id sid tid token lemma pos fcis-1100 1 1 frontiers frontier NOUN fcis-1100 1 2 in in ADP fcis-1100 1 3 computing computing NOUN fcis-1100 1 4 and and CCONJ fcis-1100 1 5 intelligent intelligent ADJ fcis-1100 1 6 systems system NOUN fcis-1100 1 7 issn issn VERB fcis-1100 1 8 : : PUNCT fcis-1100 1 9 2832 2832 NUM fcis-1100 1 10 - - SYM fcis-1100 1 11 6024 6024 NUM fcis-1100 1 12 | | NOUN fcis-1100 1 13 vol vol NOUN fcis-1100 1 14 . . PROPN fcis-1100 2 1 1 1 NUM fcis-1100 2 2 , , PUNCT fcis-1100 2 3 no no INTJ fcis-1100 2 4 . . NOUN fcis-1100 2 5 1 1 NUM fcis-1100 2 6 , , PUNCT fcis-1100 2 7 2022 2022 NUM fcis-1100 2 8 22 22 NUM fcis-1100 2 9 a a DET fcis-1100 2 10 face face NOUN fcis-1100 2 11 recognition recognition NOUN fcis-1100 2 12 algorithm algorithm NOUN fcis-1100 2 13 based base VERB fcis-1100 2 14 on on ADP fcis-1100 2 15 improved improve VERB fcis-1100 2 16 resnet resnet NOUN fcis-1100 2 17 hongrong hongrong NOUN fcis-1100 2 18 jing1 jing1 PROPN fcis-1100 2 19 , , PUNCT fcis-1100 2 20 guojun guojun PROPN fcis-1100 2 21 lin1,2 lin1,2 PROPN fcis-1100 2 22 , , PUNCT fcis-1100 2 23 * * PROPN fcis-1100 2 24 , , PUNCT fcis-1100 2 25 hongjie hongjie ADJ fcis-1100 2 26 zhang1 zhang1 PROPN fcis-1100 2 27 , , PUNCT fcis-1100 2 28 tiantian tiantian PROPN fcis-1100 2 29 chen1 chen1 PROPN fcis-1100 3 1 1 1 NUM fcis-1100 3 2 department department NOUN fcis-1100 3 3 of of ADP fcis-1100 3 4 automation automation NOUN fcis-1100 3 5 and and CCONJ fcis-1100 3 6 information information NOUN fcis-1100 3 7 engineering engineering NOUN fcis-1100 3 8 , , PUNCT fcis-1100 3 9 sichuan sichuan PROPN fcis-1100 3 10 university university PROPN fcis-1100 3 11 of of ADP fcis-1100 3 12 science science PROPN fcis-1100 3 13 & & CCONJ fcis-1100 3 14 engineering engineering PROPN fcis-1100 3 15 , , PUNCT fcis-1100 3 16 yibin yibin PROPN fcis-1100 3 17 633000 633000 NUM fcis-1100 3 18 , , PUNCT fcis-1100 3 19 sichuan sichuan PROPN fcis-1100 3 20 , , PUNCT fcis-1100 3 21 china china PROPN fcis-1100 3 22 2 2 NUM fcis-1100 3 23 artificial artificial ADJ fcis-1100 3 24 intelligence intelligence NOUN fcis-1100 3 25 key key NOUN fcis-1100 3 26 laboratory laboratory NOUN fcis-1100 3 27 of of ADP fcis-1100 3 28 sichuan sichuan PROPN fcis-1100 3 29 province province PROPN fcis-1100 3 30 , , PUNCT fcis-1100 3 31 sichuan sichuan PROPN fcis-1100 3 32 university university PROPN fcis-1100 3 33 of of ADP fcis-1100 3 34 science science PROPN fcis-1100 3 35 & & CCONJ fcis-1100 3 36 engineering engineering PROPN fcis-1100 3 37 , , PUNCT fcis-1100 3 38 yibin yibin PROPN fcis-1100 3 39 633000 633000 NUM fcis-1100 3 40 , , PUNCT fcis-1100 3 41 sichuan sichuan PROPN fcis-1100 3 42 , , PUNCT fcis-1100 3 43 china china PROPN fcis-1100 3 44 * * PUNCT fcis-1100 3 45 corresponding correspond VERB fcis-1100 3 46 author author NOUN fcis-1100 3 47 : : PUNCT fcis-1100 3 48 guojun guojun PROPN fcis-1100 3 49 lin lin PROPN fcis-1100 3 50 ( ( PUNCT fcis-1100 3 51 email email NOUN fcis-1100 3 52 : : PUNCT fcis-1100 3 53 386988463@qq.com 386988463@qq.com NUM fcis-1100 3 54 ) ) PUNCT fcis-1100 3 55 abstract abstract NOUN fcis-1100 3 56 : : PUNCT fcis-1100 3 57 regarding regard VERB fcis-1100 3 58 the the DET fcis-1100 3 59 problem problem NOUN fcis-1100 3 60 that that SCONJ fcis-1100 3 61 the the DET fcis-1100 3 62 increasing increase VERB fcis-1100 3 63 number number NOUN fcis-1100 3 64 of of ADP fcis-1100 3 65 layers layer NOUN fcis-1100 3 66 of of ADP fcis-1100 3 67 cnn cnn PROPN fcis-1100 3 68 ( ( PUNCT fcis-1100 3 69 convolutional convolutional ADJ fcis-1100 3 70 neural neural ADJ fcis-1100 3 71 network network NOUN fcis-1100 3 72 ) ) PUNCT fcis-1100 3 73 leads lead VERB fcis-1100 3 74 to to ADP fcis-1100 3 75 the the DET fcis-1100 3 76 decline decline NOUN fcis-1100 3 77 of of ADP fcis-1100 3 78 accuracy accuracy NOUN fcis-1100 3 79 , , PUNCT fcis-1100 3 80 an an DET fcis-1100 3 81 improved improved ADJ fcis-1100 3 82 loss loss NOUN fcis-1100 3 83 function function NOUN fcis-1100 3 84 algorithm algorithm NOUN fcis-1100 3 85 based base VERB fcis-1100 3 86 on on ADP fcis-1100 3 87 the the DET fcis-1100 3 88 resnet-50 resnet-50 PROPN fcis-1100 3 89 model model NOUN fcis-1100 3 90 is be AUX fcis-1100 3 91 proposed propose VERB fcis-1100 3 92 . . PUNCT fcis-1100 4 1 the the DET fcis-1100 4 2 softmax softmax NOUN fcis-1100 4 3 loss loss NOUN fcis-1100 4 4 function function NOUN fcis-1100 4 5 lacks lack VERB fcis-1100 4 6 constraints constraint NOUN fcis-1100 4 7 on on ADP fcis-1100 4 8 the the DET fcis-1100 4 9 distance distance NOUN fcis-1100 4 10 within within ADP fcis-1100 4 11 the the DET fcis-1100 4 12 same same ADJ fcis-1100 4 13 class class NOUN fcis-1100 4 14 and and CCONJ fcis-1100 4 15 between between ADP fcis-1100 4 16 different different ADJ fcis-1100 4 17 classes class NOUN fcis-1100 4 18 . . PUNCT fcis-1100 5 1 replacing replace VERB fcis-1100 5 2 the the DET fcis-1100 5 3 softmax softmax NOUN fcis-1100 5 4 layer layer NOUN fcis-1100 5 5 with with ADP fcis-1100 5 6 improved improved ADJ fcis-1100 5 7 arcface arcface NOUN fcis-1100 5 8 loss loss NOUN fcis-1100 5 9 enables enable VERB fcis-1100 5 10 the the DET fcis-1100 5 11 neural neural ADJ fcis-1100 5 12 network network NOUN fcis-1100 5 13 to to PART fcis-1100 5 14 learn learn VERB fcis-1100 5 15 more more ADJ fcis-1100 5 16 distinguishing distinguish VERB fcis-1100 5 17 features feature NOUN fcis-1100 5 18 . . PUNCT fcis-1100 6 1 experiments experiment NOUN fcis-1100 6 2 on on ADP fcis-1100 6 3 lfw lfw NOUN fcis-1100 6 4 and and CCONJ fcis-1100 6 5 agedb agedb PROPN fcis-1100 6 6 data data NOUN fcis-1100 6 7 sets set NOUN fcis-1100 6 8 show show VERB fcis-1100 6 9 that that SCONJ fcis-1100 6 10 the the DET fcis-1100 6 11 algorithm algorithm NOUN fcis-1100 6 12 can can AUX fcis-1100 6 13 not not PART fcis-1100 6 14 only only ADV fcis-1100 6 15 learn learn VERB fcis-1100 6 16 deep deep ADJ fcis-1100 6 17 - - PUNCT fcis-1100 6 18 face face NOUN fcis-1100 6 19 characteristics characteristic NOUN fcis-1100 6 20 but but CCONJ fcis-1100 6 21 also also ADV fcis-1100 6 22 efficiently efficiently ADV fcis-1100 6 23 improve improve VERB fcis-1100 6 24 the the DET fcis-1100 6 25 accuracy accuracy NOUN fcis-1100 6 26 of of ADP fcis-1100 6 27 face face NOUN fcis-1100 6 28 recognition recognition NOUN fcis-1100 6 29 compared compare VERB fcis-1100 6 30 with with ADP fcis-1100 6 31 ordinary ordinary ADJ fcis-1100 6 32 cnn cnn PROPN fcis-1100 6 33 . . PUNCT fcis-1100 7 1 in in ADP fcis-1100 7 2 the the DET fcis-1100 7 3 meantime meantime NOUN fcis-1100 7 4 , , PUNCT fcis-1100 7 5 the the DET fcis-1100 7 6 improved improved ADJ fcis-1100 7 7 resnet resnet NOUN fcis-1100 7 8 also also ADV fcis-1100 7 9 obtains obtain VERB fcis-1100 7 10 a a DET fcis-1100 7 11 higher high ADJ fcis-1100 7 12 discerning discerning NOUN fcis-1100 7 13 rate rate NOUN fcis-1100 7 14 under under ADP fcis-1100 7 15 the the DET fcis-1100 7 16 conditions condition NOUN fcis-1100 7 17 of of ADP fcis-1100 7 18 occlusions occlusion NOUN fcis-1100 7 19 , , PUNCT fcis-1100 7 20 illumination illumination NOUN fcis-1100 7 21 , , PUNCT fcis-1100 7 22 expression expression NOUN fcis-1100 7 23 , , PUNCT fcis-1100 7 24 age age NOUN fcis-1100 7 25 . . PUNCT fcis-1100 8 1 keywords keyword NOUN fcis-1100 8 2 : : PUNCT fcis-1100 8 3 deep deep ADJ fcis-1100 8 4 learning learning NOUN fcis-1100 8 5 ; ; PUNCT fcis-1100 8 6 residual residual ADJ fcis-1100 8 7 ; ; PUNCT fcis-1100 8 8 face face NOUN fcis-1100 8 9 recognition recognition NOUN fcis-1100 8 10 ; ; PUNCT fcis-1100 8 11 softmax softmax NOUN fcis-1100 8 12 ; ; PUNCT fcis-1100 8 13 improved improve VERB fcis-1100 8 14 loss loss NOUN fcis-1100 8 15 function function NOUN fcis-1100 8 16 . . PUNCT fcis-1100 9 1 1 1 X fcis-1100 9 2 . . X fcis-1100 9 3 introduction introduction NOUN fcis-1100 9 4 in in ADP fcis-1100 9 5 recent recent ADJ fcis-1100 9 6 decades decade NOUN fcis-1100 9 7 , , PUNCT fcis-1100 9 8 traditional traditional ADJ fcis-1100 9 9 face face NOUN fcis-1100 9 10 recognition recognition NOUN fcis-1100 9 11 methods method NOUN fcis-1100 9 12 are be AUX fcis-1100 9 13 used use VERB fcis-1100 9 14 to to PART fcis-1100 9 15 extract extract VERB fcis-1100 9 16 features feature NOUN fcis-1100 9 17 and and CCONJ fcis-1100 9 18 classify classify VERB fcis-1100 9 19 , , PUNCT fcis-1100 9 20 such such ADJ fcis-1100 9 21 as as ADP fcis-1100 9 22 lbp lbp NOUN fcis-1100 9 23 [ [ X fcis-1100 9 24 1 1 NUM fcis-1100 9 25 ] ] PUNCT fcis-1100 9 26 and and CCONJ fcis-1100 9 27 svm svm VERB fcis-1100 9 28 [ [ X fcis-1100 9 29 2 2 NUM fcis-1100 9 30 ] ] PUNCT fcis-1100 9 31 . . PUNCT fcis-1100 10 1 in in ADP fcis-1100 10 2 the the DET fcis-1100 10 3 case case NOUN fcis-1100 10 4 of of ADP fcis-1100 10 5 a a DET fcis-1100 10 6 small small ADJ fcis-1100 10 7 number number NOUN fcis-1100 10 8 of of ADP fcis-1100 10 9 samples sample NOUN fcis-1100 10 10 , , PUNCT fcis-1100 10 11 this this DET fcis-1100 10 12 kind kind NOUN fcis-1100 10 13 of of ADP fcis-1100 10 14 method method NOUN fcis-1100 10 15 has have VERB fcis-1100 10 16 a a DET fcis-1100 10 17 good good ADJ fcis-1100 10 18 result result NOUN fcis-1100 10 19 , , PUNCT fcis-1100 10 20 but but CCONJ fcis-1100 10 21 with with ADP fcis-1100 10 22 the the DET fcis-1100 10 23 increase increase NOUN fcis-1100 10 24 of of ADP fcis-1100 10 25 face face NOUN fcis-1100 10 26 data datum NOUN fcis-1100 10 27 , , PUNCT fcis-1100 10 28 the the DET fcis-1100 10 29 traditional traditional ADJ fcis-1100 10 30 methods method NOUN fcis-1100 10 31 are be AUX fcis-1100 10 32 far far ADV fcis-1100 10 33 from from ADP fcis-1100 10 34 meeting meet VERB fcis-1100 10 35 the the DET fcis-1100 10 36 requirements requirement NOUN fcis-1100 10 37 . . PUNCT fcis-1100 11 1 with with ADP fcis-1100 11 2 the the DET fcis-1100 11 3 vigorous vigorous ADJ fcis-1100 11 4 development development NOUN fcis-1100 11 5 of of ADP fcis-1100 11 6 computer computer NOUN fcis-1100 11 7 vision vision NOUN fcis-1100 11 8 and and CCONJ fcis-1100 11 9 deep deep ADJ fcis-1100 11 10 learning learning NOUN fcis-1100 11 11 algorithms algorithm NOUN fcis-1100 11 12 in in ADP fcis-1100 11 13 recent recent ADJ fcis-1100 11 14 years year NOUN fcis-1100 11 15 , , PUNCT fcis-1100 11 16 especially especially ADV fcis-1100 11 17 the the DET fcis-1100 11 18 development development NOUN fcis-1100 11 19 of of ADP fcis-1100 11 20 cnn cnn PROPN fcis-1100 11 21 , , PUNCT fcis-1100 11 22 a a DET fcis-1100 11 23 powerful powerful ADJ fcis-1100 11 24 classification classification NOUN fcis-1100 11 25 method method NOUN fcis-1100 11 26 commonly commonly ADV fcis-1100 11 27 used use VERB fcis-1100 11 28 to to PART fcis-1100 11 29 recognize recognize VERB fcis-1100 11 30 and and CCONJ fcis-1100 11 31 verify verify VERB fcis-1100 11 32 images image NOUN fcis-1100 11 33 [ [ X fcis-1100 11 34 3 3 NUM fcis-1100 11 35 ] ] PUNCT fcis-1100 11 36 , , PUNCT fcis-1100 11 37 face face VERB fcis-1100 11 38 recognition recognition NOUN fcis-1100 11 39 technology technology NOUN fcis-1100 12 1 [ [ X fcis-1100 12 2 4 4 X fcis-1100 12 3 ] ] PUNCT fcis-1100 12 4 has have AUX fcis-1100 12 5 also also ADV fcis-1100 12 6 developed develop VERB fcis-1100 12 7 rapidly rapidly ADV fcis-1100 12 8 and and CCONJ fcis-1100 12 9 gradually gradually ADV fcis-1100 12 10 stepped step VERB fcis-1100 12 11 out out ADP fcis-1100 12 12 to to ADP fcis-1100 12 13 the the DET fcis-1100 12 14 life life NOUN fcis-1100 12 15 . . PUNCT fcis-1100 13 1 compared compare VERB fcis-1100 13 2 with with ADP fcis-1100 13 3 the the DET fcis-1100 13 4 traditional traditional ADJ fcis-1100 13 5 face face NOUN fcis-1100 13 6 recognition recognition NOUN fcis-1100 13 7 methods method NOUN fcis-1100 13 8 , , PUNCT fcis-1100 13 9 the the DET fcis-1100 13 10 face face NOUN fcis-1100 13 11 recognition recognition NOUN fcis-1100 13 12 algorithm algorithm NOUN fcis-1100 13 13 based base VERB fcis-1100 13 14 on on ADP fcis-1100 13 15 deep deep ADJ fcis-1100 13 16 learning learning NOUN fcis-1100 13 17 has have VERB fcis-1100 13 18 better well ADJ fcis-1100 13 19 recognition recognition NOUN fcis-1100 13 20 accuracy accuracy NOUN fcis-1100 13 21 . . PUNCT fcis-1100 14 1 thus thus ADV fcis-1100 14 2 , , PUNCT fcis-1100 14 3 it it PRON fcis-1100 14 4 can can AUX fcis-1100 14 5 be be AUX fcis-1100 14 6 seen see VERB fcis-1100 14 7 that that SCONJ fcis-1100 14 8 deep deep ADJ fcis-1100 14 9 learning learning NOUN fcis-1100 14 10 plays play VERB fcis-1100 14 11 a a DET fcis-1100 14 12 great great ADJ fcis-1100 14 13 role role NOUN fcis-1100 14 14 in in ADP fcis-1100 14 15 face face NOUN fcis-1100 14 16 recognition recognition NOUN fcis-1100 15 1 [ [ X fcis-1100 15 2 5 5 NUM fcis-1100 15 3 ] ] PUNCT fcis-1100 15 4 . . PUNCT fcis-1100 16 1 in in ADP fcis-1100 16 2 view view NOUN fcis-1100 16 3 of of ADP fcis-1100 16 4 the the DET fcis-1100 16 5 convenience convenience NOUN fcis-1100 16 6 , , PUNCT fcis-1100 16 7 uniqueness uniqueness NOUN fcis-1100 16 8 , , PUNCT fcis-1100 16 9 and and CCONJ fcis-1100 16 10 non non ADJ fcis-1100 16 11 - - NOUN fcis-1100 16 12 repeatability repeatability NOUN fcis-1100 16 13 of of ADP fcis-1100 