%PDF-1.6
%
1 0 obj
<< /Metadata 3 0 R /Names 4 0 R /OpenAction 5 0 R /PageMode /UseNone /Pages 6 0 R /Type /Catalog >>
endobj
2 0 obj
<< /Author (Xinzhe Han; Shuhui Wang; Chi Su; Qingming Huang; Qi Tian) /Producer (pikepdf 3.1.0) /Subject (IEEE International Conference on Computer Vision) /Title (Greedy Gradient Ensemble for Robust Visual Question Answering) >>
endobj
3 0 obj
<< /Subtype /XML /Type /Metadata /Length 1181 >>
stream
Greedy Gradient Ensemble for Robust Visual Question AnsweringXinzhe Han Shuhui Wang Chi Su Qingming Huang Qi TianIEEE International Conference on Computer Vision
endstream
endobj
4 0 obj
<< /Dests 7 0 R >>
endobj
5 0 obj
<< /D [ 8 0 R /Fit ] /S /GoTo >>
endobj
6 0 obj
<< /Count 10 /Kids [ 9 0 R 10 0 R ] /Type /Pages >>
endobj
7 0 obj
<< /Kids [ 11 0 R 12 0 R 13 0 R 14 0 R ] /Limits [ (Doc-Start) (table.4) ] >>
endobj
8 0 obj
<< /Annots [ 15 0 R 16 0 R 17 0 R 18 0 R 19 0 R 20 0 R 21 0 R 22 0 R 23 0 R 24 0 R 25 0 R 26 0 R 27 0 R 28 0 R 29 0 R 30 0 R 31 0 R 32 0 R 33 0 R 34 0 R 35 0 R 36 0 R 37 0 R 38 0 R 39 0 R 40 0 R 41 0 R 42 0 R ] /Contents [ 43 0 R 44 0 R 45 0 R ] /MediaBox [ 0 0 612 792 ] /Parent 9 0 R /Resources 46 0 R /Type /Page >>
endobj
9 0 obj
<< /Count 6 /Kids [ 8 0 R 47 0 R 48 0 R 49 0 R 50 0 R 51 0 R ] /Parent 6 0 R /Type /Pages >>
endobj
10 0 obj
<< /Count 4 /Kids [ 52 0 R 53 0 R 54 0 R 55 0 R ] /Parent 6 0 R /Type /Pages >>
endobj
11 0 obj
<< /Kids [ 56 0 R 57 0 R 58 0 R 59 0 R 60 0 R 61 0 R ] /Limits [ (Doc-Start) (cite.2020cf-vqa) ] >>
endobj
12 0 obj
<< /Kids [ 62 0 R 63 0 R 64 0 R 65 0 R 66 0 R 67 0 R ] /Limits [ (cite.2020counterfactual) (page.1) ] >>
endobj
13 0 obj
<< /Kids [ 68 0 R 69 0 R 70 0 R 71 0 R 72 0 R 73 0 R ] /Limits [ (page.10) (table.1) ] >>
endobj
14 0 obj
<< /Kids [ 74 0 R ] /Limits [ (table.2) (table.4) ] >>
endobj
15 0 obj
<< /A << /D (cite.2017rn) /S /GoTo >> /Border [ 0 0 0 ] /C [ 0 1 0 ] /H /I /Rect [ 79.2 217.304 91.155 226.23 ] /Subtype /Link /Type /Annot >>
endobj
16 0 obj
<< /A << /D (cite.2018film) /S /GoTo >> /Border [ 0 0 0 ] /C [ 0 1 0 ] /H /I /Rect [ 94.171 217.304 106.126 226.23 ] /Subtype /Link /Type /Annot >>
endobj
17 0 obj
<< /A << /D (cite.2019murel) /S /GoTo >> /Border [ 0 0 0 ] /C [ 0 1 0 ] /H /I /Rect [ 109.142 217.304 116.116 226.23 ] /Subtype /Link /Type /Annot >>
