1,1%8%"71%7TT 7 @ \@ l L^iP-!T8  @ \@8 l L^itdimensional space. 7S18+581%+81,0 T  @ \@ JL^iThe relation between the set of landmarks in the actual face and a set of ;713)11"58371"M1183"713+1"353187S1*4+483"7131,"7131,13188313+1"35 T  k @ \@V NL^icanonical landmarks is used to map the shape of the imaged face to a standard ,1858,1187S1)4++7+17"5S18"71+71815"71S11171,1"5 1+"1871)7 Tl  @ \@ L^ipspace where the re+81,1=M71)1="71=)1Tl  @ \@ 6L^isulting aligned image is generated by ray casting the +6"81=11817=S111=*=1181)1"17=72=)12=,0+"81="70 T V @ \@A L^ixreconstructed surface.)1,58+")7,"17+7*1,1TT V @ \@A L^iP ; TTW  @ \@ L^iP ,TW  @ \@ <L^iIn contrast to existing solutions, the proposed approach is 8l,58")1+"l"5l10+"81l+57"58+l"70l8)585+17l188)51,7l+ T8 @@ \@+RL^istraightforward and easily fits into popular processing pipelines. It can also be +")117"5)M2)7!188!11+2!"+!8"5!85871)!8*5,1++81!88181*! "!,18!1+5!70 T$A @ \@$L^iextended to produce correct color im10"187178"588)577,18,5))1,"8,55)8ST, A@ \@ %L^iages for resulting depth images. The 111+85)8)1+7"818718"78R111+8;70 T+@ \@HL^iexperiments show that the approximation errors produced by the proposed 1081)S18"+/+75M/"71"/"71/188)50S1"58/0))5)+/8)677,18/72/"72/8)585+17 T,N@ \@CL^imethod are up to two orders of magnitude smaller than those using 2S1"757=1)1>78="5="N5=5)72)+=5=S118"781=+S11)="718="75+2=7+81=8TTO,o@ \@OL^iP-! Tl @ \@0L^idimensional alignment with linear interpolation.7S18+581118S18"M"7811)8"1)851"48TT @ @ \@ L^iP : TT@ \@uL^iP ,T @ \@uL^iThis work also presents a com;7+BM5)4B1+5B8)1+28"+B1B,5ST  @ \@ u#L^iparative study among four distinct 81)1"21B+!772B1S581B67)B7+"8+!Rp@CambriaD5߈g /,߈(5X5X3!parative7K@Cambri@50@5Ln0X5dv%  % T@ \@L^i|interpolation methods (V &WMFCZ8"1)851"57#S1"757+"&% Tl@ \@L^iXi.e.,.TT@ \@L^iP % Tx@ \@L^i\nearest811)1+"TT @ \@L^iP-!T ! @ \@ L^ineighbor, linear, natural811775)#811)"81"7)1TT" B @ \@" L^iP-!TC @ \@C L^ipneighbor and Thin 811775)#187#;77 T?u@ \@`;L^iPlate Spline) using the proposed alignment method (in the 391"1;2881&;7+81;"61;8)585+17;118S18!;S1"757;&8;"71;7TT@`u@ \@@`L^iP-!Tau@ \@a` L^iddimensional 7S18+581 Tv@ \@CL^idomain). Each interpolation method was applied as part of a gender 75S18&O:1,7N8"1)751"58OS1"757OM1+O18817O1+N81)"O5O1O11871( TT_@ \@JL^iPc,T@_@ \@JSL^ilassification process available in the literature. From the obtained results it is 1++,1"58"8)5,1++!021171"8""70""1)1"7(1"6)5S""71"47"1717")1+7"+"""* T@`@ \@SL^ipossible to state which interpolation approach is better to be applied with direct 85++71"5*"1"1M7,78"1)851"58188)51,7+71""1)"47108817M"77)1," TJ@ \@5L^ior iterative classifiers.5)"1)1"21,1++1)+TTJ@ \@5L^iP : TTK(@ \@L^iP ;Rp@CambriaD5|߈8߈g /|߈(5}06h6hm%05@00;0ve classifiers.nterplatio7K@Cambri@50@5Ln0X5dv%  T0W @ \@r&L^iKeywords: depth image, alignment, nonD55P9.<."<5<%;"Y545""64<Y5<%"<9<TTX y @ \@X rL^iP-"Tz @ \@z rL^i|linear interpolation, re <56."<%5.<96%9<".5T@ \@r L^i`sampling, .6Y<<4 T@ \@ L^idface image.!6/5Y645TT&@ \@L^iP :% 6i6^i6^66h6]h6]66g6\g6\66f6[f6[66e6Ze6Z66d6Yd6Y66c6Xc6X66b6Wb6W66a6Va6V6 6 `6U`6U 6  6 _6T_6T 6  6 ^6S^6S 6  6 ]6R]6R 6  6 \6Q\6Q 6 6[6P[6P66Z6OZ6O66Y6NY6N66X6MX6M66W6LW6L6  .