16 14 human human ADJ fcis-1100 16 15 faces face NOUN fcis-1100 16 16 , , PUNCT fcis-1100 16 17 face face NOUN fcis-1100 16 18 recognition recognition NOUN fcis-1100 16 19 technology technology NOUN fcis-1100 16 20 is be AUX fcis-1100 16 21 widely widely ADV fcis-1100 16 22 used use VERB fcis-1100 16 23 in in ADP fcis-1100 16 24 many many ADJ fcis-1100 16 25 aspects aspect NOUN fcis-1100 16 26 of of ADP fcis-1100 16 27 society society NOUN fcis-1100 16 28 , , PUNCT fcis-1100 16 29 such such ADJ fcis-1100 16 30 as as ADP fcis-1100 16 31 security security NOUN fcis-1100 16 32 , , PUNCT fcis-1100 16 33 finance finance NOUN fcis-1100 16 34 , , PUNCT fcis-1100 16 35 scientific scientific ADJ fcis-1100 16 36 research research NOUN fcis-1100 16 37 , , PUNCT fcis-1100 16 38 and and CCONJ fcis-1100 16 39 so so ADV fcis-1100 16 40 on on ADV fcis-1100 16 41 . . PUNCT fcis-1100 17 1 face face NOUN fcis-1100 17 2 recognition recognition NOUN fcis-1100 17 3 has have AUX fcis-1100 17 4 become become VERB fcis-1100 17 5 the the DET fcis-1100 17 6 future future ADJ fcis-1100 17 7 development development NOUN fcis-1100 17 8 direction direction NOUN fcis-1100 17 9 with with ADP fcis-1100 17 10 many many ADJ fcis-1100 17 11 potential potential ADJ fcis-1100 17 12 application application NOUN fcis-1100 17 13 prospects prospect NOUN fcis-1100 17 14 [ [ X fcis-1100 17 15 6 6 NUM fcis-1100 17 16 ] ] PUNCT fcis-1100 17 17 . . PUNCT fcis-1100 18 1 over over ADP fcis-1100 18 2 the the DET fcis-1100 18 3 past past ADJ fcis-1100 18 4 five five NUM fcis-1100 18 5 years year NOUN fcis-1100 18 6 , , PUNCT fcis-1100 18 7 it it PRON fcis-1100 18 8 has have AUX fcis-1100 18 9 made make VERB fcis-1100 18 10 a a DET fcis-1100 18 11 qualitative qualitative ADJ fcis-1100 18 12 leap leap NOUN fcis-1100 19 1 [ [ X fcis-1100 19 2 7 7 NUM fcis-1100 19 3 ] ] PUNCT fcis-1100 19 4 . . PUNCT fcis-1100 20 1 since since SCONJ fcis-1100 20 2 the the DET fcis-1100 20 3 resnet resnet NOUN fcis-1100 20 4 was be AUX fcis-1100 20 5 put put VERB fcis-1100 20 6 forward forward ADV fcis-1100 20 7 in in ADP fcis-1100 20 8 2015 2015 NUM fcis-1100 20 9 , , PUNCT fcis-1100 20 10 more more ADJ fcis-1100 20 11 and and CCONJ fcis-1100 20 12 more more ADV fcis-1100 20 13 excellent excellent ADJ fcis-1100 20 14 algorithms algorithm NOUN fcis-1100 20 15 based base VERB fcis-1100 20 16 on on ADP fcis-1100 20 17 resnet resnet NOUN fcis-1100 20 18 have have AUX fcis-1100 20 19 been be AUX fcis-1100 20 20 proposed propose VERB fcis-1100 20 21 and and CCONJ fcis-1100 20 22 good good ADJ fcis-1100 20 23 achievements achievement NOUN fcis-1100 20 24 have have AUX fcis-1100 20 25 been be AUX fcis-1100 20 26 made make VERB fcis-1100 20 27 in in ADP fcis-1100 20 28 the the DET fcis-1100 20 29 field field NOUN fcis-1100 20 30 of of ADP fcis-1100 20 31 face face NOUN fcis-1100 20 32 recognition recognition NOUN fcis-1100 20 33 . . PUNCT fcis-1100 21 1 under under ADP fcis-1100 21 2 the the DET fcis-1100 21 3 situation situation NOUN fcis-1100 21 4 that that SCONJ fcis-1100 21 5 it it PRON fcis-1100 21 6 is be AUX fcis-1100 21 7 difficult difficult ADJ fcis-1100 21 8 to to PART fcis-1100 21 9 further far ADV fcis-1100 21 10 upgrade upgrade VERB fcis-1100 21 11 and and CCONJ fcis-1100 21 12 optimize optimize VERB fcis-1100 21 13 the the DET fcis-1100 21 14 network network NOUN fcis-1100 21 15 structure structure NOUN fcis-1100 21 16 , , PUNCT fcis-1100 21 17 researchers researcher NOUN fcis-1100 21 18 gradually gradually ADV fcis-1100 21 19 turn turn VERB fcis-1100 21 20 their their PRON fcis-1100 21 21 attention attention NOUN fcis-1100 21 22 to to ADP fcis-1100 21 23 the the DET fcis-1100 21 24 field field NOUN fcis-1100 21 25 of of ADP fcis-1100 21 26 loss loss NOUN fcis-1100 21 27 function function NOUN fcis-1100 21 28 and and CCONJ fcis-1100 21 29 attention attention NOUN fcis-1100 21 30 network network NOUN fcis-1100 21 31 [ [ X fcis-1100 21 32 8 8 NUM fcis-1100 21 33 ] ] PUNCT fcis-1100 21 34 . . PUNCT fcis-1100 22 1 starting start VERB fcis-1100 22 2 from from ADP fcis-1100 22 3 the the DET fcis-1100 22 4 improved improved ADJ fcis-1100 22 5 loss loss NOUN fcis-1100 22 6 function function NOUN fcis-1100 22 7 , , PUNCT fcis-1100 22 8 this this DET fcis-1100 22 9 paper paper NOUN fcis-1100 22 10 uses use VERB fcis-1100 22 11 a a DET fcis-1100 22 12 residual residual ADJ fcis-1100 22 13 network network NOUN fcis-1100 22 14 different different ADJ fcis-1100 22 15 from from ADP fcis-1100 22 16 the the DET fcis-1100 22 17 original original ADJ fcis-1100 22 18 algorithm algorithm NOUN fcis-1100 22 19 for for ADP fcis-1100 22 20 face face NOUN fcis-1100 22 21 recognition recognition NOUN fcis-1100 22 22 . . PUNCT fcis-1100 23 1 the the DET fcis-1100 23 2 experimental experimental ADJ fcis-1100 23 3 results result NOUN fcis-1100 23 4 show show VERB fcis-1100 23 5 that that SCONJ fcis-1100 23 6 the the DET fcis-1100 23 7 improved improved ADJ fcis-1100 23 8 loss loss NOUN fcis-1100 23 9 function function NOUN fcis-1100 23 10 increases increase VERB fcis-1100 23 11 the the DET fcis-1100 23 12 accuracy accuracy NOUN fcis-1100 23 13 of of ADP fcis-1100 23 14 face face NOUN fcis-1100 23 15 recognition recognition NOUN fcis-1100 23 16 . . PUNCT fcis-1100 24 1 2 2 X fcis-1100 24 2 . . X fcis-1100 24 3 principle principle NOUN fcis-1100 24 4 2.1 2.1 NUM fcis-1100 24 5 . . PUNCT fcis-1100 25 1 residual residual ADJ fcis-1100 25 2 principle principle NOUN fcis-1100 25 3 resnet resnet NOUN fcis-1100 25 4 is be AUX fcis-1100 25 5 a a DET fcis-1100 25 6 network network NOUN fcis-1100 25 7 model model NOUN fcis-1100 25 8 raised raise VERB fcis-1100 25 9 by by ADP fcis-1100 25 10 the the DET fcis-1100 25 11 kaiming kaiming NOUN fcis-1100 26 1 [ [ X fcis-1100 26 2 9 9 NUM fcis-1100 26 3 ] ] PUNCT fcis-1100 26 4 in in ADP fcis-1100 26 5 2015 2015 NUM fcis-1100 26 6 . . PUNCT fcis-1100 27 1 directly directly ADV fcis-1100 27 2 increasing increase VERB fcis-1100 27 3 the the DET fcis-1100 27 4 network network NOUN fcis-1100 27 5 depth depth NOUN fcis-1100 27 6 to to PART fcis-1100 27 7 improve improve VERB fcis-1100 27 8 the the DET fcis-1100 27 9 accuracy accuracy NOUN fcis-1100 27 10 will will AUX fcis-1100 27 11 lead lead VERB fcis-1100 27 12 to to ADP fcis-1100 27 13 two two NUM fcis-1100 27 14 problems problem NOUN fcis-1100 27 15 , , PUNCT fcis-1100 27 16 vanishing vanish VERB fcis-1100 27 17 gradient gradient NOUN fcis-1100 27 18 and and CCONJ fcis-1100 27 19 exploding explode VERB fcis-1100 27 20 gradient gradient ADJ fcis-1100 27 21 problem problem NOUN fcis-1100 27 22 and and CCONJ fcis-1100 27 23 accuracy accuracy NOUN fcis-1100 27 24 decline decline NOUN fcis-1100 27 25 . . PUNCT fcis-1100 28 1 the the DET fcis-1100 28 2 latter latter ADJ fcis-1100 28 3 is be AUX fcis-1100 28 4 not not PART fcis-1100 28 5 due due ADJ fcis-1100 28 6 to to ADP fcis-1100 28 7 over over ADV fcis-1100 28 8 - - PUNCT fcis-1100 28 9 fitting fitting ADJ fcis-1100 28 10 . . PUNCT fcis-1100 29 1 it it PRON fcis-1100 29 2 is be AUX fcis-1100 29 3 caused cause VERB fcis-1100 29 4 by by ADP fcis-1100 29 5 saturation saturation NOUN fcis-1100 29 6 or or CCONJ fcis-1100 29 7 even even ADV fcis-1100 29 8 a a DET fcis-1100 29 9 decline decline NOUN fcis-1100 29 10 in in ADP fcis-1100 29 11 accuracy accuracy NOUN fcis-1100 29 12 . . PUNCT fcis-1100 30 1 for for ADP fcis-1100 30 2 problem problem NOUN fcis-1100 30 3 1 1 NUM fcis-1100 30 4 , , PUNCT fcis-1100 30 5 batch batch NOUN fcis-1100 30 6 norm norm NOUN fcis-1100 30 7 [ [ X fcis-1100 30 8 10 10 NUM fcis-1100 30 9 ] ] PUNCT fcis-1100 30 10 is be AUX fcis-1100 30 11 adopted adopt VERB fcis-1100 30 12 . . PUNCT fcis-1100 31 1 the the DET fcis-1100 31 2 residual residual ADJ fcis-1100 31 3 learning learning NOUN fcis-1100 31 4 mechanism mechanism NOUN fcis-1100 31 5 can can AUX fcis-1100 31 6 be be AUX fcis-1100 31 7 used use VERB fcis-1100 31 8 to to PART fcis-1100 31 9 solve solve VERB fcis-1100 31 10 performance performance NOUN fcis-1100 31 11 degradation degradation NOUN fcis-1100 31 12 problems problem NOUN fcis-1100 31 13 caused cause VERB fcis-1100 31 14 by by ADP fcis-1100 31 15 the the DET fcis-1100 31 16 increase increase NOUN fcis-1100 31 17 of of ADP fcis-1100 31 18 the the DET fcis-1100 31 19 alternating alternate VERB fcis-1100 31 20 convolutional convolutional ADJ fcis-1100 31 21 layer layer NOUN fcis-1100 31 22 like like ADP fcis-1100 31 23 problem problem NOUN fcis-1100 31 24 2 2 X fcis-1100 31 25 . . PUNCT fcis-1100 32 1 if if SCONJ fcis-1100 32 2 h h PROPN fcis-1100 32 3 ( ( PUNCT fcis-1100 32 4 x x X fcis-1100 32 5 ) ) PUNCT fcis-1100 32 6 is be AUX fcis-1100 32 7 regarded regard VERB fcis-1100 32 8 as as ADP fcis-1100 32 9 the the DET fcis-1100 32 10 desired desire VERB fcis-1100 32 11 actual actual ADJ fcis-1100 32 12 mapping mapping NOUN fcis-1100 32 13 , , PUNCT fcis-1100 32 14 that that ADV fcis-1100 32 15 is is ADV fcis-1100 32 16 , , PUNCT fcis-1100 32 17 the the DET fcis-1100 32 18 stacked stacked ADJ fcis-1100 32 19 multi multi ADJ fcis-1100 32 20 - - ADJ fcis-1100 32 21 layer layer ADJ fcis-1100 32 22 nonlinear nonlinear ADJ fcis-1100 32 23 network network NOUN fcis-1100 32 24 is be AUX fcis-1100 32 25 used use VERB fcis-1100 32 26 to to PART fcis-1100 32 27 represent represent VERB fcis-1100 32 28 the the DET fcis-1100 32 29 fitting fitting NOUN fcis-1100 32 30 of of ADP fcis-1100 32 31 the the DET fcis-1100 32 32 mapping mapping NOUN fcis-1100 32 33 relationship relationship NOUN fcis-1100 32 34 , , PUNCT fcis-1100 32 35 then then ADV fcis-1100 32 36 the the DET fcis-1100 32 37 multi multi ADJ fcis-1100 32 38 - - ADJ fcis-1100 32 39 layer layer ADJ fcis-1100 32 40 network network NOUN fcis-1100 32 41 can can AUX fcis-1100 32 42 gradually gradually ADV fcis-1100 32 43 approach approach VERB fcis-1100 32 44 a a DET fcis-1100 32 45 complex complex ADJ fcis-1100 32 46 function function NOUN fcis-1100 32 47 . . PUNCT fcis-1100 33 1 it it PRON fcis-1100 33 2 is be AUX fcis-1100 33 3 assumed assume VERB fcis-1100 33 4 to to PART fcis-1100 33 5 be be AUX fcis-1100 33 6 equivalent equivalent ADJ fcis-1100 33 7 to to ADP fcis-1100 33 8 the the DET fcis-1100 33 9 approximation approximation NOUN fcis-1100 33 10 residual residual ADJ fcis-1100 33 11 function function NOUN fcis-1100 33 12 f f PROPN fcis-1100 33 13 ( ( PUNCT fcis-1100 33 14 x x NOUN fcis-1100 33 15 ) ) PUNCT fcis-1100 33 16 , , PUNCT fcis-1100 33 17 where where SCONJ fcis-1100 33 18 x x PRON fcis-1100 33 19 refers refer VERB fcis-1100 33 20 to to ADP fcis-1100 33 21 the the DET fcis-1100 33 22 input input NOUN fcis-1100 33 23 of of ADP fcis-1100 33 24 the the DET fcis-1100 33 25 first first ADJ fcis-1100 33 26 layer layer NOUN fcis-1100 33 27 and and CCONJ fcis-1100 33 28 f(x f(x PROPN fcis-1100 33 29 ) ) PUNCT fcis-1100 33 30 represents represent VERB fcis-1100 33 31 the the DET fcis-1100 33 32 residual residual ADJ fcis-1100 33 33 function function NOUN fcis-1100 33 34 , , PUNCT fcis-1100 33 35 then then ADV fcis-1100 33 36 the the DET fcis-1100 33 37 actual actual ADJ fcis-1100 33 38 mapping mapping NOUN fcis-1100 33 39 relationship relationship NOUN fcis-1100 33 40 can can AUX fcis-1100 33 41 be be AUX fcis-1100 33 42 expressed express VERB fcis-1100 33 43 as as ADP fcis-1100 33 44 : : PUNCT fcis-1100 33 45 𝐻(π‘₯ 𝐻(π‘₯ NUM fcis-1100 33 46 ) ) PUNCT fcis-1100 33 47 = = PRON fcis-1100 33 48 𝐹(π‘₯ 𝐹(π‘₯ X fcis-1100 33 49 ) ) PUNCT fcis-1100 34 1 + + CCONJ fcis-1100 34 2 π‘₯. π‘₯. ADJ fcis-1100 34 3 ( ( PUNCT fcis-1100 34 4 1 1 X fcis-1100 34 5 ) ) PUNCT fcis-1100 34 6 the the DET fcis-1100 34 7 advantage advantage NOUN fcis-1100 34 8 of of ADP fcis-1100 34 9 using use VERB fcis-1100 34 10 residual residual ADJ fcis-1100 34 11 block block NOUN fcis-1100 34 12 is be AUX fcis-1100 34 13 that that SCONJ fcis-1100 34 14 there there PRON fcis-1100 34 15 are be VERB fcis-1100 34 16 no no DET fcis-1100 34 17 redundant redundant ADJ fcis-1100 34 18 parameters parameter NOUN fcis-1100 34 19 and and CCONJ fcis-1100 34 20 the the DET fcis-1100 34 21 computational computational ADJ fcis-1100 34 22 complexity complexity NOUN fcis-1100 34 23 will will AUX fcis-1100 34 24 not not PART fcis-1100 34 25 be be AUX fcis-1100 34 26 increased increase VERB fcis-1100 34 27 . . PUNCT fcis-1100 35 1 the the DET fcis-1100 35 2 following follow VERB fcis-1100 35 3 figure figure NOUN fcis-1100 35 4 is be AUX fcis-1100 35 5 the the DET fcis-1100 35 6 most most ADV fcis-1100 35 7 basic basic ADJ fcis-1100 35 8 residual residual ADJ fcis-1100 35 9 learning learn VERB fcis-1100 35 10 unit unit NOUN fcis-1100 35 11 whose whose DET fcis-1100 35 12 idea idea NOUN fcis-1100 35 13 is be AUX fcis-1100 35 14 to to PART fcis-1100 35 15 assume assume VERB fcis-1100 35 16 that that SCONJ fcis-1100 35 17 there there PRON fcis-1100 35 18 is be VERB fcis-1100 35 19 identity identity NOUN fcis-1100 35 20 mapping mapping NOUN fcis-1100 35 21 between between ADP fcis-1100 35 22 models model NOUN fcis-1100 35 23 , , PUNCT fcis-1100 35 24 namely namely ADV fcis-1100 35 25 , , PUNCT fcis-1100 35 26 to to PART fcis-1100 35 27 solve solve VERB fcis-1100 