endobj
18 0 obj
<< /A << /D (cite.gao2019multi) /S /GoTo >> /Border [ 0 0 0 ] /C [ 0 1 0 ] /H /I /Rect [ 119.132 217.304 131.087 226.23 ] /Subtype /Link /Type /Annot >>
endobj
19 0 obj
<< /A << /D (cite.2017n2nmn) /S /GoTo >> /Border [ 0 0 0 ] /C [ 0 1 0 ] /H /I /Rect [ 134.103 217.304 146.058 226.23 ] /Subtype /Link /Type /Annot >>
endobj
20 0 obj
<< /A << /D (cite.2018stacknmn) /S /GoTo >> /Border [ 0 0 0 ] /C [ 0 1 0 ] /H /I /Rect [ 149.074 217.369 161.029 226.23 ] /Subtype /Link /Type /Annot >>
endobj
21 0 obj
<< /A << /D (cite.2019lcgn) /S /GoTo >> /Border [ 0 0 0 ] /C [ 0 1 0 ] /H /I /Rect [ 164.045 217.224 176.001 226.23 ] /Subtype /Link /Type /Annot >>
endobj
22 0 obj
<< /A << /D (cite.2019nscl) /S /GoTo >> /Border [ 0 0 0 ] /C [ 0 1 0 ] /H /I /Rect [ 179.017 217.224 190.972 226.23 ] /Subtype /Link /Type /Annot >>
endobj
23 0 obj
<< /A << /D (cite.2015vqa) /S /GoTo >> /Border [ 0 0 0 ] /C [ 0 1 0 ] /H /I /Rect [ 219.634 193.394 226.608 202.32 ] /Subtype /Link /Type /Annot >>
endobj
24 0 obj
<< /A << /D (cite.2017mfh) /S /GoTo >> /Border [ 0 0 0 ] /C [ 0 1 0 ] /H /I /Rect [ 229.597 193.394 241.552 202.32 ] /Subtype /Link /Type /Annot >>
endobj
25 0 obj
<< /A << /D (cite.2017clevr) /S /GoTo >> /Border [ 0 0 0 ] /C [ 0 1 0 ] /H /I /Rect [ 244.541 193.394 256.496 202.32 ] /Subtype /Link /Type /Annot >>
endobj
26 0 obj
<< /A << /D (cite.2019gqa) /S /GoTo >> /Border [ 0 0 0 ] /C [ 0 1 0 ] /H /I /Rect [ 259.485 193.394 271.44 202.32 ] /Subtype /Link /Type /Annot >>
endobj
27 0 obj
<< /A << /D (cite.2017mfh) /S /GoTo >> /Border [ 0 0 0 ] /C [ 0 1 0 ] /H /I /Rect [ 114.971 157.5 126.926 166.426 ] /Subtype /Link /Type /Annot >>
endobj
28 0 obj
<< /A << /D (cite.2017analysis) /S /GoTo >> /Border [ 0 0 0 ] /C [ 0 1 0 ] /H /I /Rect [ 131.115 157.64 143.07 166.426 ] /Subtype /Link /Type /Annot >>
endobj
29 0 obj
<< /A << /D (cite.2016yin) /S /GoTo >> /Border [ 0 0 0 ] /C [ 0 1 0 ] /H /I /Rect [ 147.258 157.5 159.214 166.426 ] /Subtype /Link /Type /Annot >>
endobj
30 0 obj
<< /A << /D (cite.2018overcoming) /S /GoTo >> /Border [ 0 0 0 ] /C [ 0 1 0 ] /H /I /Rect [ 456.126 163.551 468.081 172.477 ] /Subtype /Link /Type /Annot >>
endobj
31 0 obj
<< /A << /D (cite.2019rubi) /S /GoTo >> /Border [ 0 0 0 ] /C [ 0 1 0 ] /H /I /Rect [ 471.238 163.615 478.212 172.337 ] /Subtype /Link /Type /Annot >>