@Cambria- .2 Dissertao de Mestradoe     2  G@Cambria-V2 2Aluno: Giancarlo Vasconcelos Taveira do Nascimento             2 ~ G s2 ETtulo: Automatic Alignment and Reconstruction of Facial Depth Imagese                    2 ' G  2  G 2 APFace, gender, ethnic and age group classification systems often work through an                     82 Zalignment, feature extraction,         X2 Z3 and identification pipeline. For that reason, the h           2 rQclassification capacity depends on the quality of the source data. For instance,                     2 Nthe classification can be affected by color or depth data with small flaws or                P2 .damaged regions. Appropriate image reconstruct          =2 d!ion is therefore crucial for the ,         @2 #correct operation of those systems.n          2  G  2  G@q2 DThis work presents an elegant and effective method for aligning and             |2 Kreconstructing facial depth images from damaged depth data in a completely o                  L2 +automatic way. The approach uses informatiou           =2 P!n extracted from valid pixels to s         2 Madjust a smooth interpolating function that naturally reconstructs the depth                   2 9Minformation of missing pixels and computes smooth transitions among existing                         s2 QEones. The approach also explores facial landmarks in order to determie              2 Q@ne the  d2 j;actual position and orientation of the imaged face in the 3                2 j-G(2 jdimensional space. r    z2 JThe relation between the set of landmarks in the actual face and a set of               2 Ncanonical landmarks is used to map the shape of the imaged face to a standard                   &2 space where the re    \2 c6sulting aligned image is generated by ray casting the        ,2 reconstructed surface.      2 i G  2  G@e2 <In contrast to existing solutions, the proposed approach is                2 Rstraightforward and easily fits into popular processing pipelines. It can also be                       A2 $extended to produce correct color im      C2 %ages for resulting depth images. The n       w2 1Hexperiments show that the approximation errors produced by the proposed              p2 JCmethod are up to two orders of magnitude smaller than those using 2o             2 Ju-GS2 b0dimensional alignment with linear interpolation.             2 bV G  2 { G@72 {This work also presents a com,         @2 {#parative study among four distinct e          @Cambria--.2 interpolation methods (e      -2 |i.e.,e  2  G-2 nearest  2 -G12 neighbor, linear, natural      2 -G&2 neighbor and Thin     d2 ;Plate Spline) using the proposed alignment method (in the 3              2 -G2   dimensional   p2 Cdomain). Each interpolation method was applied as part of a gender o                   2 cG 2 Slassification process available in the literature. From the obtained results it is                    2 Spossible to state which interpolation approach is better to be applied with direct                        12 or iterative classifiers.       2 g G  2 ) G @Cambria-D2 S&Keywords: depth image, alignment, non         2 S1-G/2 S8linear interpolation, re      2 S sampling,   2 l face image.   2 l  G "System-՜.+,0 hp  Home ' Dissertao de Mestrado Ttulo  !"#$%&'()*+,-./0123456789:;<=>?@ABCDEFGHIJKLMNOPQRSTUVWYZ[\]^_bRoot Entry F`Bd1Table WordDocument0SummaryInformation(pDocumentSummaryInformation8XCompObju  F#Documento do Microsoft Office Word MSWordDocWord.Document.89q