35 28 the the DET fcis-1100 35 29 identity identity NOUN fcis-1100 35 30 mapping mapping NOUN fcis-1100 35 31 function function NOUN fcis-1100 35 32 . . PUNCT fcis-1100 36 1 because because SCONJ fcis-1100 36 2 it it PRON fcis-1100 36 3 is be AUX fcis-1100 36 4 difficult difficult ADJ fcis-1100 36 5 to to PART fcis-1100 36 6 obtain obtain VERB fcis-1100 36 7 h(x h(x PROPN fcis-1100 36 8 ) ) PUNCT fcis-1100 36 9 directly directly ADV fcis-1100 36 10 , , PUNCT fcis-1100 36 11 the the DET fcis-1100 36 12 residual residual ADJ fcis-1100 36 13 unit unit NOUN fcis-1100 36 14 is be AUX fcis-1100 36 15 used use VERB fcis-1100 36 16 to to PART fcis-1100 36 17 pass pass VERB fcis-1100 36 18 through through ADP fcis-1100 36 19 the the DET fcis-1100 36 20 lines line NOUN fcis-1100 36 21 of of ADP fcis-1100 36 22 short short ADJ fcis-1100 36 23 - - PUNCT fcis-1100 36 24 cut cut NOUN fcis-1100 36 25 connections connection NOUN fcis-1100 36 26 . . PUNCT fcis-1100 37 1 formula formula NOUN fcis-1100 37 2 ( ( PUNCT fcis-1100 37 3 1 1 X fcis-1100 37 4 ) ) PUNCT fcis-1100 37 5 can can AUX fcis-1100 37 6 be be AUX fcis-1100 37 7 realized realize VERB fcis-1100 37 8 by by ADP fcis-1100 37 9 adding add VERB fcis-1100 37 10 a a DET fcis-1100 37 11 feed feed NOUN fcis-1100 37 12 - - PUNCT fcis-1100 37 13 forward forward NOUN fcis-1100 37 14 neural neural ADJ fcis-1100 37 15 network network NOUN fcis-1100 37 16 of of ADP fcis-1100 37 17 short short ADJ fcis-1100 37 18 - - PUNCT fcis-1100 37 19 cut cut NOUN fcis-1100 37 20 connections connection NOUN fcis-1100 37 21 . . PUNCT fcis-1100 38 1 in in ADP fcis-1100 38 2 other other ADJ fcis-1100 38 3 words word NOUN fcis-1100 38 4 , , PUNCT fcis-1100 38 5 the the DET fcis-1100 38 6 input input NOUN fcis-1100 38 7 of of ADP fcis-1100 38 8 each each DET fcis-1100 38 9 layer layer NOUN fcis-1100 38 10 is be AUX fcis-1100 38 11 the the DET fcis-1100 38 12 superposition superposition NOUN fcis-1100 38 13 of of ADP fcis-1100 38 14 mapping mapping NOUN fcis-1100 38 15 and and CCONJ fcis-1100 38 16 input input NOUN fcis-1100 38 17 instead instead ADV fcis-1100 38 18 of of ADP fcis-1100 38 19 the the DET fcis-1100 38 20 input input NOUN fcis-1100 38 21 mapping mapping NOUN fcis-1100 38 22 of of ADP fcis-1100 38 23 a a DET fcis-1100 38 24 traditional traditional ADJ fcis-1100 38 25 neural neural ADJ fcis-1100 38 26 network network NOUN fcis-1100 38 27 . . PUNCT fcis-1100 39 1 2.2 2.2 NUM fcis-1100 39 2 . . PUNCT fcis-1100 39 3 batch batch NOUN fcis-1100 39 4 normalization normalization NOUN fcis-1100 39 5 batch batch NOUN fcis-1100 39 6 normalization normalization NOUN fcis-1100 39 7 , , PUNCT fcis-1100 39 8 also also ADV fcis-1100 39 9 known know VERB fcis-1100 39 10 as as ADP fcis-1100 39 11 the the DET fcis-1100 39 12 bn bn NOUN fcis-1100 39 13 layer layer NOUN fcis-1100 39 14 , , PUNCT fcis-1100 39 15 is be AUX fcis-1100 39 16 a a DET fcis-1100 39 17 data data NOUN fcis-1100 39 18 pre pre ADJ fcis-1100 39 19 - - ADJ fcis-1100 39 20 processing processing ADJ fcis-1100 39 21 method method NOUN fcis-1100 39 22 . . PUNCT fcis-1100 40 1 after after SCONJ fcis-1100 40 2 the the DET fcis-1100 40 3 input input NOUN fcis-1100 40 4 data data NOUN fcis-1100 40 5 is be AUX fcis-1100 40 6 normalized normalize VERB fcis-1100 40 7 , , PUNCT fcis-1100 40 8 the the DET fcis-1100 40 9 output output NOUN fcis-1100 40 10 range range NOUN fcis-1100 40 11 is be AUX fcis-1100 40 12 between between ADP fcis-1100 40 13 0 0 NUM fcis-1100 40 14 - - SYM fcis-1100 40 15 1 1 NUM fcis-1100 40 16 . . PUNCT fcis-1100 40 17 bn bn NOUN fcis-1100 40 18 layer layer NOUN fcis-1100 40 19 eliminates eliminate VERB fcis-1100 40 20 the the DET fcis-1100 40 21 differences difference NOUN fcis-1100 40 22 , , PUNCT fcis-1100 40 23 reduces reduce VERB fcis-1100 40 24 the the DET fcis-1100 40 25 interference interference NOUN fcis-1100 40 26 of of ADP fcis-1100 40 27 useless useless ADJ fcis-1100 40 28 data datum NOUN fcis-1100 40 29 , , PUNCT fcis-1100 40 30 and and CCONJ fcis-1100 40 31 speeds speed VERB fcis-1100 40 32 up up ADP fcis-1100 40 33 network network NOUN fcis-1100 40 34 convergence convergence NOUN fcis-1100 40 35 . . PUNCT fcis-1100 41 1 2.3 2.3 NUM fcis-1100 41 2 . . PUNCT fcis-1100 41 3 attention attention NOUN fcis-1100 41 4 mechanism mechanism NOUN fcis-1100 41 5 the the DET fcis-1100 41 6 se se PROPN fcis-1100 41 7 module module NOUN fcis-1100 41 8 proposed propose VERB fcis-1100 41 9 by by ADP fcis-1100 41 10 cvpr cvpr NOUN fcis-1100 41 11 in in ADP fcis-1100 41 12 2017 2017 NUM fcis-1100 41 13 only only ADV fcis-1100 41 14 solved solve VERB fcis-1100 41 15 the the DET fcis-1100 41 16 shortcomings shortcoming NOUN fcis-1100 41 17 of of ADP fcis-1100 41 18 traditional traditional ADJ fcis-1100 41 19 convolution convolution NOUN fcis-1100 41 20 from from ADP fcis-1100 41 21 the the DET fcis-1100 41 22 perspective perspective NOUN fcis-1100 41 23 23 23 NUM fcis-1100 41 24 of of ADP fcis-1100 41 25 channel channel NOUN fcis-1100 41 26 , , PUNCT fcis-1100 41 27 not not PART fcis-1100 41 28 from from ADP fcis-1100 41 29 the the DET fcis-1100 41 30 perspective perspective NOUN fcis-1100 41 31 of of ADP fcis-1100 41 32 space space NOUN fcis-1100 41 33 . . PUNCT fcis-1100 42 1 in in ADP fcis-1100 42 2 2018 2018 NUM fcis-1100 42 3 , , PUNCT fcis-1100 42 4 the the DET fcis-1100 42 5 author author NOUN fcis-1100 42 6 of of ADP fcis-1100 42 7 cvpr cvpr NOUN fcis-1100 42 8 proposed propose VERB fcis-1100 42 9 the the DET fcis-1100 42 10 sse sse NOUN fcis-1100 42 11 attention attention NOUN fcis-1100 42 12 mechanism mechanism NOUN fcis-1100 42 13 based base VERB fcis-1100 42 14 on on ADP fcis-1100 42 15 the the DET fcis-1100 42 16 se se PROPN fcis-1100 42 17 attention attention NOUN fcis-1100 42 18 mechanism mechanism NOUN fcis-1100 42 19 to to PART fcis-1100 42 20 solve solve VERB fcis-1100 42 21 the the DET fcis-1100 42 22 shortcomings shortcoming NOUN fcis-1100 42 23 of of ADP fcis-1100 42 24 convolution convolution NOUN fcis-1100 42 25 from from ADP fcis-1100 42 26 the the DET fcis-1100 42 27 perspective perspective NOUN fcis-1100 42 28 of of ADP fcis-1100 42 29 space space NOUN fcis-1100 42 30 . . PUNCT fcis-1100 43 1 x x PUNCT fcis-1100 44 1 w w NOUN fcis-1100 44 2 h h PROPN fcis-1100 44 3 w w PROPN fcis-1100 44 4 hc hc PROPN fcis-1100 44 5 1x1 1x1 NUM fcis-1100 44 6 figure figure NOUN fcis-1100 44 7 1 1 NUM fcis-1100 44 8 . . PUNCT fcis-1100 45 1 sse sse NOUN fcis-1100 45 2 attention attention NOUN fcis-1100 45 3 mechanism mechanism NOUN fcis-1100 45 4 the the DET fcis-1100 45 5 difference difference NOUN fcis-1100 45 6 between between ADP fcis-1100 45 7 the the DET fcis-1100 45 8 two two NUM fcis-1100 45 9 attention attention NOUN fcis-1100 45 10 mechanisms mechanism NOUN fcis-1100 45 11 is be AUX fcis-1100 45 12 their their PRON fcis-1100 45 13 different different ADJ fcis-1100 45 14 operations operation NOUN fcis-1100 45 15 on on ADP fcis-1100 45 16 dimensionality dimensionality NOUN fcis-1100 45 17 reduction reduction NOUN fcis-1100 45 18 . . PUNCT fcis-1100 46 1 the the DET fcis-1100 46 2 se se PROPN fcis-1100 46 3 attention attention NOUN fcis-1100 46 4 mechanism mechanism NOUN fcis-1100 46 5 uses use VERB fcis-1100 46 6 global global ADJ fcis-1100 46 7 average average ADJ fcis-1100 46 8 pooling pooling NOUN fcis-1100 46 9 for for ADP fcis-1100 46 10 dimensionality dimensionality NOUN fcis-1100 46 11 reduction reduction NOUN fcis-1100 46 12 , , PUNCT fcis-1100 46 13 while while SCONJ fcis-1100 46 14 the the DET fcis-1100 46 15 sse sse NOUN fcis-1100 46 16 attention attention NOUN fcis-1100 46 17 mechanism mechanism NOUN fcis-1100 46 18 uses use VERB fcis-1100 46 19 conv conv ADJ fcis-1100 46 20 1 1 NUM fcis-1100 46 21 * * SYM fcis-1100 46 22 1 1 NUM fcis-1100 46 23 for for ADP fcis-1100 46 24 dimensionality dimensionality NOUN fcis-1100 46 25 reduction reduction NOUN fcis-1100 46 26 . . PUNCT fcis-1100 47 1 however however ADV fcis-1100 47 2 , , PUNCT fcis-1100 47 3 their their PRON fcis-1100 47 4 similarity similarity NOUN fcis-1100 47 5 is be AUX fcis-1100 47 6 that that SCONJ fcis-1100 47 7 they they PRON fcis-1100 47 8 all all PRON fcis-1100 47 9 use use VERB fcis-1100 47 10 the the DET fcis-1100 47 11 sigmoid sigmoid NOUN fcis-1100 47 12 function function NOUN fcis-1100 47 13 activation activation NOUN fcis-1100 47 14 to to PART fcis-1100 47 15 compress compress VERB fcis-1100 47 16 the the DET fcis-1100 47 17 weight weight NOUN fcis-1100 47 18 between between ADP fcis-1100 47 19 0 0 NUM fcis-1100 47 20 and and CCONJ fcis-1100 47 21 1 1 NUM fcis-1100 47 22 , , PUNCT fcis-1100 47 23 which which PRON fcis-1100 47 24 is be AUX fcis-1100 47 25 convenient convenient ADJ fcis-1100 47 26 for for ADP fcis-1100 47 27 multiplying multiply VERB fcis-1100 47 28 and and CCONJ fcis-1100 47 29 stacking stacking NOUN fcis-1100 47 30 with with ADP fcis-1100 47 31 the the DET fcis-1100 47 32 original original ADJ fcis-1100 47 33 feature feature NOUN fcis-1100 47 34 . . PUNCT fcis-1100 48 1 input input NOUN fcis-1100 48 2 batchnorm batchnorm NOUN fcis-1100 48 3 3x3 3x3 NUM fcis-1100 48 4 conv conv ADJ fcis-1100 48 5 , , PUNCT fcis-1100 48 6 s=1 s=1 ADP fcis-1100 48 7 batchnorm batchnorm NOUN fcis-1100 48 8 prelu prelu NOUN fcis-1100 48 9 3x3 3x3 NUM fcis-1100 48 10 conv conv ADJ fcis-1100 48 11 , , PUNCT fcis-1100 48 12 s=2 s=2 DET fcis-1100 48 13 batchnorm batchnorm NOUN fcis-1100 48 14 ssemodule ssemodule NOUN fcis-1100 48 15 output output NOUN fcis-1100 48 16 figure figure NOUN fcis-1100 48 17 2 2 NUM fcis-1100 48 18 . . PUNCT fcis-1100 48 19 network network NOUN fcis-1100 48 20 structure structure NOUN fcis-1100 48 21 3 3 NUM fcis-1100 48 22 . . PUNCT fcis-1100 48 23 improved improve VERB fcis-1100 48 24 loss loss NOUN fcis-1100 48 25 function function NOUN fcis-1100 48 26 3.1 3.1 NUM fcis-1100 48 27 . . PUNCT fcis-1100 48 28 softmax softmax NOUN fcis-1100 48 29 function function PROPN fcis-1100 48 30 sigmoid sigmoid NOUN fcis-1100 48 31 [ [ X fcis-1100 48 32 13 13 NUM fcis-1100 48 33 ] ] PUNCT fcis-1100 48 34 function function NOUN fcis-1100 48 35 , , PUNCT fcis-1100 48 36 also also ADV fcis-1100 48 37 called call VERB fcis-1100 48 38 logistic logistic ADJ fcis-1100 48 39 function function NOUN fcis-1100 48 40 , , PUNCT fcis-1100 48 41 is be AUX fcis-1100 48 42 a a DET fcis-1100 48 43 binary binary ADJ fcis-1100 48 44 classification classification NOUN fcis-1100 48 45 problem problem NOUN fcis-1100 48 46 . . PUNCT fcis-1100 49 1 the the DET fcis-1100 49 2 logistic logistic ADJ fcis-1100 49 3 function function NOUN fcis-1100 49 4 fails fail VERB fcis-1100 49 5 to to PART fcis-1100 49 6 meet meet VERB fcis-1100 49 7 the the DET fcis-1100 49 8 requirement requirement NOUN fcis-1100 49 9 in in ADP fcis-1100 49 10 the the DET fcis-1100 49 11 case case NOUN fcis-1100 49 12 of of ADP fcis-1100 49 13 multi multi NOUN fcis-1100 49 14 - - NOUN fcis-1100 49 15 classification classification NOUN fcis-1100 49 16 of of ADP fcis-1100 49 17 face face NOUN fcis-1100 49 18 images image NOUN fcis-1100 49 19 . . PUNCT fcis-1100 50 1 softmax[14 softmax[14 PROPN fcis-1100 50 2 ] ] PUNCT fcis-1100 50 3 function function NOUN fcis-1100 50 4 is be AUX fcis-1100 50 5 usually usually ADV fcis-1100 50 6 viewed view VERB fcis-1100 50 7 as as ADP fcis-1100 50 8 the the DET fcis-1100 50 9 last last ADJ fcis-1100 50 10 classifier classifier NOUN fcis-1100 50 11 in in ADP fcis-1100 50 12 cnn cnn PROPN fcis-1100 50 13 for for ADP fcis-1100 50 14 multi multi ADJ fcis-1100 50 15 - - ADJ fcis-1100 50 16 classification classification ADJ fcis-1100 50 17 tasks task NOUN fcis-1100 50 18 . . PUNCT fcis-1100 51 1 softmax softmax NOUN fcis-1100 51 2 loss loss NOUN fcis-1100 51 3 function function NOUN fcis-1100 51 4 can can AUX fcis-1100 51 5 ensure ensure VERB fcis-1100 51 6 a a DET fcis-1100 51 7 good good ADJ fcis-1100 51 8 separability separability NOUN fcis-1100 51 9 between between ADP fcis-1100 51 10 classes class NOUN fcis-1100 51 11 . . PUNCT fcis-1100 52 1 however however ADV fcis-1100 52 2 , , PUNCT fcis-1100 52 3 the the DET fcis-1100 52 4 within within ADP fcis-1100 52 5 - - PUNCT fcis-1100 52 6 class class NOUN fcis-1100 52 7 distance distance NOUN fcis-1100 52 8 of of ADP fcis-1100 52 9 features feature NOUN fcis-1100 52 10 is be AUX fcis-1100 52 11 scattered scatter VERB fcis-1100 52 12 in in ADP fcis-1100 52 13 a a DET fcis-1100 52 14 large large ADJ fcis-1100 52 15 range range NOUN fcis-1100 52 16 , , PUNCT fcis-1100 52 17 in in ADP fcis-1100 52 18 - - PUNCT fcis-1100 52 19 class class NOUN fcis-1100 52 20 features feature NOUN fcis-1100 52 21 are be AUX fcis-1100 52 22 not not PART fcis-1100 52 23 compact compact ADJ fcis-1100 52 24 enough enough ADV fcis-1100 52 25 , , PUNCT fcis-1100 52 26 and and CCONJ fcis-1100 52 27 the the DET fcis-1100 52 28 distance distance NOUN fcis-1100 52 29 between between ADP fcis-1100 52 30 some some DET fcis-1100 52 31 in in ADP fcis-1100 52 32 - - PUNCT fcis-1100 52 33 class class NOUN fcis-1100 52 34 features feature NOUN fcis-1100 52 35 is be AUX fcis-1100 52 36 even even ADV fcis-1100 52 37 longer long ADJ fcis-1100 52 38 than