endobj
32 0 obj
<< /A << /D (cite.2019don) /S /GoTo >> /Border [ 0 0 0 ] /C [ 0 1 0 ] /H /I /Rect [ 481.37 163.551 493.325 172.477 ] /Subtype /Link /Type /Annot >>
endobj
33 0 obj
<< /A << /D (cite.2019hint) /S /GoTo >> /Border [ 0 0 0 ] /C [ 0 1 0 ] /H /I /Rect [ 336.665 151.516 348.62 160.522 ] /Subtype /Link /Type /Annot >>
endobj
34 0 obj
<< /A << /D (cite.2019SCR) /S /GoTo >> /Border [ 0 0 0 ] /C [ 0 1 0 ] /H /I /Rect [ 351.916 151.735 363.871 160.522 ] /Subtype /Link /Type /Annot >>
endobj
35 0 obj
<< /A << /D (cite.2020DLP) /S /GoTo >> /Border [ 0 0 0 ] /C [ 0 1 0 ] /H /I /Rect [ 367.166 151.735 379.122 160.522 ] /Subtype /Link /Type /Annot >>
endobj
36 0 obj
<< /A << /D (cite.2020counterfactual) /S /GoTo >> /Border [ 0 0 0 ] /C [ 0 1 0 ] /H /I /Rect [ 489.43 151.595 496.404 160.522 ] /Subtype /Link /Type /Annot >>
endobj
37 0 obj
<< /A << /D (cite.2020bbn) /S /GoTo >> /Border [ 0 0 0 ] /C [ 0 1 0 ] /H /I /Rect [ 390.693 127.685 402.648 136.611 ] /Subtype /Link /Type /Annot >>
endobj
38 0 obj
<< /A << /D (cite.2020decoupling) /S /GoTo >> /Border [ 0 0 0 ] /C [ 0 1 0 ] /H /I /Rect [ 405.821 127.685 417.776 136.611 ] /Subtype /Link /Type /Annot >>
endobj
39 0 obj
<< /A << /D (cite.2018annotation) /S /GoTo >> /Border [ 0 0 0 ] /C [ 0 1 0 ] /H /I /Rect [ 420.949 127.685 432.904 136.611 ] /Subtype /Link /Type /Annot >>
endobj
40 0 obj
<< /A << /D (cite.2019right) /S /GoTo >> /Border [ 0 0 0 ] /C [ 0 1 0 ] /H /I /Rect [ 436.077 127.685 448.032 136.611 ] /Subtype /Link /Type /Annot >>
endobj
41 0 obj
<< /A << /D (cite.2017hat) /S /GoTo >> /Border [ 0 0 0 ] /C [ 0 1 0 ] /H /I /Rect [ 329.763 80.004 341.718 88.791 ] /Subtype /Link /Type /Annot >>
endobj
42 0 obj
<< /A << /D (cite.2018vqax) /S /GoTo >> /Border [ 0 0 0 ] /C [ 0 1 0 ] /H /I /Rect [ 345.295 80.004 357.25 88.791 ] /Subtype /Link /Type /Annot >>
endobj
43 0 obj
<< /Length 10086 >>
stream
0 g 0 G
0 g 0 G
0 g 0 G
0 g 0 G
0 g 0 G
BT
/F50 14.3462 Tf 94.092 675.067 Td [(Gr)18(eedy)-250(Gradient)-250(Ensemble)-250(f)25(or)-250(Rob)20(ust)-250(V)37(isual)-250(Question)-250(Answering)]TJ/F51 11.9552 Tf 20.613 -37.858 Td [(Xinzhe)-250(Han)]TJ/F22 7.9701 Tf 57.432 4.339 Td [(2)]TJ/F25 7.9701 Tf 4.234 0 Td [(;)]TJ/F22 7.9701 Tf 2.353 0 Td [(1)]TJ/F51 11.9552 Tf 16.687 -4.339 Td [(Shuhui)-250(W)80(ang)]TJ/F22 7.9701 Tf 64.462 4.339 Td [(1)]TJ/F28 7.9701 Tf 4.732 0 Td [<03>]TJ/F51 11.9552 Tf 11.955 -4.339 Td [(Chi)-250(Su)]TJ/F22 7.9701 Tf 32.889 4.339 Td [(3)]TJ/F51 11.9552 Tf 16.687 -4.339 Td [(Qingming)-250(Huang)]TJ/F22 7.9701 Tf 83.351 4.339 Td [(1)]TJ/F25 7.9701 Tf 4.234 0 Td [(;)]TJ/F22 7.9701 Tf 2.353 0 Td [(2)]TJ/F25 7.9701 Tf 4.234 0 Td [(;)]TJ/F22 7.9701 Tf 2.352 0 Td [(4)]TJ/F51 11.9552 Tf 16.688 -4.339 Td [(Qi)-250(T)35(ian)]TJ/F22 7.9701 Tf 36.439 4.339 Td [(5)]TJ -367.937 -13.948 Td [(1)]TJ/F51 11.9552 Tf 4.732 -4.338 Td [(K)25(e)15(y)-250(Lab)-250(of)-250(Intell.)-310(Info.)-310(Process.,)-250(Inst.)-310(of)-250(Comput.)-310(T)70(ech.,)-250(CAS,)-250(Beijing,)-250(China)]TJ/F22 7.9701 Tf -43.943 -9.61 Td [(2)]TJ/F51 11.9552 Tf 4.732 -4.338 Td [(Uni)25(v)15(ersity)-250(of)-250(Chinese)-250(Academy)-250(of)-250(Sciences,)-250(Beijing,)-250(China)]TJ/F22 7.9701 Tf 298.697 4.338 Td [(3)]TJ/F51 11.9552 Tf 4.732 -4.338 Td [(Kingsoft)-250(Cloud,)-250(Beijing,)-250(China)]TJ/F22 7.9701 Tf -314.235 -9.609 Td [(4)]TJ/F51 11.9552 Tf 4.732 -4.339 Td [(Peng)-250(Cheng)-250(Laboratory)65(,)-250(Shenzhen,)-250(China)]TJ/F22 7.9701 Tf 212.394 4.339 Td [(5)]TJ/F51 11.9552 Tf 4.733 -4.339 Td [(Cloud)-250(B)10(U,)-250(Hua)15(wei)-250(T)70(echnologies,)-250(Shenzhen,)-250(China.)]TJ/F52 8.9664 Tf -175.106 -13.948 Td [(hanxinzhe17@mails.ucas.ac.cn,)-600(wangshuhui@ict.ac.cn,)-600(suchi@kingsoft.com)]TJ 83.387 -13.947 Td [(qmhuang@ucas.ac.cn,)-600(tian.qi1@huawei.com)]TJ
0 g 0 G
/F50 11.9552 Tf -46.71 -41.046 Td [(Abstract)]TJ/F54 9.9626 Tf -83.928 -23.939 Td [(Langua)10(g)10(e)-277(bias)-278(is)-277(a)-277(critical)-278(i)1(ssue)-278(in)-277(V)74(isual)-277(Question)-278(An-)]TJ -11.955 -11.955 Td [(swering)-408(\(VQA\),)-408(wher)37(e)-408(models)-408(often)-408(e)20(xploit)-408(dataset)-408(bias-)]TJ 0 -11.955 Td [(es)-322(for)-323(the)-322<026e616c>-322(decision)-323(w)1(ithout)-323(considering)-322(the)-322(ima)10(g)10(e)-323(in-)]TJ 0 -11.955 Td [(formation.)-536(As)-326(a)-325(r)37(esult,)-345(the)30(y)-325(suf)18(fer)-326(fr)45(om)-325(performance)-326(dr)45(op)]TJ 0 -11.955 Td [(on)-292(out-of-distrib)20(ution)-293(data)-292(and)-292(inadequate)-292(visual)-293(e)20(xplana-)]TJ 0 -11.955 Td [(tion.)