than ADP fcis-1100 52 39 that that PRON fcis-1100 52 40 between between ADP fcis-1100 52 41 classes class NOUN fcis-1100 52 42 . . PUNCT fcis-1100 53 1 the the DET fcis-1100 53 2 function function NOUN fcis-1100 53 3 is be AUX fcis-1100 53 4 expressed express VERB fcis-1100 53 5 as as SCONJ fcis-1100 53 6 follows follow VERB fcis-1100 53 7 : : PUNCT fcis-1100 54 1 losssoftmax losssoftmax NOUN fcis-1100 54 2 = = PUNCT fcis-1100 54 3 βˆ’ βˆ’ PROPN fcis-1100 54 4 1 1 NUM fcis-1100 54 5 𝑁 𝑁 PROPN fcis-1100 54 6 βˆ‘ βˆ‘ PUNCT fcis-1100 54 7 π‘™π‘œπ‘” π‘™π‘œπ‘” X fcis-1100 54 8 ( ( PUNCT fcis-1100 54 9 𝑒 𝑒 PROPN fcis-1100 54 10 𝑀𝑦𝑖 𝑀𝑦𝑖 ADV fcis-1100 54 11 𝑇 𝑇 PROPN fcis-1100 54 12 π‘₯𝑖+𝑏𝑦𝑖 π‘₯𝑖+𝑏𝑦𝑖 X fcis-1100 54 13 βˆ‘ βˆ‘ PUNCT fcis-1100 54 14 𝑒 𝑒 VERB fcis-1100 54 15 𝑀𝑦𝑗 𝑀𝑦𝑗 DET fcis-1100 54 16 𝑇 𝑇 PROPN fcis-1100 54 17 π‘₯𝑖+𝑏𝑗𝑛 π‘₯𝑖+𝑏𝑗𝑛 DET fcis-1100 54 18 𝑗=1 𝑗=1 PROPN fcis-1100 54 19 ) ) PUNCT fcis-1100 55 1 𝑁 𝑁 PROPN fcis-1100 55 2 𝑖=1 𝑖=1 PROPN fcis-1100 55 3 ( ( PUNCT fcis-1100 55 4 2 2 NUM fcis-1100 55 5 ) ) PUNCT fcis-1100 55 6 n n CCONJ fcis-1100 55 7 : : PUNCT fcis-1100 55 8 batch batch NOUN fcis-1100 55 9 size size NOUN fcis-1100 55 10 n n CCONJ fcis-1100 55 11 : : PUNCT fcis-1100 55 12 number number NOUN fcis-1100 55 13 of of ADP fcis-1100 55 14 categories category NOUN fcis-1100 55 15 in in ADP fcis-1100 55 16 formula formula NOUN fcis-1100 55 17 ( ( PUNCT fcis-1100 55 18 2 2 NUM fcis-1100 55 19 ) ) PUNCT fcis-1100 55 20 , , PUNCT fcis-1100 55 21 ji ji PROPN fcis-1100 55 22 t t PROPN fcis-1100 55 23 y y PROPN fcis-1100 55 24 bxw bxw PROPN fcis-1100 55 25 j j PROPN fcis-1100 55 26 + + CCONJ fcis-1100 55 27 represents represent VERB fcis-1100 55 28 the the DET fcis-1100 55 29 output output NOUN fcis-1100 55 30 of of ADP fcis-1100 55 31 the the DET fcis-1100 55 32 fully fully ADV fcis-1100 55 33 connected connected ADJ fcis-1100 55 34 layer layer NOUN fcis-1100 55 35 . . PUNCT fcis-1100 56 1 the the DET fcis-1100 56 2 value value NOUN fcis-1100 56 3 of of ADP fcis-1100 56 4 ji ji PROPN fcis-1100 56 5 t t PROPN fcis-1100 56 6 y y PROPN fcis-1100 56 7 bxw bxw PROPN fcis-1100 56 8 j j PROPN fcis-1100 57 1 + + PRON fcis-1100 57 2 has have VERB fcis-1100 57 3 to to PART fcis-1100 57 4 be be AUX fcis-1100 57 5 increased increase VERB fcis-1100 57 6 to to PART fcis-1100 57 7 reduce reduce VERB fcis-1100 57 8 loss loss NOUN fcis-1100 57 9 and and CCONJ fcis-1100 57 10 all all DET fcis-1100 57 11 faces face NOUN fcis-1100 57 12 belonging belong VERB fcis-1100 57 13 to to ADP fcis-1100 57 14 this this DET fcis-1100 57 15 kind kind NOUN fcis-1100 57 16 of of ADP fcis-1100 57 17 sample sample NOUN fcis-1100 57 18 are be AUX fcis-1100 57 19 included include VERB fcis-1100 57 20 in in ADP fcis-1100 57 21 this this DET fcis-1100 57 22 kind kind NOUN fcis-1100 57 23 of of ADP fcis-1100 57 24 decision decision NOUN fcis-1100 57 25 boundary boundary NOUN fcis-1100 57 26 . . PUNCT fcis-1100 58 1 softmax softmax NOUN fcis-1100 58 2 mainly mainly ADV fcis-1100 58 3 takes take VERB fcis-1100 58 4 correct correct ADJ fcis-1100 58 5 classification classification NOUN fcis-1100 58 6 into into ADP fcis-1100 58 7 consideration consideration NOUN fcis-1100 58 8 while while SCONJ fcis-1100 58 9 lacking lack VERB fcis-1100 58 10 constraints constraint NOUN fcis-1100 58 11 on on ADP fcis-1100 58 12 in in ADP fcis-1100 58 13 - - PUNCT fcis-1100 58 14 class class NOUN fcis-1100 58 15 and and CCONJ fcis-1100 58 16 inter inter ADJ fcis-1100 58 17 - - ADJ fcis-1100 58 18 class class ADJ fcis-1100 58 19 distances distance NOUN fcis-1100 58 20 . . PUNCT fcis-1100 59 1 generally generally ADV fcis-1100 59 2 speaking speak VERB fcis-1100 59 3 , , PUNCT fcis-1100 59 4 it it PRON fcis-1100 59 5 is be AUX fcis-1100 59 6 poor poor ADJ fcis-1100 59 7 in in ADP fcis-1100 59 8 classifying classify VERB fcis-1100 59 9 similar similar ADJ fcis-1100 59 10 faces face NOUN fcis-1100 59 11 . . PUNCT fcis-1100 60 1 wen wen PROPN fcis-1100 60 2 yandong[15 yandong[15 PROPN fcis-1100 60 3 ] ] X fcis-1100 60 4 , , PUNCT fcis-1100 60 5 the the DET fcis-1100 60 6 author author NOUN fcis-1100 60 7 of of ADP fcis-1100 60 8 this this DET fcis-1100 60 9 thesis thesis NOUN fcis-1100 60 10 , , PUNCT fcis-1100 60 11 finds find VERB fcis-1100 60 12 that that SCONJ fcis-1100 60 13 there there PRON fcis-1100 60 14 is be VERB fcis-1100 60 15 still still ADV fcis-1100 60 16 a a DET fcis-1100 60 17 large large ADJ fcis-1100 60 18 in in ADP fcis-1100 60 19 - - PUNCT fcis-1100 60 20 class class NOUN fcis-1100 60 21 distance distance NOUN fcis-1100 60 22 in in ADP fcis-1100 60 23 traditional traditional ADJ fcis-1100 60 24 softmax softmax NOUN fcis-1100 60 25 through through ADP fcis-1100 60 26 experiments experiment NOUN fcis-1100 60 27 . . PUNCT fcis-1100 61 1 that that PRON fcis-1100 61 2 is be AUX fcis-1100 61 3 to to PART fcis-1100 61 4 say say VERB fcis-1100 61 5 , , PUNCT fcis-1100 61 6 by by ADP fcis-1100 61 7 improving improve VERB fcis-1100 61 8 loss loss NOUN fcis-1100 61 9 function function NOUN fcis-1100 61 10 , , PUNCT fcis-1100 61 11 it it PRON fcis-1100 61 12 can can AUX fcis-1100 61 13 add add VERB fcis-1100 61 14 constraints constraint NOUN fcis-1100 61 15 on on ADP fcis-1100 61 16 in in ADP fcis-1100 61 17 - - PUNCT fcis-1100 61 18 class class NOUN fcis-1100 61 19 distance distance NOUN fcis-1100 61 20 and and CCONJ fcis-1100 61 21 improve improve VERB fcis-1100 61 22 network network NOUN fcis-1100 61 23 performance performance NOUN fcis-1100 61 24 as as ADV fcis-1100 61 25 well well ADV fcis-1100 61 26 . . PUNCT fcis-1100 62 1 3.2 3.2 NUM fcis-1100 62 2 . . PUNCT fcis-1100 62 3 improved improve VERB fcis-1100 62 4 softmax softmax NOUN fcis-1100 62 5 function function NOUN fcis-1100 62 6 because because SCONJ fcis-1100 62 7 softmax softmax NOUN fcis-1100 62 8 mainly mainly ADV fcis-1100 62 9 considers consider VERB fcis-1100 62 10 whether whether SCONJ fcis-1100 62 11 the the DET fcis-1100 62 12 samples sample NOUN fcis-1100 62 13 can can AUX fcis-1100 62 14 be be AUX fcis-1100 62 15 correctly correctly ADV fcis-1100 62 16 classified classified ADJ fcis-1100 62 17 , , PUNCT fcis-1100 62 18 but but CCONJ fcis-1100 62 19 lacks lack VERB fcis-1100 62 20 the the DET fcis-1100 62 21 distance distance NOUN fcis-1100 62 22 limitation limitation NOUN fcis-1100 62 23 within within ADP fcis-1100 62 24 and and CCONJ fcis-1100 62 25 between between ADP fcis-1100 62 26 classes class NOUN fcis-1100 63 1 , , PUNCT fcis-1100 63 2 this this DET fcis-1100 63 3 paper paper NOUN fcis-1100 63 4 raises raise VERB fcis-1100 63 5 an an DET fcis-1100 63 6 arcface[16 arcface[16 NOUN fcis-1100 63 7 ] ] PUNCT fcis-1100 63 8 . . PUNCT fcis-1100 64 1 arcface arcface NOUN fcis-1100 64 2 improves improve VERB fcis-1100 64 3 the the DET fcis-1100 64 4 recognition recognition NOUN fcis-1100 64 5 capability capability NOUN fcis-1100 64 6 of of ADP fcis-1100 64 7 the the DET fcis-1100 64 8 training training NOUN fcis-1100 64 9 model model NOUN fcis-1100 64 10 by by ADP fcis-1100 64 11 reducing reduce VERB fcis-1100 64 12 the the DET fcis-1100 64 13 in in ADP fcis-1100 64 14 - - PUNCT fcis-1100 64 15 class class NOUN fcis-1100 64 16 distance distance NOUN fcis-1100 64 17 and and CCONJ fcis-1100 64 18 increasing increase VERB fcis-1100 64 19 the the DET fcis-1100 64 20 inter inter ADJ fcis-1100 64 21 - - ADJ fcis-1100 64 22 class class ADJ fcis-1100 64 23 distance distance NOUN fcis-1100 64 24 . . PUNCT fcis-1100 65 1 by by ADP fcis-1100 65 2 observing observe VERB fcis-1100 65 3 the the DET fcis-1100 65 4 relationship relationship NOUN fcis-1100 65 5 between between ADP fcis-1100 65 6 weight weight NOUN fcis-1100 65 7 and and CCONJ fcis-1100 65 8 class class NOUN fcis-1100 65 9 center center NOUN fcis-1100 65 10 , , PUNCT fcis-1100 65 11 sphereface[17 sphereface[17 PROPN fcis-1100 65 12 ] ] X fcis-1100 65 13 , , PUNCT fcis-1100 65 14 an an DET fcis-1100 65 15 epoch epoch NOUN fcis-1100 65 16 - - PUNCT fcis-1100 65 17 making make VERB fcis-1100 65 18 paper paper NOUN fcis-1100 65 19 is be AUX fcis-1100 65 20 put put VERB fcis-1100 65 21 forward forward ADV fcis-1100 65 22 . . PUNCT fcis-1100 66 1 moreover moreover ADV fcis-1100 66 2 , , PUNCT fcis-1100 66 3 the the DET fcis-1100 66 4 important important ADJ fcis-1100 66 5 concept concept NOUN fcis-1100 66 6 of of ADP fcis-1100 66 7 angular angular ADJ fcis-1100 66 8 margin margin NOUN fcis-1100 66 9 is be AUX fcis-1100 66 10 introduced introduce VERB fcis-1100 66 11 . . PUNCT fcis-1100 67 1 however however ADV fcis-1100 67 2 , , PUNCT fcis-1100 67 3 some some DET fcis-1100 67 4 approximate approximate ADJ fcis-1100 67 5 calculations calculation NOUN fcis-1100 67 6 needed need VERB fcis-1100 67 7 will will AUX fcis-1100 67 8 lead lead VERB fcis-1100 67 9 to to ADP fcis-1100 67 10 training training NOUN fcis-1100 67 11 instability instability NOUN fcis-1100 67 12 . . PUNCT fcis-1100 68 1 the the DET fcis-1100 68 2 arccosine arccosine NOUN fcis-1100 68 3 function function NOUN fcis-1100 68 4 is be AUX fcis-1100 68 5 used use VERB fcis-1100 68 6 to to PART fcis-1100 68 7 calculate calculate VERB fcis-1100 68 8 the the DET fcis-1100 68 9 angle angle NOUN fcis-1100 68 10 between between ADP fcis-1100 68 11 the the DET fcis-1100 68 12 current current ADJ fcis-1100 68 13 feature feature NOUN fcis-1100 68 14 and and CCONJ fcis-1100 68 15 the the DET fcis-1100 68 16 weight weight NOUN fcis-1100 68 17 and and CCONJ fcis-1100 68 18 an an DET fcis-1100 68 19 extra extra ADJ fcis-1100 68 20 angular angular ADJ fcis-1100 68 21 margin margin NOUN fcis-1100 68 22 m m VERB fcis-1100 68 23 is be AUX fcis-1100 68 24 added add VERB fcis-1100 68 25 to to ADP fcis-1100 68 26 the the DET fcis-1100 68 27 target target NOUN fcis-1100 68 28 angle angle NOUN fcis-1100 68 29 . . PUNCT fcis-1100 69 1 to to PART fcis-1100 69 2 reduce reduce VERB fcis-1100 69 3 the the DET fcis-1100 69 4 computational computational ADJ fcis-1100 69 5 complexity complexity NOUN fcis-1100 69 6 , , PUNCT fcis-1100 69 7 the the DET fcis-1100 69 8 offset offset NOUN fcis-1100 69 9 jb jb PROPN fcis-1100 69 10 = = SYM fcis-1100 69 11 0 0 PROPN fcis-1100 69 12 , , PUNCT fcis-1100 69 13 the the DET fcis-1100 69 14 inner inner ADJ fcis-1100 69 15 product product NOUN fcis-1100 69 16 of of ADP fcis-1100 69 17 weight weight NOUN fcis-1100 69 18 and and CCONJ fcis-1100 69 19 input input NOUN fcis-1100 69 20 features feature NOUN fcis-1100 69 21 is be AUX fcis-1100 69 22 as as SCONJ fcis-1100 69 23 follows follow VERB fcis-1100 69 24 : : PUNCT fcis-1100 69 25 .cos|||||||| .cos|||||||| PUNCT fcis-1100 70 1 jiji jiji PROPN fcis-1100 70 2 t t PROPN fcis-1100 70 3 j j PROPN fcis-1100 70 4 xwxw xwxw VERB fcis-1100 71 1 = = PROPN fcis-1100 71 2 ( ( PUNCT fcis-1100 71 3 3 3 X fcis-1100 71 4 ) ) PUNCT fcis-1100 71 5 regularizing regularize VERB fcis-1100 71 6 the the DET fcis-1100 71 7 weight weight NOUN fcis-1100 71 8 and and CCONJ fcis-1100 71 9 feature feature NOUN fcis-1100 71 10 l2 l2 NOUN fcis-1100 71 11 , , PUNCT fcis-1100 71 12 namely namely ADV fcis-1100 71 13 , , PUNCT fcis-1100 71 14 1|||| 1|||| NUM fcis-1100 71 15 = = PROPN fcis-1100 71 16 jw jw PROPN fcis-1100 71 17 and and CCONJ fcis-1100 71 18 1|||| 1|||| NUM fcis-1100 71 19 = = NOUN fcis-1100 71 20 ix ix ADJ fcis-1100 71 21 , , PUNCT fcis-1100 71 22 and and CCONJ fcis-1100 71 23 j j PROPN fcis-1100 71 24 is be AUX fcis-1100 71 25 the the DET fcis-1100 71 26 angle angle NOUN fcis-1100 71 27 between between ADP fcis-1100 71 28 jw jw PROPN fcis-1100 71 29 and and CCONJ fcis-1100 71 30 ix ix ADV fcis-1100 71 31 , , PUNCT fcis-1100 71 32 therefore therefore ADV fcis-1100 71 33 , , PUNCT fcis-1100 71 34 the the DET fcis-1100 71 35 learned learn VERB fcis-1100 71 36 embedding embed VERB fcis-1100 71 37 features feature NOUN fcis-1100 71 38 are be AUX fcis-1100 71 39 distributed distribute VERB fcis-1100 71 40 on on ADP fcis-1100 71 41 the the DET fcis-1100 71 42 hyper hyper NOUN fcis-1100 71 43 - - NOUN fcis-1100 71 44 sphere sphere NOUN fcis-1100 71 45 with with ADP fcis-1100 71 46 radius radius NOUN fcis-1100 71 47 s. s. PROPN fcis-1100 71 48 since since SCONJ fcis-1100 71 49 the the DET fcis-1100 71 50 embedding embed VERB fcis-1100 71 51 features feature NOUN fcis-1100 71 52 are be AUX fcis-1100 71 53 distributed distribute VERB fcis-1100 71 54 around around ADP fcis-1100 71 55 each each DET fcis-1100 71 56 feature feature NOUN fcis-1100 71 57 center center NOUN fcis-1100 71 58 on on ADP fcis-1100 71 59 the the DET fcis-1100 71 60 hyper hyper NOUN fcis-1100 71 61 - - NOUN fcis-1100 71 62 sphere sphere NOUN fcis-1100 71 63 , , PUNCT fcis-1100 71 64 an an DET fcis-1100 71 65 extra extra ADJ fcis-1100 71 66 m m VERB fcis-1100 71 67 is be AUX fcis-1100 71 68 added add VERB fcis-1100 71 69 between between ADP fcis-1100 71 70 the the DET fcis-1100 71 71 weight weight NOUN fcis-1100 71 72 and and CCONJ fcis-1100 71 73 feature feature NOUN fcis-1100 71 74 . . PUNCT fcis-1100 72 1 arcface arcface NOUN fcis-1100 72 2 loss loss NOUN fcis-1100 72 3 function function NOUN fcis-1100 72 4 is be AUX fcis-1100 72 5 shown show VERB fcis-1100 72 6 as as ADP fcis-1100 72 7 : : PUNCT