-785(Based)-409(on)-408(e)20(xperimental)-408(analysis)-409(for)-408(e)20(xisting)-409(r)45(ob)20(ust)]TJ 0 -11.956 Td [(VQA)-435(methods,)-482(we)-436(str)37(ess)-435(the)-436(langua)10(g)10(e)-435(bias)-436(in)-435(VQA)-436(that)]TJ 0 -11.955 Td [(comes)-271(fr)45(om)-272(two)-271(aspects,)-277(i.e)15(.,)-277(distrib)20(ution)-271(bias)-271(and)-272(shortcut)]TJ 0 -11.955 Td [(bias.)-339(W)92(e)-259(further)-260(pr)45(opose)-259(a)-260(ne)15(w)-259(de-bias)-260(fr)15(ame)15(work,)-262(Gr)37(eedy)]TJ 0 -11.955 Td [(Gr)15(adient)-203(Ensemble)-203(\(GGE\),)-203(whic)15(h)-203(combines)-204(multiple)-203(biased)]TJ 0 -11.955 Td [(models)-290(for)-290(unbiased)-290(base)-291(model)-290(learning)15(.)-430(W)55(ith)-291(the)-290(gr)37(eedy)]TJ 0 -11.956 Td [(str)15(ate)40(gy)55(,)-386(GGE)-358(for)37(ces)-359(the)-359(biased)-358(models)-359(to)-358(o)10(ver)20<2d0274>-359(the)-359(bi-)]TJ 0 -11.955 Td [(ased)-240(data)-239(distrib)20(ution)-240(in)-240(priority)55(,)-242(thus)-239(mak)10(es)-240(the)-240(base)-240(mod-)]TJ 0 -11.955 Td [(el)-396(pay)-396(mor)37(e)-395(attention)-396(to)-396(e)20(xamples)-396(that)-395(ar)37(e)-396(har)37(d)-396(to)-396(solve)]TJ 0 -11.955 Td [(by)-370(biased)-370(models.)-670(The)-370(e)20(xperiments)-370(demonstr)15(ate)-370(that)-370(our)]TJ 0 -11.955 Td [(method)-226(mak)10(es)-227(better)-226(use)-227(of)-226(visual)-227(information)-226(and)-227(ac)15(hie)15(ves)]TJ 0 -11.955 Td [(state-of-the-art)-389(performance)-390(on)-390(dia)10(gnosing)-389(dataset)-390(VQA-)]TJ 0 -11.956 Td [(CP)-250(without)-250(using)-250(e)20(xtr)15(a)-250(annotations.)]TJ/F50 11.9552 Tf 0 -25.986 Td [(1.)-250(Intr)18(oduction)]TJ/F51 9.9626 Tf 11.955 -18.957 Td [(V)60(isual)-249(Question)-249(Answering)-249(\(VQA\))-249(is)-250(a)-249(challenging)-249(task)]TJ -11.955 -11.955 Td [(that)-257(requires)-257(both)-258(language-a)15(w)10(are)-257(reasoning)-257(and)-257(image)-258(un-)]TJ 0 -11.955 Td [(derstanding.)-680(W)40(ith)-373(adv)25(ances)-374(in)-373(deep)-373(learning,)-404(neural)-374(net-)]TJ 0 -11.956 Td [(w)10(orks)-253([)]TJ
0 1 0 rg 0 1 0 RG
[(37)]TJ
0 g 0 G
[(,)]TJ
0 1 0 rg 0 1 0 RG
[-252(34)]TJ
0 g 0 G
[(,)]TJ
0 1 0 rg 0 1 0 RG
[-253(6)]TJ
0 g 0 G
[(,)]TJ
0 1 0 rg 0 1 0 RG
[-253(13)]TJ
0 g 0 G
[(,)]TJ
0 1 0 rg 0 1 0 RG
[-253(18)]TJ
0 g 0 G
[(,)]TJ
0 1 0 rg 0 1 0 RG
[-252(17)]TJ
0 g 0 G
[(,)]TJ
0 1 0 rg 0 1 0 RG
[-253(19)]TJ
0 g 0 G
[(,)]TJ
0 1 0 rg 0 1 0 RG
[-253(29)]TJ
0 g 0 G
[(])-253(that)-252(model)-253(the)-253(correla-)]TJ 0 -11.955 Td [(tions)-286(between)-286(vision)-285(and)-286(language)-286(ha)20(v)15(e)-286(sho)25(wn)-286(remarkable)]TJ 0 -11.955 Td [(results)-250(on)-250(lar)18(ge-scale)-250(benchmark)-250(datasets)-250([)]TJ
0 1 0 rg 0 1 0 RG
[(3)]TJ
0 g 0 G
[(,)]TJ
0 1 0 rg 0 1 0 RG
[-250(15)]TJ
0 g 0 G
[(,)]TJ
0 1 0 rg 0 1 0 RG
[-250(23)]TJ
0 g 0 G
[(,)]TJ
0 1 0 rg 0 1 0 RG
[-250(20)]TJ
0 g 0 G
[(].)]TJ 11.955 -11.983 Td [(Ho)25(we)25(v)15(er)40(,)-509(recent)-458(studies)-457(ha)20(v)15(e)-458(demonstrated)-457(that)-458(most)]TJ -11.955 -11.955 Td [(VQA)-293(methods)-292(tend)-293(to)-293(rely)-293(on)-292(e)15(xisting)-293(idiosyncratic)-293(biases)]TJ 0 -11.955 Td [(in)-370(the)-371(datasets)-370([)]TJ
0 1 0 rg 0 1 0 RG
[(15)]TJ
0 g 0 G
[(,)]TJ
0 1 0 rg 0 1 0 RG
[-371(24)]TJ
0 g 0 G
[(,)]TJ
0 1 0 rg 0 1 0 RG
[-370(43)]TJ
0 g 0 G
[(].)-671(The)15(y)-371(often)-370(le)25(v)15(erage)-371(super\002-)]TJ 0 -11.956 Td [(cial)-243(correlations)-242(between)-243(questions)-242(and)-243(answers)-243(to)-242(train)-243(the)]TJ 0 -11.955 Td [(model)-381(without)-380(considering)-381(e)15(xact)-381(vision)-381(informat)1(ion.)-703(F)15(or)]TJ 0 -11.955 Td [(e)15(xample,)-221(a)-213(model)-213(may)-214(blindly)-213(answer)-214(\223tennis\224)-213(for)-213(the)-214(ques-)]TJ 0 -11.955 Td [(tion)-194(\223What)-194(sports)-194(...)70<94>-193(just)-194(based)-194(on)-194(the)-194(most)-194(common)-194(te)15(xtual)]TJ 0 -11.955 Td [(QA)-284(pairs)-284(in)-283(the)-284(train)-284(set.)-412(Unfortunately)65(,)-292(models)-284(e)15(xploiting)]TJ
0 g 0 G
ET
q
1 0 0 1 50.112 90.415 cm
[]0 d 0 J 0.398 w 0 0 m 94.499 0 l S
Q
BT
/F29 5.9776 Tf 60.141 83.813 Td [<03>]TJ/F51 7.9701 Tf 4.317 -2.813 Td [(Corresponding)-250(author)55(.)]TJ
0 g 0 G
0 g 0 G
0 g 0 G
0 g 0 G
0 g 0 G
0 g 0 G
ET
1 0 0 1 320.675 422.687 cm
q
.34743 0 0 .34743 0 0 cm
q
1 0 0 1 0 -34.589 cm
/Im1 Do
Q
Q
1 0 0 1 -320.675 -422.687 cm
BT
/F51 7.9701 Tf 393.654 410.732 Td [(\(a\))-413(Distri)1(b)19(ut)1(ion)-250(Bias)]TJ
ET
1 0 0 1 311.353 315.277 cm
q
.28905 0 0 .28905 0 0 cm
q