fcis-1100 72 8 lossarcface lossarcface NOUN fcis-1100 72 9 = = PUNCT fcis-1100 72 10 βˆ’ βˆ’ PROPN fcis-1100 72 11 1 1 NUM fcis-1100 72 12 𝑁 𝑁 PROPN fcis-1100 72 13 βˆ‘ βˆ‘ PUNCT fcis-1100 72 14 π‘™π‘œπ‘” π‘™π‘œπ‘” NOUN fcis-1100 72 15 𝑒 𝑒 PROPN fcis-1100 72 16 𝑠(π‘π‘œπ‘ (πœƒπ‘¦π‘– 𝑠(π‘π‘œπ‘ (πœƒπ‘¦π‘– ADJ fcis-1100 72 17 + + NOUN fcis-1100 72 18 π‘š π‘š NOUN fcis-1100 72 19 ) ) PUNCT fcis-1100 72 20 ) ) PUNCT fcis-1100 73 1 𝑒 𝑒 ADP fcis-1100 73 2 𝑠(π‘π‘œπ‘ (πœƒπ‘¦π‘– 𝑠(π‘π‘œπ‘ (πœƒπ‘¦π‘– PROPN fcis-1100 73 3 + + NOUN fcis-1100 73 4 π‘š π‘š NOUN fcis-1100 73 5 ) ) PUNCT fcis-1100 73 6 ) ) PUNCT fcis-1100 74 1 + + ADP fcis-1100 74 2 βˆ‘ βˆ‘ PROPN fcis-1100 74 3 𝑒 𝑒 PROPN fcis-1100 74 4 𝑠 𝑠 INTJ fcis-1100 74 5 cos cos PROPN fcis-1100 74 6 πœƒπ‘—π‘› πœƒπ‘—π‘› AUX fcis-1100 74 7 𝑗=1,𝑗≠𝑦𝑖 𝑗=1,𝑗≠𝑦𝑖 VERB fcis-1100 74 8 𝑁 𝑁 PROPN fcis-1100 74 9 𝑖=1 𝑖=1 PROPN fcis-1100 74 10 ( ( PUNCT fcis-1100 74 11 4 4 X fcis-1100 74 12 ) ) PUNCT fcis-1100 74 13 formula formula NOUN fcis-1100 74 14 ( ( PUNCT fcis-1100 74 15 4 4 X fcis-1100 74 16 ) ) PUNCT fcis-1100 74 17 enhances enhance VERB fcis-1100 74 18 the the DET fcis-1100 74 19 in in ADP fcis-1100 74 20 - - PUNCT fcis-1100 74 21 class class NOUN fcis-1100 74 22 compactness compactness NOUN fcis-1100 74 23 and and CCONJ fcis-1100 74 24 expands expand VERB fcis-1100 74 25 the the DET fcis-1100 74 26 inner inner ADJ fcis-1100 74 27 - - PUNCT fcis-1100 74 28 class class NOUN fcis-1100 74 29 differences difference NOUN fcis-1100 74 30 almost almost ADV fcis-1100 74 31 simultaneously simultaneously ADV fcis-1100 74 32 . . PUNCT fcis-1100 75 1 improved improved ADJ fcis-1100 75 2 softmax softmax NOUN fcis-1100 75 3 loss loss NOUN fcis-1100 75 4 function function NOUN fcis-1100 75 5 pays pay VERB fcis-1100 75 6 attention attention NOUN fcis-1100 75 7 to to ADP fcis-1100 75 8 maximizing maximize VERB fcis-1100 75 9 the the DET fcis-1100 75 10 classification classification NOUN fcis-1100 75 11 boundary boundary NOUN fcis-1100 75 12 directly directly ADV fcis-1100 75 13 in in ADP fcis-1100 75 14 the the DET fcis-1100 75 15 angular angular ADJ fcis-1100 75 16 space space NOUN fcis-1100 75 17 . . PUNCT fcis-1100 76 1 combined combine VERB fcis-1100 76 2 with with ADP fcis-1100 76 3 cosface[18 cosface[18 PROPN fcis-1100 76 4 ] ] PUNCT fcis-1100 76 5 and and CCONJ fcis-1100 76 6 the the DET fcis-1100 76 7 usage usage NOUN fcis-1100 76 8 of of ADP fcis-1100 76 9 angular angular ADJ fcis-1100 76 10 margin margin NOUN fcis-1100 76 11 m m VERB fcis-1100 76 12 in in ADP fcis-1100 76 13 arcface arcface NOUN fcis-1100 76 14 , , PUNCT fcis-1100 76 15 this this DET fcis-1100 76 16 paper paper NOUN fcis-1100 76 17 points point VERB fcis-1100 76 18 out out ADP fcis-1100 76 19 a a DET fcis-1100 76 20 modified modify VERB fcis-1100 76 21 loss loss NOUN fcis-1100 76 22 function function NOUN fcis-1100 76 23 for for ADP fcis-1100 76 24 arcface arcface NOUN fcis-1100 76 25 loss loss NOUN fcis-1100 76 26 . . PUNCT fcis-1100 77 1 the the DET fcis-1100 77 2 details detail NOUN fcis-1100 77 3 are be AUX fcis-1100 77 4 as as SCONJ fcis-1100 77 5 follows follow VERB fcis-1100 77 6 : : PUNCT fcis-1100 77 7 loss loss NOUN fcis-1100 77 8 = = SYM fcis-1100 77 9 βˆ’ βˆ’ NOUN fcis-1100 77 10 1 1 NUM fcis-1100 78 1 𝑁 𝑁 PROPN fcis-1100 78 2 βˆ‘ βˆ‘ PUNCT fcis-1100 78 3 π‘™π‘œπ‘” π‘™π‘œπ‘” NOUN fcis-1100 78 4 𝑒 𝑒 PROPN fcis-1100 78 5 𝑠(π‘π‘œπ‘ (πœƒπ‘¦π‘– 𝑠(π‘π‘œπ‘ (πœƒπ‘¦π‘– PROPN fcis-1100 78 6 + + NOUN fcis-1100 78 7 π‘š)+𝑐 π‘š)+𝑐 NOUN fcis-1100 78 8 ) ) PUNCT fcis-1100 78 9 𝑒 𝑒 PROPN fcis-1100 78 10 𝑠(π‘π‘œπ‘ (πœƒπ‘¦π‘– 𝑠(π‘π‘œπ‘ (πœƒπ‘¦π‘– PROPN fcis-1100 78 11 + + NOUN fcis-1100 78 12 π‘š)+𝑐 π‘š)+𝑐 NOUN fcis-1100 78 13 ) ) PUNCT fcis-1100 79 1 + + ADP fcis-1100 79 2 βˆ‘ βˆ‘ PROPN fcis-1100 79 3 𝑒 𝑒 PROPN fcis-1100 79 4 𝑠 𝑠 INTJ fcis-1100 79 5 cos cos PROPN fcis-1100 79 6 πœƒπ‘—π‘› πœƒπ‘—π‘› VERB fcis-1100 79 7 𝑗=1 𝑗=1 PROPN fcis-1100 79 8 𝑁 𝑁 PROPN fcis-1100 79 9 𝑖=1 𝑖=1 PROPN fcis-1100 79 10 ( ( PUNCT fcis-1100 79 11 5 5 NUM fcis-1100 79 12 ) ) PUNCT fcis-1100 79 13 s s PART fcis-1100 79 14 : : PUNCT fcis-1100 79 15 radius radius NOUN fcis-1100 79 16 of of ADP fcis-1100 79 17 hypersphere hypersphere PROPN fcis-1100 79 18 m m PROPN fcis-1100 79 19 : : PUNCT fcis-1100 79 20 angle angle NOUN fcis-1100 79 21 margin margin NOUN fcis-1100 79 22 c c NOUN fcis-1100 79 23 : : PUNCT fcis-1100 79 24 cosine cosine NOUN fcis-1100 79 25 distance distance NOUN fcis-1100 79 26 . . PUNCT fcis-1100 80 1 𝑠((π‘π‘œπ‘  𝑠((π‘π‘œπ‘  PROPN fcis-1100 80 2 ( ( PUNCT fcis-1100 80 3 πœƒ1 πœƒ1 VERB fcis-1100 80 4 + + CCONJ fcis-1100 80 5 π‘š π‘š X fcis-1100 80 6 ) ) PUNCT fcis-1100 80 7 + + NUM fcis-1100 80 8 𝑐 𝑐 X fcis-1100 80 9 ) ) PUNCT fcis-1100 80 10 βˆ’ βˆ’ NOUN fcis-1100 80 11 π‘π‘œπ‘ πœƒ2 π‘π‘œπ‘ πœƒ2 NOUN fcis-1100 80 12 ) ) PUNCT fcis-1100 80 13 = = SYM fcis-1100 81 1 0 0 X fcis-1100 81 2 . . PUNCT fcis-1100 82 1 ( ( PUNCT fcis-1100 82 2 6 6 NUM fcis-1100 82 3 ) ) PUNCT fcis-1100 82 4 formula formula NOUN fcis-1100 82 5 ( ( PUNCT fcis-1100 82 6 5 5 NUM fcis-1100 82 7 ) ) PUNCT fcis-1100 82 8 is be AUX fcis-1100 82 9 the the DET fcis-1100 82 10 final final ADJ fcis-1100 82 11 loss loss NOUN fcis-1100 82 12 function function NOUN fcis-1100 82 13 and and CCONJ fcis-1100 82 14 formula formula NOUN fcis-1100 82 15 ( ( PUNCT fcis-1100 82 16 6 6 NUM fcis-1100 82 17 ) ) PUNCT fcis-1100 82 18 is be AUX fcis-1100 82 19 the the DET fcis-1100 82 20 24 24 NUM fcis-1100 82 21 corresponding correspond VERB fcis-1100 82 22 classification classification NOUN fcis-1100 82 23 boundary boundary NOUN fcis-1100 82 24 . . PUNCT fcis-1100 83 1 c c NOUN fcis-1100 83 2 is be AUX fcis-1100 83 3 a a DET fcis-1100 83 4 cosine cosine ADJ fcis-1100 83 5 distance distance NOUN fcis-1100 83 6 with with ADP fcis-1100 83 7 a a DET fcis-1100 83 8 value value NOUN fcis-1100 83 9 of of ADP fcis-1100 83 10 0.3 0.3 NUM fcis-1100 83 11 . . PUNCT fcis-1100 84 1 adding add VERB fcis-1100 84 2 c c NOUN fcis-1100 84 3 can can AUX fcis-1100 84 4 further far ADV fcis-1100 84 5 compress compress VERB fcis-1100 84 6 the the DET fcis-1100 84 7 intraclass intraclass NOUN fcis-1100 84 8 distance distance NOUN fcis-1100 84 9 , , PUNCT fcis-1100 84 10 and and CCONJ fcis-1100 84 11 compressing compress VERB fcis-1100 84 12 the the DET fcis-1100 84 13 intra intra ADJ fcis-1100 84 14 - - ADJ fcis-1100 84 15 class class ADJ fcis-1100 84 16 distance distance NOUN fcis-1100 84 17 is be AUX fcis-1100 84 18 equivalent equivalent ADJ fcis-1100 84 19 to to ADP fcis-1100 84 20 expanding expand VERB fcis-1100 84 21 the the DET fcis-1100 84 22 inter inter ADJ fcis-1100 84 23 - - ADJ fcis-1100 84 24 class class ADJ fcis-1100 84 25 distance distance NOUN fcis-1100 84 26 . . PUNCT fcis-1100 85 1 if if SCONJ fcis-1100 85 2 π‘π‘œπ‘  π‘π‘œπ‘  PROPN fcis-1100 85 3 ( ( PUNCT fcis-1100 85 4 ( ( PUNCT fcis-1100 85 5 πœƒ1 πœƒ1 VERB fcis-1100 85 6 + + CCONJ fcis-1100 85 7 π‘š π‘š X fcis-1100 85 8 ) ) PUNCT fcis-1100 85 9 + + NUM fcis-1100 85 10 𝑐 𝑐 X fcis-1100 85 11 ) ) PUNCT fcis-1100 85 12 > > X fcis-1100 86 1 π‘π‘œπ‘ πœƒ2 π‘π‘œπ‘ πœƒ2 PROPN fcis-1100 86 2 , , PUNCT fcis-1100 86 3 then then ADV fcis-1100 86 4 the the DET fcis-1100 86 5 face face NOUN fcis-1100 86 6 input input NOUN fcis-1100 86 7 belongs belong VERB fcis-1100 86 8 to to ADP fcis-1100 86 9 category category NOUN fcis-1100 86 10 1 1 NUM fcis-1100 86 11 , , PUNCT fcis-1100 86 12 otherwise otherwise ADV fcis-1100 86 13 , , PUNCT fcis-1100 86 14 it it PRON fcis-1100 86 15 belongs belong VERB fcis-1100 86 16 to to ADP fcis-1100 86 17 category category NOUN fcis-1100 86 18 2 2 NUM fcis-1100 86 19 . . NOUN fcis-1100 86 20 4 4 NUM fcis-1100 86 21 . . X fcis-1100 86 22 experimental experimental ADJ fcis-1100 86 23 results result NOUN fcis-1100 86 24 and and CCONJ fcis-1100 86 25 analysis analysis NOUN fcis-1100 86 26 4.1 4.1 NUM fcis-1100 86 27 . . PUNCT fcis-1100 87 1 introduction introduction NOUN fcis-1100 87 2 to to ADP fcis-1100 87 3 data datum NOUN fcis-1100 87 4 set set VERB fcis-1100 87 5 lfw lfw NOUN fcis-1100 87 6 data datum NOUN fcis-1100 87 7 set set VERB fcis-1100 87 8 including include VERB fcis-1100 87 9 a a DET fcis-1100 87 10 total total NOUN fcis-1100 87 11 of of ADP fcis-1100 87 12 13000 13000 NUM fcis-1100 87 13 face face NOUN fcis-1100 87 14 images image NOUN fcis-1100 87 15 of of ADP fcis-1100 87 16 about about ADV fcis-1100 87 17 5000 5000 NUM fcis-1100 87 18 people people NOUN fcis-1100 87 19 . . PUNCT fcis-1100 88 1 casia casia NOUN fcis-1100 88 2 - - PUNCT fcis-1100 88 3 webface[19 webface[19 NOUN fcis-1100 88 4 ] ] PUNCT fcis-1100 88 5 date date NOUN fcis-1100 88 6 set set NOUN fcis-1100 88 7 is be AUX fcis-1100 88 8 a a DET fcis-1100 88 9 largescale largescale ADJ fcis-1100 88 10 face face NOUN fcis-1100 88 11 data datum NOUN fcis-1100 88 12 set set VERB fcis-1100 88 13 published publish VERB fcis-1100 88 14 in in ADP fcis-1100 88 15 2014 2014 NUM fcis-1100 88 16 . . PUNCT fcis-1100 89 1 face face NOUN fcis-1100 89 2 images image NOUN fcis-1100 89 3 collected collect VERB fcis-1100 89 4 from from ADP fcis-1100 89 5 the the DET fcis-1100 89 6 internet internet NOUN fcis-1100 89 7 , , PUNCT fcis-1100 89 8 including include VERB fcis-1100 89 9 494414 494414 NUM fcis-1100 89 10 images image NOUN fcis-1100 89 11 of of ADP fcis-1100 89 12 10575 10575 NUM fcis-1100 89 13 people people NOUN fcis-1100 89 14 , , PUNCT fcis-1100 89 15 are be AUX fcis-1100 89 16 used use VERB fcis-1100 89 17 as as ADP fcis-1100 89 18 the the DET fcis-1100 89 19 training training NOUN fcis-1100 89 20 set set NOUN fcis-1100 89 21 of of ADP fcis-1100 89 22 resnet resnet NOUN fcis-1100 89 23 in in ADP fcis-1100 89 24 this this DET fcis-1100 89 25 experiment experiment NOUN fcis-1100 89 26 . . PUNCT fcis-1100 90 1 there there PRON fcis-1100 90 2 are be VERB fcis-1100 90 3 4000 4000 NUM fcis-1100 90 4 face face NOUN fcis-1100 90 5 images image NOUN fcis-1100 90 6 of of ADP fcis-1100 90 7 100 100 NUM fcis-1100 90 8 people people NOUN fcis-1100 90 9 in in ADP fcis-1100 90 10 the the DET fcis-1100 90 11 ar ar PROPN fcis-1100 90 12 face face NOUN fcis-1100 90 13 database database NOUN fcis-1100 90 14 . . PUNCT fcis-1100 91 1 it it PRON fcis-1100 91 2 not not PART fcis-1100 91 3 only only ADV fcis-1100 91 4 contains contain VERB fcis-1100 91 5 four four NUM fcis-1100 91 6 basic basic ADJ fcis-1100 91 7 expressions expression NOUN fcis-1100 91 8 like like ADP fcis-1100 91 9 nature nature NOUN fcis-1100 91 10 , , PUNCT fcis-1100 91 11 joy joy NOUN fcis-1100 91 12 , , PUNCT fcis-1100 91 13 anger anger NOUN fcis-1100 91 14 , , PUNCT fcis-1100 91 15 and and CCONJ fcis-1100 91 16 surprise surprise NOUN fcis-1100 91 17 but but CCONJ fcis-1100 91 18 also also ADV fcis-1100 91 19 includes include VERB fcis-1100 91 20 face face NOUN fcis-1100 91 21 images image NOUN fcis-1100 91 22 under under ADP fcis-1100 91 23 various various ADJ fcis-1100 91 24 illumination illumination NOUN fcis-1100 91 25 conditions condition NOUN fcis-1100 91 26 . . PUNCT fcis-1100 92 1 besides besides SCONJ fcis-1100 92 2 , , PUNCT fcis-1100 92 3 it it PRON fcis-1100 92 4 also also ADV fcis-1100 92 5 contains contain VERB fcis-1100 92 6 partially partially ADV fcis-1100 92 7 occluded occlude VERB fcis-1100 92 8 face face VERB fcis-1100 92 9 images image NOUN fcis-1100 92 10 with with ADP fcis-1100 92 11 sunglasses sunglass NOUN fcis-1100 92 12 or or CCONJ fcis-1100 92 13 scarves scarf NOUN fcis-1100 92 14 . . PUNCT fcis-1100 93 1 normal normal ADJ fcis-1100 93 2 images image NOUN fcis-1100 93 3 and and CCONJ fcis-1100 93 4 partially partially ADV fcis-1100 93 5 occluded occlude VERB fcis-1100 93 6 images image NOUN fcis-1100 93 7 with with ADP fcis-1100 93 8 sunglasses sunglass NOUN fcis-1100 93 9 or or CCONJ fcis-1100 93 10 scarves scarf NOUN fcis-1100 93 11 of of ADP fcis-1100 93 12 100 100 NUM fcis-1100 93 13 people people NOUN fcis-1100 93 14 are be AUX fcis-1100 93 15 selected select VERB fcis-1100 93 16 from from ADP fcis-1100 93 17 the the DET fcis-1100 93 18 ar ar PROPN fcis-1100 93 19 database database NOUN fcis-1100 93 20 . . PUNCT fcis-1100 94 1 agedb agedb PROPN fcis-1100 94 2 [ [ X fcis-1100 94 3 20 20 NUM fcis-1100 94 4 ] ] PUNCT fcis-1100 94 5 is be AUX fcis-1100 94 6 a a DET fcis-1100 94 7 dataset dataset NOUN fcis-1100 94 8 of of ADP fcis-1100 94 9 different different ADJ fcis-1100 94 10 ages age NOUN fcis-1100 94 11 , , PUNCT fcis-1100 94 12 containing contain VERB fcis-1100 94 13 about about ADV fcis-1100 94 14 16,000 16,000 NUM fcis-1100 94 15 images image NOUN fcis-1100 94 16 of of ADP fcis-1100 94 17 human human ADJ fcis-1100 94 18 faces face NOUN fcis-1100 94 19 ranging range VERB fcis-1100 94 20 in in ADP fcis-1100 94 21 age age NOUN fcis-1100 94 22 from from ADP fcis-1100 94 23 1 1 NUM fcis-1100 94 24 to to ADP fcis-1100 94 25 101 101 NUM fcis-1100 94 26 , , PUNCT fcis-1100 94 27 with with ADP fcis-1100 94 28 an an DET fcis-1100 94 29 average average ADJ fcis-1100 94 30 age age NOUN fcis-1100 94 31 of of ADP fcis-1100 94 32 30 30 NUM fcis-1100 94 33 for for ADP fcis-1100 94 34 each each DET fcis-1100 94 35 face face NOUN fcis-1100 94 36 . . PUNCT fcis-1100 95 1 4.2 4.2 NUM fcis-1100 