1 0 0 1 -17.334 -142.658 cm
/Im2 Do
Q
Q
1 0 0 1 -311.353 -315.277 cm
BT
/F51 7.9701 Tf 401.6 303.322 Td [(\(b\))-413(Shortcut)-250(Bias)]TJ/F51 8.9664 Tf -92.738 -13.57 Td [(Figure)-274(1.)-274(T)80(w)10(o)-274(aspects)-274(of)-274(language)-274(bias)-274(in)-274(VQA.)]TJ/F50 8.9664 Tf 176.687 0 Td [(\(a\))-274(Distrib)20(ution)]TJ -176.687 -10.959 Td [(Bias:)]TJ/F51 8.9664 Tf 25.308 0 Td [(The)-423(answer)-422(distrib)20(ution)-423(for)-423(certain)-423(question)-422(type)-423(is)-423(sig-)]TJ -25.308 -10.959 Td [(ni\002cantly)-395(long-tailed.)]TJ/F50 8.9664 Tf 85.205 0 Td [(\(b\))-396(Sh)1(ortcut)-396(Bias:)]TJ/F51 8.9664 Tf 76.24 0 Td [(The)-395(correct)-396(answers)]TJ -161.445 -10.959 Td [(produced)-292(by)-293(the)-292(model)-292(may)-292(rely)-293(on)-292(the)-292(question-answer)-292(shortcut)]TJ 0 -10.958 Td [(rather)-250(than)-250(proper)-250(visual)-250(grounding.)]TJ
0 g 0 G
/F51 9.9626 Tf 0 -29.562 Td [(statistical)-299(shortcuts)-299(during)-300(training)-299(often)-299(sho)25(w)-299(poor)-300(gener)20(-)]TJ 0 -11.956 Td [(alization)-313(ability)-312(to)-313(out-of-domain)-312(data,)-329(and)-312(hardly)-313(pro)15(vide)]TJ 0 -11.955 Td [(proper)-250(visual)-250(e)25(vidence)-250(for)-250(a)-250(certain)-250(answer)55(.)]TJ 11.955 -15.803 Td [(Currently)65(,)-349(the)-329(pre)25(v)25(ailing)-330(solutions)-329(for)-329(this)-329(problem)-330(can)]TJ -11.955 -11.955 Td [(be)-267(cate)15(gorized)-267(into)-267(ensemble-based)-267([)]TJ
0 1 0 rg 0 1 0 RG
[(36)]TJ
0 g 0 G
[(,)]TJ
0 1 0 rg 0 1 0 RG
[-267(7)]TJ
0 g 0 G
[(,)]TJ
0 1 0 rg 0 1 0 RG
[-267(10)]TJ
0 g 0 G
[(],)-271(grounding-)]TJ 0 -11.955 Td [(based)-281([)]TJ
0 1 0 rg 0 1 0 RG
[(39)]TJ
0 g 0 G
[(,)]TJ
0 1 0 rg 0 1 0 RG
[-281(42)]TJ
0 g 0 G
[(,)]TJ
0 1 0 rg 0 1 0 RG
[-280(22)]TJ
0 g 0 G
[(])-281(and)-281(counterf)10(actual-based)-281([)]TJ
0 1 0 rg 0 1 0 RG
[(8)]TJ
0 g 0 G
[(].)-402(Similar)-281(to)]TJ 0 -11.955 Td [(re-weighting)-270(and)-270(re-sampling)-270(strate)15(gies)-270(in)-271(traditional)-270(long-)]TJ 0 -11.955 Td [(tailed)-268(classi\002cation)-269([)]TJ
0 1 0 rg 0 1 0 RG
[(45)]TJ
0 g 0 G
[(,)]TJ
0 1 0 rg 0 1 0 RG
[-268(25)]TJ
0 g 0 G
[(,)]TJ
0 1 0 rg 0 1 0 RG
[-269(16)]TJ
0 g 0 G
[(,)]TJ
0 1 0 rg 0 1 0 RG
[-268(31)]TJ
0 g 0 G