95 2 . . PUNCT fcis-1100 96 1 experimental experimental ADJ fcis-1100 96 2 results result NOUN fcis-1100 96 3 based base VERB fcis-1100 96 4 on on ADP fcis-1100 96 5 the the DET fcis-1100 96 6 pytorch pytorch NOUN fcis-1100 96 7 framework framework NOUN fcis-1100 96 8 of of ADP fcis-1100 96 9 deep deep ADJ fcis-1100 96 10 learning learning NOUN fcis-1100 96 11 , , PUNCT fcis-1100 96 12 the the DET fcis-1100 96 13 experiment experiment NOUN fcis-1100 96 14 is be AUX fcis-1100 96 15 carried carry VERB fcis-1100 96 16 out out ADP fcis-1100 96 17 in in ADP fcis-1100 96 18 the the DET fcis-1100 96 19 windows windows PROPN fcis-1100 96 20 environment environment NOUN fcis-1100 96 21 with with ADP fcis-1100 96 22 rtx2060 rtx2060 NOUN fcis-1100 96 23 gpu gpu NOUN fcis-1100 96 24 and and CCONJ fcis-1100 96 25 6 6 NUM fcis-1100 96 26 gb gb NOUN fcis-1100 96 27 ram ram NOUN fcis-1100 96 28 . . PUNCT fcis-1100 97 1 the the DET fcis-1100 97 2 batch batch NOUN fcis-1100 97 3 size size NOUN fcis-1100 97 4 = = SYM fcis-1100 97 5 64 64 NUM fcis-1100 97 6 , , PUNCT fcis-1100 97 7 the the DET fcis-1100 97 8 learning learning NOUN fcis-1100 97 9 rate rate NOUN fcis-1100 97 10 = = NOUN fcis-1100 97 11 1 1 NUM fcis-1100 97 12 / / SYM fcis-1100 97 13 e e NOUN fcis-1100 97 14 , , PUNCT fcis-1100 97 15 and and CCONJ fcis-1100 97 16 the the DET fcis-1100 97 17 optimizer optimizer NOUN fcis-1100 97 18 adopts adopt VERB fcis-1100 97 19 sgd sgd PROPN fcis-1100 97 20 . . PROPN fcis-1100 97 21 accuracy accuracy NOUN fcis-1100 97 22 : : PUNCT fcis-1100 98 1 𝐴𝐢𝐢 𝐴𝐢𝐢 PROPN fcis-1100 98 2 = = SYM fcis-1100 98 3 𝑇𝑃+𝑇𝑁 𝑇𝑃+𝑇𝑁 NOUN fcis-1100 98 4 𝑇𝑃+𝑇𝑁+𝐹𝑁+𝐹𝑃 𝑇𝑃+𝑇𝑁+𝐹𝑁+𝐹𝑃 NUM fcis-1100 98 5 ( ( PUNCT fcis-1100 98 6 7 7 NUM fcis-1100 98 7 ) ) PUNCT fcis-1100 98 8 the the DET fcis-1100 98 9 meanings meaning NOUN fcis-1100 98 10 represented represent VERB fcis-1100 98 11 by by ADP fcis-1100 98 12 the the DET fcis-1100 98 13 symbols symbol NOUN fcis-1100 98 14 of of ADP fcis-1100 98 15 formulas formula NOUN fcis-1100 98 16 ( ( PUNCT fcis-1100 98 17 7 7 NUM fcis-1100 98 18 ) ) PUNCT fcis-1100 98 19 are be AUX fcis-1100 98 20 as as SCONJ fcis-1100 98 21 follows follow VERB fcis-1100 98 22 : : PUNCT fcis-1100 98 23 tp tp X fcis-1100 98 24 : : PUNCT fcis-1100 98 25 true true ADJ fcis-1100 98 26 - - PUNCT fcis-1100 98 27 negative negative ADJ fcis-1100 98 28 fn fn NOUN fcis-1100 98 29 : : PUNCT fcis-1100 98 30 true true ADJ fcis-1100 98 31 - - PUNCT fcis-1100 98 32 negative negative ADJ fcis-1100 98 33 fp fp ADJ fcis-1100 98 34 : : PUNCT fcis-1100 98 35 false false ADJ fcis-1100 98 36 - - PUNCT fcis-1100 98 37 positive positive ADJ fcis-1100 98 38 tn tn NOUN fcis-1100 98 39 : : PUNCT fcis-1100 98 40 true true ADJ fcis-1100 98 41 - - PUNCT fcis-1100 98 42 negative negative ADJ fcis-1100 98 43 table table NOUN fcis-1100 98 44 1 1 NUM fcis-1100 98 45 . . PUNCT fcis-1100 98 46 accuracy accuracy NOUN fcis-1100 98 47 rate rate NOUN fcis-1100 98 48 of of ADP fcis-1100 98 49 various various ADJ fcis-1100 98 50 algorithms algorithm NOUN fcis-1100 98 51 in in ADP fcis-1100 98 52 lfw lfw NOUN fcis-1100 98 53 face face NOUN fcis-1100 98 54 recognition recognition NOUN fcis-1100 98 55 algorithm algorithm NOUN fcis-1100 98 56 accuracy accuracy NOUN fcis-1100 98 57 rate rate NOUN fcis-1100 98 58 ( ( PUNCT fcis-1100 98 59 % % INTJ fcis-1100 98 60 ) ) PUNCT fcis-1100 98 61 deepface deepface NOUN fcis-1100 99 1 [ [ X fcis-1100 99 2 21 21 NUM fcis-1100 99 3 ] ] PUNCT fcis-1100 99 4 97.35 97.35 NUM fcis-1100 99 5 % % NOUN fcis-1100 99 6 facenet facenet NOUN fcis-1100 100 1 [ [ X fcis-1100 100 2 22 22 NUM fcis-1100 100 3 ] ] PUNCT fcis-1100 100 4 98.87 98.87 NUM fcis-1100 100 5 % % NOUN fcis-1100 100 6 softmax+center softmax+center PROPN fcis-1100 100 7 loss loss NOUN fcis-1100 100 8 [ [ X fcis-1100 100 9 23 23 NUM fcis-1100 100 10 ] ] PUNCT fcis-1100 100 11 98.78 98.78 NUM fcis-1100 100 12 % % NOUN fcis-1100 100 13 arcface[16 arcface[16 X fcis-1100 100 14 ] ] PUNCT fcis-1100 100 15 99.53 99.53 NUM fcis-1100 100 16 % % NOUN fcis-1100 100 17 table table NOUN fcis-1100 100 18 2 2 NUM fcis-1100 100 19 . . PUNCT fcis-1100 100 20 accuracy accuracy NOUN fcis-1100 100 21 rate rate NOUN fcis-1100 100 22 of of ADP fcis-1100 100 23 the the DET fcis-1100 100 24 algorithm algorithm NOUN fcis-1100 100 25 ( ( PUNCT fcis-1100 100 26 s=30 s=30 NOUN fcis-1100 100 27 , , PUNCT fcis-1100 100 28 m=0.3 m=0.3 NOUN fcis-1100 100 29 , , PUNCT fcis-1100 100 30 c=0.3 c=0.3 NOUN fcis-1100 100 31 ) ) PUNCT fcis-1100 100 32 in in ADP fcis-1100 100 33 various various ADJ fcis-1100 100 34 data datum NOUN fcis-1100 100 35 sets set NOUN fcis-1100 100 36 network+ network+ ADP fcis-1100 100 37 loss loss NOUN fcis-1100 100 38 function function NOUN fcis-1100 100 39 data datum NOUN fcis-1100 100 40 set set VERB fcis-1100 100 41 accuracy accuracy NOUN fcis-1100 100 42 rate rate NOUN fcis-1100 100 43 ( ( PUNCT fcis-1100 100 44 % % INTJ fcis-1100 100 45 ) ) PUNCT fcis-1100 100 46 network network NOUN fcis-1100 100 47 structure structure NOUN fcis-1100 100 48 + + CCONJ fcis-1100 100 49 formula formula NOUN fcis-1100 100 50 ( ( PUNCT fcis-1100 100 51 4 4 X fcis-1100 100 52 ) ) PUNCT fcis-1100 100 53 lfw lfw VERB fcis-1100 100 54 99.57 99.57 NUM fcis-1100 100 55 % % NOUN fcis-1100 100 56 network network NOUN fcis-1100 100 57 structure structure NOUN fcis-1100 100 58 + + CCONJ fcis-1100 100 59 formula formula NOUN fcis-1100 100 60 ( ( PUNCT fcis-1100 100 61 5 5 NUM fcis-1100 100 62 ) ) PUNCT fcis-1100 100 63 lfw lfw VERB fcis-1100 100 64 99.60 99.60 NUM fcis-1100 100 65 % % NOUN fcis-1100 100 66 network network NOUN fcis-1100 100 67 structure structure NOUN fcis-1100 100 68 + + CCONJ fcis-1100 100 69 formula formula NOUN fcis-1100 100 70 ( ( PUNCT fcis-1100 100 71 5 5 NUM fcis-1100 100 72 ) ) PUNCT fcis-1100 100 73 agedb agedb NOUN fcis-1100 100 74 94.93 94.93 NUM fcis-1100 100 75 % % NOUN fcis-1100 100 76 network network NOUN fcis-1100 100 77 structure structure NOUN fcis-1100 100 78 + + CCONJ fcis-1100 100 79 formula formula NOUN fcis-1100 100 80 ( ( PUNCT fcis-1100 100 81 5 5 NUM fcis-1100 100 82 ) ) PUNCT fcis-1100 100 83 ar(illumination ar(illumination PROPN fcis-1100 100 84 and and CCONJ fcis-1100 100 85 expression expression NOUN fcis-1100 100 86 ) ) PUNCT fcis-1100 100 87 99.99 99.99 NUM fcis-1100 100 88 % % NOUN fcis-1100 100 89 network network NOUN fcis-1100 100 90 structure structure NOUN fcis-1100 100 91 + + CCONJ fcis-1100 100 92 formula formula NOUN fcis-1100 100 93 ( ( PUNCT fcis-1100 100 94 5 5 NUM fcis-1100 100 95 ) ) PUNCT fcis-1100 100 96 ar(occlusions ar(occlusion NOUN fcis-1100 100 97 ) ) PUNCT fcis-1100 100 98 98.8 98.8 NUM fcis-1100 100 99 % % NOUN fcis-1100 100 100 table table NOUN fcis-1100 100 101 2 2 NUM fcis-1100 100 102 shows show VERB fcis-1100 100 103 the the DET fcis-1100 100 104 comparative comparative ADJ fcis-1100 100 105 experiments experiment NOUN fcis-1100 100 106 of of ADP fcis-1100 100 107 this this DET fcis-1100 100 108 paper paper NOUN fcis-1100 100 109 and and CCONJ fcis-1100 100 110 the the DET fcis-1100 100 111 accuracy accuracy NOUN fcis-1100 100 112 of of ADP fcis-1100 100 113 the the DET fcis-1100 100 114 improved improve VERB fcis-1100 100 115 loss loss NOUN fcis-1100 100 116 function function NOUN fcis-1100 100 117 algorithm algorithm NOUN fcis-1100 100 118 on on ADP fcis-1100 100 119 four four NUM fcis-1100 100 120 data datum NOUN fcis-1100 100 121 sets set NOUN fcis-1100 100 122 . . PUNCT fcis-1100 101 1 by by ADP fcis-1100 101 2 comparing compare VERB fcis-1100 101 3 the the DET fcis-1100 101 4 accuracy accuracy NOUN fcis-1100 101 5 , , PUNCT fcis-1100 101 6 it it PRON fcis-1100 101 7 can can AUX fcis-1100 101 8 be be AUX fcis-1100 101 9 seen see VERB fcis-1100 101 10 that that SCONJ fcis-1100 101 11 the the DET fcis-1100 101 12 improved improved ADJ fcis-1100 101 13 loss loss NOUN fcis-1100 101 14 function function NOUN fcis-1100 101 15 slightly slightly ADV fcis-1100 101 16 improves improve VERB fcis-1100 101 17 the the DET fcis-1100 101 18 accuracy accuracy NOUN fcis-1100 101 19 compared compare VERB fcis-1100 101 20 to to ADP fcis-1100 101 21 the the DET fcis-1100 101 22 arcface arcface NOUN fcis-1100 101 23 loss loss NOUN fcis-1100 101 24 function function NOUN fcis-1100 101 25 set set NOUN fcis-1100 101 26 , , PUNCT fcis-1100 101 27 which which PRON fcis-1100 101 28 proves prove VERB fcis-1100 101 29 the the DET fcis-1100 101 30 effectiveness effectiveness NOUN fcis-1100 101 31 of of ADP fcis-1100 101 32 the the DET fcis-1100 101 33 improved improve VERB fcis-1100 101 34 loss loss NOUN fcis-1100 101 35 function function NOUN fcis-1100 101 36 . . PUNCT fcis-1100 102 1 5 5 X fcis-1100 102 2 . . X fcis-1100 102 3 conclusion conclusion NOUN fcis-1100 102 4 since since SCONJ fcis-1100 102 5 the the DET fcis-1100 102 6 lack lack NOUN fcis-1100 102 7 of of ADP fcis-1100 102 8 distance distance NOUN fcis-1100 102 9 constraints constraint NOUN fcis-1100 102 10 within within ADP fcis-1100 102 11 the the DET fcis-1100 102 12 same same ADJ fcis-1100 102 13 class class NOUN fcis-1100 102 14 and and CCONJ fcis-1100 102 15 between between ADP fcis-1100 102 16 different different ADJ fcis-1100 102 17 classes class NOUN fcis-1100 102 18 , , PUNCT fcis-1100 102 19 softmax softmax NOUN fcis-1100 102 20 loss loss NOUN fcis-1100 102 21 fails fail VERB fcis-1100 102 22 to to PART fcis-1100 102 23 improve improve VERB fcis-1100 102 24 accuracy accuracy NOUN fcis-1100 102 25 no no ADV fcis-1100 102 26 matter matter ADV fcis-1100 102 27 how how SCONJ fcis-1100 102 28 good good ADJ fcis-1100 102 29 the the DET fcis-1100 102 30 model model NOUN fcis-1100 102 31 is be AUX fcis-1100 102 32 used use VERB fcis-1100 102 33 in in ADP fcis-1100 102 34 training training NOUN fcis-1100 102 35 . . PUNCT fcis-1100 103 1 the the DET fcis-1100 103 2 algorithm algorithm NOUN fcis-1100 103 3 proposed propose VERB fcis-1100 103 4 in in ADP fcis-1100 103 5 this this DET fcis-1100 103 6 paper paper NOUN fcis-1100 103 7 uses use VERB fcis-1100 103 8 the the DET fcis-1100 103 9 improved improve VERB fcis-1100 103 10 arcface arcface NOUN fcis-1100 103 11 loss loss NOUN fcis-1100 103 12 function function NOUN fcis-1100 103 13 to to PART fcis-1100 103 14 solve solve VERB fcis-1100 103 15 the the DET fcis-1100 103 16 non non ADJ fcis-1100 103 17 - - ADJ fcis-1100 103 18 ideal ideal ADJ fcis-1100 103 19 classification classification NOUN fcis-1100 103 20 effect effect NOUN fcis-1100 103 21 of of ADP fcis-1100 103 22 the the DET fcis-1100 103 23 softmax softmax NOUN fcis-1100 103 24 layer layer NOUN fcis-1100 103 25 based base VERB fcis-1100 103 26 on on ADP fcis-1100 103 27 resnet resnet NOUN fcis-1100 103 28 and and CCONJ fcis-1100 103 29 has have AUX fcis-1100 103 30 achieved achieve VERB fcis-1100 103 31 great great ADJ fcis-1100 103 32 results result NOUN fcis-1100 103 33 . . PUNCT fcis-1100 104 1 a a DET fcis-1100 104 2 large large ADJ fcis-1100 104 3 number number NOUN fcis-1100 104 4 of of ADP fcis-1100 104 5 experiments experiment NOUN fcis-1100 104 6 shall shall AUX fcis-1100 104 7 be be AUX fcis-1100 104 8 conducted conduct VERB fcis-1100 104 9 to to PART fcis-1100 104 10 set set VERB fcis-1100 104 11 super super ADJ fcis-1100 104 12 - - NOUN fcis-1100 104 13 parameter parameter NOUN fcis-1100 104 14 m m VERB fcis-1100 104 15 reasonably reasonably ADV fcis-1100 104 16 . . PUNCT fcis-1100 105 1 as as SCONJ fcis-1100 105 2 the the DET fcis-1100 105 3 angular angular ADJ fcis-1100 105 4 margin margin NOUN fcis-1100 105 5 m m NOUN fcis-1100 105 6 increases increase NOUN fcis-1100 105 7 , , PUNCT fcis-1100 105 8 the the DET fcis-1100 105 9 model model NOUN fcis-1100 105 10 is be AUX fcis-1100 105 11 hard hard ADJ fcis-1100 105 12 to to PART fcis-1100 105 13 be be AUX fcis-1100 105 14 trained train VERB fcis-1100 105 15 . . PUNCT fcis-1100 106 1 further further ADJ fcis-1100 106 2 study study NOUN fcis-1100 106 3 is be AUX fcis-1100 106 4 needed need VERB fcis-1100 106 5 on on ADP fcis-1100 106 6 how how SCONJ fcis-1100 106 7 to to PART fcis-1100 106 8 select select VERB fcis-1100 106 9 m. m. NOUN fcis-1100 106 10 last last ADJ fcis-1100 106 11 but but CCONJ fcis-1100 106 12 not not PART fcis-1100 106 13 least least ADJ fcis-1100 106 14 , , PUNCT fcis-1100 106 15 improving improve VERB fcis-1100 106 16 the the DET fcis-1100 106 17 low low ADJ fcis-1100 106 18 accuracy accuracy NOUN fcis-1100 106 19 of of ADP fcis-1100 106 20 the the DET fcis-1100 106 21 agedb agedb PROPN fcis-1100 106 22 data datum NOUN fcis-1100 106 23 set set VERB fcis-1100 106 24 also also ADV fcis-1100 106 25 needs need VERB fcis-1100 106 26 to to PART fcis-1100 106 27 be be AUX fcis-1100 106 28 solved solve VERB fcis-1100 106 29 . . PUNCT fcis-1100 107 1 references reference NOUN fcis-1100 107 2 [ [ X fcis-1100 107 3 1 1 NUM fcis-1100 107 4 ] ] X fcis-1100 107 5 ojala ojala PROPN fcis-1100 107 6 t t PROPN fcis-1100 107 7 , , PUNCT fcis-1100 107 8 pietikainen pietikainen PROPN fcis-1100 107 9 m m PROPN fcis-1100 107 10 , , PUNCT fcis-1100 107 11 maenpaa maenpaa ADJ fcis-1100 107 12 t. t. NOUN fcis-1100 107 13 multiresolution multiresolution NOUN fcis-1100 107 14 gray gray ADJ fcis-1100 107 15 - - PUNCT fcis-1100 107 16 scale scale NOUN fcis-1100 107 17 and and CCONJ fcis-1100 107 18 rotation rotation NOUN fcis-1100 107 19 invariant invariant ADJ fcis-1100 107 20 texture texture NOUN fcis-1100 107 21 classification classification NOUN fcis-1100 107 22 with with ADP fcis-1100 107 23 local local ADJ fcis-1100 107 24 binary binary ADJ fcis-1100 107 25 patterns[j patterns[j PROPN fcis-1100 107 26 ] ] PUNCT fcis-1100 107 27 . . PUNCT fcis-1100 108 1 ieee ieee NOUN fcis-1100 108 2 transactions transaction NOUN fcis-1100 108 3 on on ADP fcis-1100 108 4 pattern pattern NOUN fcis-1100 108 5 analysis analysis NOUN fcis-1100 108 6 and and CCONJ fcis-1100 108 7 machine machine NOUN fcis-1100 108 8 intelligence intelligence NOUN fcis-1100 108 9 , , PUNCT fcis-1100 108 10 2002 2002 NUM fcis-1100 108 11 , , PUNCT fcis-1100 108 12 24(7 24(7 NUM fcis-1100 108 13 ): ): PUNCT fcis-1100 108 14 971 971 NUM fcis-1100 108 15 - - SYM fcis-1100 108 16 987 987 NUM fcis-1100 108 17 . . PUNCT fcis-1100 109 1 [ [ X fcis-1100 109 2 2 2 X fcis-1100 109 3 ] ] X fcis-1100 109 4 platt platt PROPN fcis-1100 109 5 j. j. PROPN fcis-1100 109 6 sequential sequential PROPN fcis-1100 109 7 minimal minimal ADJ fcis-1100 109 8 optimization optimization NOUN fcis-1100 109 9 : : PUNCT fcis-1100 109 10 a a DET fcis-1100 109 11 fast fast ADJ fcis-1100 109 12 algorithm algorithm NOUN fcis-1100 109 13 for for ADP fcis-1100 109 14 training training NOUN fcis-1100 109 15 support support NOUN fcis-1100 109 16 vector vector NOUN fcis-1100 109 17 machines[j machines[j PROPN fcis-1100 109 18 ] ] PUNCT fcis-1100 109 19 . . PUNCT fcis-1100 110 1 1998 1998 NUM fcis-1100 110 2 . . PUNCT fcis-1100 111 1 [ [ X fcis-1100 111 2 3 3 X fcis-1100 111 3 ] ] X fcis-1100 111 4 alzu’bi alzu’bi NOUN fcis-1100 111 5 a a X fcis-1100 111 6 , , PUNCT fcis-1100 111 7 albalas albalas PROPN fcis-1100 111 8 f f PROPN fcis-1100 111 9 , , PUNCT fcis-1100 111 10 al al PROPN fcis-1100 111 11 - - PUNCT fcis-1100 111 12 hadhrami hadhrami PROPN fcis-1100 111 13 t t PROPN fcis-1100 111 14 , , PUNCT fcis-1100 111 15 et et PROPN fcis-1100 111 16 al al PROPN fcis-1100 111 17 . . PROPN fcis-1100 112 1 masked mask VERB fcis-1100 112 2 face face NOUN fcis-1100 112 3 recognition recognition NOUN fcis-1100 112 4 using use VERB fcis-1100 112 5 deep deep ADJ fcis-1100 112 6 learning learning NOUN fcis-1100 112 7 : : PUNCT fcis-1100 112 8 a a DET fcis-1100 112 9 review[j review[j PROPN fcis-1100 112 10 ] ] PUNCT fcis-1100 112 11 . . PUNCT fcis-1100 113 1 electronics electronic NOUN fcis-1100 113 2 , , PUNCT fcis-1100 113 3 2021 2021 NUM fcis-1100 113 4 , , PUNCT fcis-1100 113 5 10(21 10(21 NUM fcis-1100 113 6 ): ): PUNCT fcis-1100 113 7 2666 2666 NUM fcis-1100 113 8 . . PUNCT fcis-1100 114 1 [ [ X fcis-1100 114 2 4 4 NUM fcis-1100 114 3 ] ] PUNCT fcis-1100 114 4 yu yu PROPN fcis-1100 114 5 cuicancui cuicancui PROPN fcis-1100 114 6 & & CCONJ fcis-1100 114 7 li li PROPN fcis-1100 114 8 huibin huibin PROPN fcis-1100 114 9 . . PUNCT fcis-1100 115 1 review review NOUN fcis-1100 115 2 of of ADP fcis-1100 115 3 face face NOUN fcis-1100 115 4 recognition recognition NOUN fcis-1100 115 5 methods method NOUN fcis-1100 115 6 based base VERB fcis-1100 115 7 on on ADP fcis-1100 115 8 deep deep ADJ fcis-1100 115 9 learning learning NOUN fcis-1100 115 10 [ [ X fcis-1100 115 11 j j X fcis-1100 115 12 ] ] X fcis-1100 115 13 . . PUNCT fcis-1100 116 1 chinese chinese ADJ fcis-1100 116 2 journal journal PROPN fcis-1100 116 3 of of ADP fcis-1100 116 4 engineering engineering NOUN fcis-1100 116 5 mathematics mathematic NOUN fcis-1100 116 6 , , PUNCT fcis-1100 116 7 2021,38(4 2021,38(4 NUM fcis-1100 116 8 ): ): PUNCT fcis-1100 116 9 451 451 NUM fcis-1100 116 10 - - SYM fcis-1100 116 11 469 469 NUM fcis-1100 116 12 . . PUNCT fcis-1100 117 1 [ [ X fcis-1100 117 2 5 5 X fcis-1100 117 3 ] ] PUNCT fcis-1100 117 4 tiantian tiantian PROPN fcis-1100 117 5 chen chen PROPN fcis-1100 117 6 , , PUNCT fcis-1100 117 7 hongrong hongrong PROPN fcis-1100 117 8 jing jing PROPN fcis-1100 117 9 , , PUNCT fcis-1100 117 10 hongjie hongjie PROPN fcis-1100 117 11 zhang zhang PROPN fcis-1100 117 12 . . PUNCT fcis-1100 117 13 research research NOUN fcis-1100 117 14 on on ADP fcis-1100 117 15 face face NOUN fcis-1100 117 16 recognition recognition NOUN fcis-1100 117 17 method method NOUN fcis-1100 117 18 based base VERB fcis-1100 117 19 on on ADP fcis-1100 117 20 deep deep ADJ fcis-1100 117 21 learning learning NOUN fcis-1100 117 22 [ [ X fcis-1100 117 23 j j X fcis-1100 117 24 ] ] X fcis-1100 117 25 . . PUNCT fcis-1100 118 1 scientific scientific ADJ fcis-1100 118 2 journal journal NOUN fcis-1100 118 3 of of ADP fcis-1100 118 4 intelligent intelligent ADJ fcis-1100 118 5 systems system NOUN fcis-1100 118 6 research,2021,3(7 research,2021,3(7 NOUN fcis-1100 118 7 ): ): PUNCT fcis-1100 118 8 [ [ X fcis-1100 118 9 6 6 NUM fcis-1100 118 10 ] ] SYM fcis-1100 118 11 li li PROPN fcis-1100 118 12 l l PROPN fcis-1100 118 13 , , PUNCT fcis-1100 118 14 mu mu PROPN fcis-1100 118 15 x x PROPN fcis-1100 118 16 , , PUNCT fcis-1100 118 17 li li PROPN fcis-1100 118 18 s s PROPN fcis-1100 118 19 , , PUNCT fcis-1100 118 20 et et PROPN fcis-1100 118 21 al al PROPN fcis-1100 118 22 . . PUNCT fcis-1100 119 1 a a DET fcis-1100 119 2 review review NOUN fcis-1100 119 3 of of ADP fcis-1100 119 4 face face NOUN fcis-1100 119 5 recognition recognition NOUN fcis-1100 119 6 technology technology NOUN fcis-1100 120 1 [ [ X fcis-1100 120 2 j j X fcis-1100 120 3 ] ] X fcis-1100 120 4 . . PUNCT fcis-1100 121 1 ieee ieee NOUN fcis-1100 121 2 access access NOUN fcis-1100 121 3 , , PUNCT fcis-1100 121 4 2020 2020 NUM fcis-1100 121 5 , , PUNCT fcis-1100 121 6 8 8 NUM fcis-1100 121 7 : : SYM fcis-1100 121 8 139110 139110 NUM fcis-1100 121 9 - - SYM fcis-1100 121 10 139120 139120 NUM fcis-1100 121 11 . . PUNCT fcis-1100 122 1 [ [ X fcis-1100 122 2 7 7 X fcis-1100 122 3 ] ] X fcis-1100 122 4 wang wang PROPN fcis-1100 122 5 m m PROPN fcis-1100 122 6 , , PUNCT fcis-1100 122 7 deng deng PROPN fcis-1100 122 8 w. w. PROPN fcis-1100 122 9 deep deep PROPN fcis-1100 122 10 face face NOUN fcis-1100 122 11 recognition recognition NOUN fcis-1100 122 12 : : PUNCT fcis-1100 122 13 a a DET fcis-1100 122 14 survey[j survey[j PROPN fcis-1100 122 15 ] ] X fcis-1100 122 16 . . PUNCT fcis-1100 123 1 neurocomputing neurocomputing NOUN fcis-1100 123 2 , , PUNCT fcis-1100 123 3 2021 2021 NUM fcis-1100 123 4 , , PUNCT fcis-1100 123 5 429 429 NUM fcis-1100 123 6 : : SYM fcis-1100 123 7 215 215 NUM fcis-1100 123 8 - - SYM fcis-1100 123 9 244 244 NUM fcis-1100 123 10 . . PUNCT fcis-1100 124 1 [ [ X fcis-1100 124 2 8 8 NUM fcis-1100 124 3 ] ] X fcis-1100 124 4 han han PROPN fcis-1100 124 5 x. x. PROPN fcis-1100 124 6 , , PUNCT fcis-1100 124 7 zhang zhang PROPN fcis-1100 124 8 h.y h.y PROPN fcis-1100 124 9 . . PROPN fcis-1100 124 10 , , PUNCT fcis-1100 124 11 zhang zhang PROPN fcis-1100 124 12 y.y y.y PROPN fcis-1100 124 13 . . PROPN fcis-1100 124 14 facial facial ADJ fcis-1100 124 15 expression expression NOUN fcis-1100 124 16 recognition recognition NOUN fcis-1100 124 17 based base VERB fcis-1100 124 18 on on ADP fcis-1100 124 19 efficient efficient ADJ fcis-1100 124 20 channel channel NOUN fcis-1100 124 21 attention attention NOUN fcis-1100 124 22 network[j network[j NOUN fcis-1100 124 23 ] ] PUNCT fcis-1100 124 24 . . PUNCT fcis-1100 125 1 transducer transducer NOUN fcis-1100 125 2 and and CCONJ fcis-1100 125 3 microsystem microsystem NOUN fcis-1100 125 4 technologies technology NOUN fcis-1100 125 5 , , PUNCT fcis-1100 125 6 2021(1 2021(1 NUM fcis-1100 125 7 ) ) PUNCT fcis-1100 125 8 . . PUNCT fcis-1100 126 1 [ [ X fcis-1100 126 2 9 9 NUM fcis-1100 126 3 ] ] PUNCT fcis-1100 126 4 k. k. NOUN fcis-1100 126 5 he he PROPN fcis-1100 126 6 , , PUNCT fcis-1100 126 7 x. x. PROPN fcis-1100 126 8 zhang zhang PROPN fcis-1100 126 9 , , PUNCT fcis-1100 126 10 s. s. PROPN fcis-1100 126 11 ren ren PROPN fcis-1100 126 12 , , PUNCT fcis-1100 126 13 et et PROPN fcis-1100 126 14 al al PROPN fcis-1100 126 15 . . PUNCT fcis-1100 127 1 deep deep ADJ fcis-1100 127 2 residual residual ADJ fcis-1100 127 3 learning learning NOUN fcis-1100 127 4 for for ADP fcis-1100 127 5 image image NOUN fcis-1100 127 6 recognition recognition NOUN fcis-1100 128 1 [ [ X fcis-1100 128 2 c c X fcis-1100 128 3 ] ] PUNCT fcis-1100 128 4 . . PUNCT fcis-1100 129 1 proceedings proceeding NOUN fcis-1100 129 2 of of ADP fcis-1100 129 3 the the DET fcis-1100 129 4 ieee ieee NOUN fcis-1100 129 5 conference conference NOUN fcis-1100 129 6 on on ADP fcis-1100 129 7 computer computer NOUN fcis-1100 129 8 vision vision NOUN fcis-1100 129 9 and and CCONJ fcis-1100 129 10 pattern pattern NOUN fcis-1100 129 11 recognition recognition NOUN fcis-1100 129 12 , , PUNCT fcis-1100 129 13 las las PROPN fcis-1100 129 14 vegas vegas PROPN fcis-1100 129 15 , , PUNCT fcis-1100 129 16 usa usa PROPN fcis-1100 129 17 , , PUNCT fcis-1100 129 18 2016 2016 NUM fcis-1100 129 19 , , PUNCT fcis-1100 129 20 770 770 NUM fcis-1100 129 21 - - SYM fcis-1100 129 22 778 778 NUM fcis-1100 129 23 . . PUNCT fcis-1100 130 1 [ [ X fcis-1100 130 2 10 10 NUM fcis-1100 130 3 ] ] X fcis-1100 130 4 ioffe ioffe PROPN fcis-1100 130 5 s s PART fcis-1100 130 6 , , PUNCT fcis-1100 130 7 szegedy szegedy VERB fcis-1100 130 8 c. c. NOUN fcis-1100 130 9 batch batch NOUN fcis-1100 130 10 normalization normalization NOUN fcis-1100 130 11 : : PUNCT fcis-1100 130 12 accelerating accelerate VERB fcis-1100 130 13 deep deep ADJ fcis-1100 130 14 network network NOUN fcis-1100 130 15 training training NOUN fcis-1100 130 16 by by ADP fcis-1100 130 17 reducing reduce VERB fcis-1100 130 18 internal internal ADJ fcis-1100 130 19 covariate covariate ADJ fcis-1100 130 20 shift shift NOUN fcis-1100 131 1 [ [ X fcis-1100 131 2 c c X fcis-1100 131 3 ] ] PUNCT fcis-1100 131 4 . . PUNCT fcis-1100 132 1 international international ADJ fcis-1100 132 2 conference conference NOUN fcis-1100 132 3 on on ADP fcis-1100 132 4 machine machine NOUN fcis-1100 132 5 learning learning NOUN fcis-1100 132 6 . . PUNCT fcis-1100 133 1 pmlr pmlr NOUN fcis-1100 133 2 , , PUNCT fcis-1100 133 3 2015 2015 NUM fcis-1100 133 4 : : PUNCT fcis-1100 133 5 448 448 NUM fcis-1100 133 6 - - SYM fcis-1100 133 7 456 456 NUM fcis-1100 133 8 . . PUNCT fcis-1100 134 1 [ [ X fcis-1100 134 2 11 11 NUM fcis-1100 134 3 ] ] SYM fcis-1100 134 4 hu hu PROPN fcis-1100 134 5 j j PROPN fcis-1100 134 6 , , PUNCT fcis-1100 134 7 shen shen PROPN fcis-1100 134 8 l l PROPN fcis-1100 134 9 , , PUNCT fcis-1100 134 10 sun sun PROPN fcis-1100 134 11 g. g. PROPN fcis-1100 134 12 squeeze squeeze PROPN fcis-1100 134 13 - - PUNCT fcis-1100 134 14 and and CCONJ fcis-1100 134 15 - - PUNCT fcis-1100 134 16 excitation excitation NOUN fcis-1100 134 17 networks network NOUN fcis-1100 134 18 [ [ X fcis-1100 134 19 c c X fcis-1100 134 20 ] ] PUNCT fcis-1100 134 21 . . PUNCT fcis-1100 135 1 proceedings proceeding NOUN fcis-1100 135 2 of of ADP fcis-1100 135 3 the the DET fcis-1100 135 4 ieee ieee NOUN fcis-1100 135 5 conference conference NOUN fcis-1100 135 6 on on ADP fcis-1100 135 7 computer computer NOUN fcis-1100 135 8 vision vision NOUN fcis-1100 135 9 and and CCONJ fcis-1100 135 10 pattern pattern NOUN fcis-1100 135 11 recognition recognition NOUN fcis-1100 135 12 . . PUNCT fcis-1100 136 1 2018 2018 NUM fcis-1100 136 2 : : PUNCT fcis-1100 136 3 7132 7132 NUM fcis-1100 136 4 - - SYM fcis-1100 136 5 7141 7141 NUM fcis-1100 136 6 . . PUNCT fcis-1100 137 1 [ [ X fcis-1100 137 2 12 12 NUM fcis-1100 137 3 ] ] X fcis-1100 137 4 woo woo NOUN fcis-1100 137 5 s s PROPN fcis-1100 137 6 , , PUNCT fcis-1100 137 7 park park PROPN fcis-1100 137 8 j j PROPN fcis-1100 137 9 , , PUNCT fcis-1100 137 10 lee lee PROPN fcis-1100 137 11 j j PROPN fcis-1100 137 12 y y PROPN fcis-1100 137 13 , , PUNCT fcis-1100 137 14 et et PROPN fcis-1100 137 15 al al PROPN fcis-1100 137 16 . . PUNCT fcis-1100 137 17 cbam cbam NOUN fcis-1100 137 18 : : PUNCT fcis-1100 137 19 convolutional convolutional ADJ fcis-1100 137 20 block block NOUN fcis-1100 137 21 attention attention NOUN fcis-1100 137 22 module module NOUN fcis-1100 137 23 [ [ X fcis-1100 137 24 c c X fcis-1100 137 25 ] ] PUNCT fcis-1100 137 26 . . PUNCT fcis-1100 138 1 proceedings proceeding NOUN fcis-1100 138 2 of of ADP fcis-1100 138 3 the the DET fcis-1100 138 4 european european PROPN fcis-1100 138 5 conference conference PROPN fcis-1100 138 6 on on ADP fcis-1100 138 7 computer computer NOUN fcis-1100 138 8 vision vision NOUN fcis-1100 138 9 ( ( PUNCT fcis-1100 138 10 eccv eccv ADV fcis-1100 138 11 ) ) PUNCT fcis-1100 138 12 . . PUNCT fcis-1100 139 1 2018 2018 NUM fcis-1100 139 2 : : PUNCT fcis-1100 139 3 3 3 NUM fcis-1100 139 4 - - SYM fcis-1100 139 5 19 19 NUM fcis-1100 139 6 . . PUNCT fcis-1100 140 1 [ [ X fcis-1100 140 2 13 13 NUM fcis-1100 140 3 ] ] SYM fcis-1100 140 4 yin yin PROPN fcis-1100 140 5 x x X fcis-1100 140 6 , , PUNCT fcis-1100 140 7 goudriaan goudriaan PROPN fcis-1100 140 8 j j PROPN fcis-1100 140 9 a a PRON fcis-1100 140 10 n n CCONJ fcis-1100 140 11 , , PUNCT fcis-1100 140 12 lantinga lantinga ADV fcis-1100 140 13 e e PROPN fcis-1100 140 14 a a X fcis-1100 140 15 , , PUNCT fcis-1100 140 16 et et PROPN fcis-1100 140 17 al al PROPN fcis-1100 140 18 . . PUNCT fcis-1100 141 1 a a DET fcis-1100 141 2 flexible flexible ADJ fcis-1100 141 3 sigmoid sigmoid NOUN fcis-1100 141 4 function function NOUN fcis-1100 141 5 of of ADP fcis-1100 141 6 determinate determinate ADJ fcis-1100 141 7 growth growth NOUN fcis-1100 142 1 [ [ X fcis-1100 142 2 j j X fcis-1100 142 3 ] ] X fcis-1100 142 4 . . PUNCT fcis-1100 143 1 annals annal NOUN fcis-1100 143 2 of of ADP fcis-1100 143 3 botany botany NOUN fcis-1100 143 4 , , PUNCT fcis-1100 143 5 2003 2003 NUM fcis-1100 143 6 , , PUNCT fcis-1100 143 7 91(3 91(3 NUM fcis-1100 143 8 ): ): PUNCT fcis-1100 143 9 361 361 NUM fcis-1100 143 10 - - SYM fcis-1100 143 11 371 371 NUM fcis-1100 143 12 . . PUNCT fcis-1100 144 1 25 25 NUM fcis-1100 145 1 [ [ X fcis-1100 145 2 14 14 NUM fcis-1100 145 3 ] ] X fcis-1100 145 4 jang jang PROPN fcis-1100 145 5 e e PROPN fcis-1100 145 6 , , PUNCT fcis-1100 145 7 gu gu NOUN fcis-1100 145 8 s s PROPN fcis-1100 145 9 , , PUNCT fcis-1100 145 10 poole poole PROPN fcis-1100 145 11 b. b. PROPN fcis-1100 145 12 categorical categorical ADJ fcis-1100 145 13 reparameterization reparameterization NOUN fcis-1100 145 14 with with ADP fcis-1100 145 15 gumbel gumbel NOUN fcis-1100 145 16 - - PUNCT fcis-1100 145 17 softmax softmax NOUN fcis-1100 146 1 [ [ X fcis-1100 146 2 j j X fcis-1100 146 3 ] ] X fcis-1100 146 4 . . PUNCT fcis-1100 147 1 arxiv arxiv PROPN fcis-1100 147 2 preprint preprint VERB fcis-1100 147 3 arxiv:1611.01144 arxiv:1611.01144 ADV fcis-1100 147 4 , , PUNCT fcis-1100 147 5 2016 2016 NUM fcis-1100 147 6 . . PUNCT fcis-1100 148 1 [ [ X fcis-1100 148 2 15 15 NUM fcis-1100 148 3 ] ] X fcis-1100 148 4 wen wen PROPN fcis-1100 148 5 y y PROPN fcis-1100 148 6 , , PUNCT fcis-1100 148 7 zhang zhang PROPN fcis-1100 148 8 k k PROPN fcis-1100 148 9 , , PUNCT fcis-1100 148 10 li li PROPN fcis-1100 148 11 z z PROPN fcis-1100 148 12 , , PUNCT fcis-1100 148 13 et et PROPN fcis-1100 148 14 al al PROPN fcis-1100 148 15 . . PUNCT fcis-1100 149 1 a a DET fcis-1100 149 2 discriminative discriminative NOUN fcis-1100 149 3 feature feature NOUN fcis-1100 149 4 learning learn VERB fcis-1100 149 5 approach approach NOUN fcis-1100 149 6 for for ADP fcis-1100 149 7 deep deep ADJ fcis-1100 149 8 face face NOUN fcis-1100 149 9 recognition[c recognition[c PROPN fcis-1100 149 10 ] ] PUNCT fcis-1100 149 11 . . PUNCT fcis-1100 150 1 european european ADJ fcis-1100 150 2 conference conference PROPN fcis-1100 150 3 on on ADP fcis-1100 150 4 computer computer NOUN fcis-1100 150 5 vision vision NOUN fcis-1100 150 6 . . PUNCT fcis-1100 151 1 springer springer NOUN fcis-1100 151 2 , , PUNCT fcis-1100 151 3 cham cham PROPN fcis-1100 151 4 , , PUNCT fcis-1100 151 5 2016 2016 NUM fcis-1100 151 6 : : PUNCT fcis-1100 151 7 499 499 NUM fcis-1100 151 8 - - SYM fcis-1100 151 9 515 515 NUM fcis-1100 151 10 . . PUNCT fcis-1100 152 1 [ [ X fcis-1100 152 2 16 16 NUM fcis-1100 152 3 ] ] X fcis-1100 152 4 deng deng PROPN fcis-1100 152 5 j j PROPN fcis-1100 152 6 , , PUNCT fcis-1100 152 7 guo guo PROPN fcis-1100 152 8 j j PROPN fcis-1100 152 9 , , PUNCT fcis-1100 152 10 xue xue PROPN fcis-1100 152 11 n n CCONJ fcis-1100 152 12 , , PUNCT fcis-1100 152 13 et et PROPN fcis-1100 152 14 al al PROPN fcis-1100 152 15 . . PROPN fcis-1100 152 16 arcface arcface PROPN fcis-1100 152 17 : : PUNCT fcis-1100 152 18 additive additive ADJ fcis-1100 152 19 angular angular ADJ fcis-1100 152 20 margin margin NOUN fcis-1100 152 21 loss loss NOUN fcis-1100 152 22 for for ADP fcis-1100 152 23 deep deep ADJ fcis-1100 152 24 face face NOUN fcis-1100 152 25 recognition[c recognition[c PROPN fcis-1100 152 26 ] ] PUNCT fcis-1100 152 27 . . PUNCT fcis-1100 153 1 proceedings proceeding NOUN fcis-1100 153 2 of of ADP fcis-1100 153 3 the the DET fcis-1100 153 4 ieee ieee NOUN fcis-1100 153 5 / / SYM fcis-1100 153 6 cvf cvf NOUN fcis-1100 153 7 conference conference NOUN fcis-1100 153 8 on on ADP fcis-1100 153 9 computer computer NOUN fcis-1100 153 10 vision vision NOUN fcis-1100 153 11 and and CCONJ fcis-1100 153 12 pattern pattern NOUN fcis-1100 153 13 recognition recognition NOUN fcis-1100 153 14 . . PUNCT fcis-1100 154 1 2019 2019 NUM fcis-1100 154 2 : : PUNCT fcis-1100 154 3 4690 4690 NUM fcis-1100 154 4 - - SYM fcis-1100 154 5 4699 4699 NUM fcis-1100 154 6 . . PUNCT fcis-1100 155 1 [ [ X fcis-1100 155 2 17 17 NUM fcis-1100 155 3 ] ] X fcis-1100 155 4 liu liu PROPN fcis-1100 155 5 w w PROPN fcis-1100 155 6 , , PUNCT fcis-1100 155 7 wen wen PROPN fcis-1100 155 8 y y PROPN fcis-1100 155 9 , , PUNCT fcis-1100 155 10 yu yu PROPN fcis-1100 155 11 z z PROPN fcis-1100 155 12 , , PUNCT fcis-1100 155 13 et et PROPN fcis-1100 155 14 al al PROPN fcis-1100 155 15 . . PROPN fcis-1100 155 16 sphereface sphereface PROPN fcis-1100 155 17 : : PUNCT fcis-1100 155 18 deep deep ADJ fcis-1100 155 19 hypersphere hypersphere NOUN fcis-1100 155 20 embedding embed VERB fcis-1100 155 21 for for ADP fcis-1100 155 22 face face NOUN fcis-1100 155 23 recognition[c recognition[c PROPN fcis-1100 155 24 ] ] PUNCT fcis-1100 155 25 . . PUNCT fcis-1100 156 1 proceedings proceeding NOUN fcis-1100 156 2 of of ADP fcis-1100 156 3 the the DET fcis-1100 156 4 ieee ieee NOUN fcis-1100 156 5 conference conference NOUN fcis-1100 156 6 on on ADP fcis-1100 156 7 computer computer NOUN fcis-1100 156 8 vision vision NOUN fcis-1100 156 9 and and CCONJ fcis-1100 156 10 pattern pattern NOUN fcis-1100 156 11 recognition recognition NOUN fcis-1100 156 12 . . PUNCT fcis-1100 157 1 2017 2017 NUM fcis-1100 157 2 : : PUNCT fcis-1100 157 3 212 212 NUM fcis-1100 157 4 - - SYM fcis-1100 157 5 220 220 NUM fcis-1100 157 6 . . PUNCT fcis-1100 158 1 [ [ X fcis-1100 158 2 18 18 NUM fcis-1100 158 3 ] ] X fcis-1100 158 4 wang wang PROPN fcis-1100 158 5 h h PROPN fcis-1100 158 6 , , PUNCT fcis-1100 158 7 wang wang PROPN fcis-1100 158 8 y y PROPN fcis-1100 158 9 , , PUNCT fcis-1100 159 1 zhou zhou PROPN fcis-1100 159 2 z z PROPN fcis-1100 159 3 , , PUNCT fcis-1100 159 4 et et PROPN fcis-1100 159 5 al al PROPN fcis-1100 159 6 . . PROPN fcis-1100 159 7 cosface cosface PROPN fcis-1100 159 8 : : PUNCT fcis-1100 159 9 large large ADJ fcis-1100 159 10 margin margin NOUN fcis-1100 159 11 cosine cosine NOUN fcis-1100 159 12 loss loss NOUN fcis-1100 159 13 for for ADP fcis-1100 159 14 deep deep ADJ fcis-1100 159 15 face face NOUN fcis-1100 159 16 recognition[c recognition[c PROPN fcis-1100 159 17 ] ] PUNCT fcis-1100 159 18 . . PUNCT fcis-1100 160 1 proceedings proceeding NOUN fcis-1100 160 2 of of ADP fcis-1100 160 3 the the DET fcis-1100 160 4 ieee ieee NOUN fcis-1100 160 5 conference conference NOUN fcis-1100 160 6 on on ADP fcis-1100 160 7 computer computer NOUN fcis-1100 160 8 vision vision NOUN fcis-1100 160 9 and and CCONJ fcis-1100 160 10 pattern pattern NOUN fcis-1100 160 11 recognition recognition NOUN fcis-1100 160 12 . . PUNCT fcis-1100 161 1 2018 2018 NUM fcis-1100 161 2 : : PUNCT fcis-1100 161 3 5265 5265 NUM fcis-1100 161 4 - - SYM fcis-1100 161 5 5274 5274 NUM fcis-1100 161 6 . . PUNCT fcis-1100 162 1 [ [ X fcis-1100 162 2 19 19 NUM fcis-1100 162 3 ] ] X fcis-1100 162 4 yi yi PROPN fcis-1100 162 5 d d PROPN fcis-1100 162 6 , , PUNCT fcis-1100 162 7 lei lei PROPN fcis-1100 162 8 z z PROPN fcis-1100 162 9 , , PUNCT fcis-1100 162 10 liao liao PROPN fcis-1100 162 11 s s PROPN fcis-1100 162 12 , , PUNCT fcis-1100 162 13 et et PROPN fcis-1100 162 14 al al PROPN fcis-1100 162 15 . . PUNCT fcis-1100 163 1 learning learn VERB fcis-1100 163 2 face face NOUN fcis-1100 163 3 representation representation NOUN fcis-1100 163 4 from from ADP fcis-1100 163 5 scratch[j scratch[j PROPN fcis-1100 163 6 ] ] PUNCT fcis-1100 163 7 . . PUNCT fcis-1100 164 1 arxiv arxiv PROPN fcis-1100 164 2 preprint preprint NOUN fcis-1100 164 3 arxiv:1411.7923 arxiv:1411.7923 PROPN fcis-1100 164 4 , , PUNCT fcis-1100 164 5 2014 2014 NUM fcis-1100 164 6 . . PUNCT fcis-1100 165 1 [ [ X fcis-1100 165 2 20 20 NUM fcis-1100 165 3 ] ] X fcis-1100 165 4 li li PROPN fcis-1100 165 5 j.y.(24 j.y.(24 PROPN fcis-1100 165 6 ) ) PUNCT fcis-1100 165 7 , , PUNCT fcis-1100 165 8 li li PROPN fcis-1100 165 9 zh.h zh.h PROPN fcis-1100 165 10 . . PROPN fcis-1100 165 11 , , PUNCT fcis-1100 165 12 xie xie PROPN fcis-1100 165 13 l.c l.c PROPN fcis-1100 165 14 . . PROPN fcis-1100 165 15 , , PUNCT fcis-1100 165 16 etc etc X fcis-1100 165 17 . . X fcis-1100 165 18 research research NOUN fcis-1100 165 19 progress progress NOUN fcis-1100 165 20 of of ADP fcis-1100 165 21 cross cross ADJ fcis-1100 165 22 - - ADJ fcis-1100 165 23 age age ADJ fcis-1100 165 24 face face NOUN fcis-1100 165 25 recognition recognition NOUN fcis-1100 165 26 based base VERB fcis-1100 165 27 on on ADP fcis-1100 165 28 aging age VERB fcis-1100 165 29 model[j model[j PROPN fcis-1100 165 30 ] ] PUNCT fcis-1100 165 31 . . PUNCT fcis-1100 166 1 computer computer NOUN fcis-1100 166 2 engineering engineering NOUN fcis-1100 166 3 and and CCONJ fcis-1100 166 4 applications application NOUN fcis-1100 166 5 , , PUNCT fcis-1100 166 6 2021(24 2021(24 NUM fcis-1100 166 7 ) ) PUNCT fcis-1100 166 8 . . PUNCT fcis-1100 167 1 [ [ X fcis-1100 167 2 21 21 NUM fcis-1100 167 3 ] ] PUNCT fcis-1100 167 4 taigman taigman PROPN fcis-1100 167 5 y y PROPN fcis-1100 167 6 , , PUNCT fcis-1100 167 7 yang yang PROPN fcis-1100 167 8 m m PROPN fcis-1100 167 9 , , PUNCT fcis-1100 167 10 ranzato ranzato PROPN fcis-1100 167 11 m m PROPN fcis-1100 167 12 , , PUNCT fcis-1100 167 13 et et PROPN fcis-1100 167 14 a1 a1 PROPN fcis-1100 167 15 . . PROPN fcis-1100 167 16 deepface deepface PROPN fcis-1100 167 17 : : PUNCT fcis-1100 167 18 closing close VERB fcis-1100 167 19 the the DET fcis-1100 167 20 gap gap NOUN fcis-1100 167 21 to to ADP fcis-1100 167 22 human human ADJ fcis-1100 167 23 - - PUNCT fcis-1100 167 24 level level NOUN fcis-1100 167 25 performance performance NOUN fcis-1100 167 26 in in ADP fcis-1100 167 27 face face NOUN fcis-1100 167 28 verification verification NOUN fcis-1100 167 29 [ [ X fcis-1100 167 30 c c X fcis-1100 167 31 ] ] PUNCT fcis-1100 167 32 . . PUNCT fcis-1100 168 1 computer computer NOUN fcis-1100 168 2 vision vision NOUN fcis-1100 168 3 and and CCONJ fcis-1100 168 4 pattern pattern NOUN fcis-1100 168 5 recognition recognition NOUN fcis-1100 168 6 . . PUNCT fcis-1100 169 1 ieee ieee PROPN fcis-1100 169 2 , , PUNCT fcis-1100 169 3 2014 2014 NUM fcis-1100 169 4 : : PUNCT fcis-1100 169 5 17011708 17011708 NUM fcis-1100 169 6 . . PUNCT fcis-1100 170 1 [ [ X fcis-1100 170 2 22 22 NUM fcis-1100 170 3 ] ] SYM fcis-1100 170 4 schroff schroff NOUN fcis-1100 170 5 f f PROPN fcis-1100 170 6 , , PUNCT fcis-1100 170 7 kalenichenko kalenichenko PROPN fcis-1100 170 8 d d PROPN fcis-1100 170 9 , , PUNCT fcis-1100 170 10 philbin philbin PROPN fcis-1100 170 11 j. j. PROPN fcis-1100 170 12 facenet facenet PROPN fcis-1100 170 13 : : PUNCT fcis-1100 170 14 a a DET fcis-1100 170 15 unifiedembedding unifiedembedding NOUN fcis-1100 170 16 for for ADP fcis-1100 170 17 face face NOUN fcis-1100 170 18 recognition recognition NOUN fcis-1100 170 19 and and CCONJ fcis-1100 170 20 clustering cluster VERB fcis-1100 170 21 [ [ X fcis-1100 170 22 c c X fcis-1100 170 23 ] ] PUNCT fcis-1100 170 24 . . PUNCT fcis-1100 171 1 ieee ieee PROPN fcis-1100 171 2 conference conference PROPN fcis-1100 171 3 on on ADP fcis-1100 171 4 computer computer NOUN fcis-1100 171 5 vision vision NOUN fcis-1100 171 6 and and CCONJ fcis-1100 171 7 pattern pattern NOUN fcis-1100 171 8 recognition recognition NOUN fcis-1100 171 9 . . PUNCT fcis-1100 172 1 ieee ieee NOUN fcis-1100 172 2 computer computer NOUN fcis-1100 172 3 society society NOUN fcis-1100 172 4 , , PUNCT fcis-1100 172 5 2015 2015 NUM fcis-1100 172 6 : : PUNCT fcis-1100 172 7 815 815 NUM fcis-1100 172 8 - - SYM fcis-1100 172 9 823 823 NUM fcis-1100 172 10 . . PUNCT fcis-1100 173 1 [ [ X fcis-1100 173 2 23 23 NUM fcis-1100 173 3 ] ] X fcis-1100 173 4 sun sun PROPN fcis-1100 173 5 yi yi PROPN fcis-1100 173 6 , , PUNCT fcis-1100 173 7 wang wang PROPN fcis-1100 173 8 xiaogang xiaogang PROPN fcis-1100 173 9 , , PUNCT fcis-1100 173 10 tang tang PROPN fcis-1100 173 11 xiaoou xiaoou PROPN fcis-1100 173 12 . . PUNCT fcis-1100 174 1 deep deep ADJ fcis-1100 174 2 learning learning NOUN fcis-1100 174 3 face face VERB fcis-1100 174 4 representation representation NOUN fcis-1100 174 5 from from ADP fcis-1100 174 6 predicting predict VERB fcis-1100 174 7 10 10 NUM fcis-1100 174 8 , , PUNCT fcis-1100 174 9 000 000 NUM fcis-1100 174 10 classes class NOUN fcis-1100 174 11 [ [ X fcis-1100 174 12 c c X fcis-1100 174 13 ] ] PUNCT fcis-1100 174 14 . . PUNCT fcis-1100 175 1 ieee ieee PROPN fcis-1100 175 2 conference conference PROPN fcis-1100 175 3 on on ADP fcis-1100 175 4 computer computer NOUN fcis-1100 175 5 vision vision NOUN fcis-1100 175 6 and and CCONJ fcis-1100 175 7 pattern pattern NOUN fcis-1100 175 8 recognition recognition NOUN fcis-1100 175 9 , , PUNCT fcis-1100 175 10 2014 2014 NUM fcis-1100 175 11 : : PUNCT fcis-1100 175 12 1891 1891 NUM fcis-1100 175 13 - - SYM fcis-1100 175 14 1898 1898 NUM fcis-1100 175 15 . . PUNCT