[(],)-273(ensemble-based)-269(meth-)]TJ 0 -11.956 Td [(ods)-479(re-weight)-480(the)-479(samples)-480(by)-479(the)-479(question-only)-480(branch.)]TJ 0 -11.955 Td [(Grounding-based)-402(models)-402(stress)-403(a)-402(better)-402(use)-402(of)-402(image)-403(in-)]TJ 0 -11.955 Td [(formation)-458(according)-457(to)-458(human-annotated)-458(visual)-458(e)15(xplana-)]TJ 0 -11.955 Td [(tion)-309([)]TJ
0 1 0 rg 0 1 0 RG
[(11)]TJ
0 g 0 G
[(,)]TJ
0 1 0 rg 0 1 0 RG
[-309(21)]TJ
0 g 0 G
[(].)-487(Ne)25(wly)-309(proposed)-309(counterf)10(actual-based)-309(meth-)]TJ
0 g 0 G
0 g 0 G
ET
endstream
endobj
44 0 obj
<< /Length 46 >>
stream
q
1 0 0 1 0 0 cm
/ItcEIK82bU0uqQinTGv--A Do
Q
endstream
endobj
45 0 obj
<< /Length 46 >>
stream
q
1 0 0 1 0 0 cm
/b_Ez7txKMWl7zKNKo1XclA Do
Q
endstream
endobj
46 0 obj
<< /Font << /F22 75 0 R /F25 76 0 R /F28 77 0 R /F29 78 0 R /F50 79 0 R /F51 80 0 R /F52 81 0 R /F54 82 0 R >> /ProcSet [ /PDF /Text ] /XObject << /Im1 83 0 R /Im2 84 0 R /ItcEIK82bU0uqQinTGv--A 85 0 R /b_Ez7txKMWl7zKNKo1XclA 86 0 R >> >>
endobj
47 0 obj
<< /Annots [ 87 0 R 88 0 R 89 0 R 90 0 R 91 0 R 92 0 R 93 0 R 94 0 R 95 0 R 96 0 R 97 0 R 98 0 R 99 0 R 100 0 R 101 0 R 102 0 R 103 0 R 104 0 R 105 0 R 106 0 R 107 0 R 108 0 R 109 0 R 110 0 R 111 0 R 112 0 R 113 0 R 114 0 R 115 0 R 116 0 R 117 0 R 118 0 R 119 0 R 120 0 R 121 0 R 122 0 R 123 0 R ] /Contents [ 124 0 R 125 0 R ] /MediaBox [ 0 0 612 792 ] /Parent 9 0 R /Resources 126 0 R /Type /Page >>
endobj
48 0 obj
<< /Annots [ 127 0 R 128 0 R 129 0 R 130 0 R 131 0 R 132 0 R 133 0 R 134 0 R 135 0 R 136 0 R 137 0 R 138 0 R 139 0 R 140 0 R 141 0 R 142 0 R 143 0 R 144 0 R 145 0 R 146 0 R 147 0 R 148 0 R ] /Contents [ 149 0 R 150 0 R ] /MediaBox [ 0 0 612 792 ] /Parent 9 0 R /Resources 151 0 R /Type /Page >>
endobj
49 0 obj
<< /Annots [ 152 0 R 153 0 R ] /Contents [ 154 0 R 155 0 R ] /MediaBox [ 0 0 612 792 ] /Parent 9 0 R /Resources 156 0 R /Type /Page >>
endobj
50 0 obj
<< /Annots [ 157 0 R 158 0 R 159 0 R 160 0 R 161 0 R 162 0 R 163 0 R 164 0 R 165 0 R 166 0 R 167 0 R 168 0 R 169 0 R 170 0 R ] /Contents [ 171 0 R 172 0 R ] /MediaBox [ 0 0 612 792 ] /Parent 9 0 R /Resources 173 0 R /Type /Page >>
endobj
51 0 obj
<< /Annots [ 174 0 R 175 0 R 176 0 R 177 0 R 178 0 R 179 0 R 180 0 R 181 0 R 182 0 R 183 0 R 184 0 R 185 0 R 186 0 R 187 0 R 188 0 R 189 0 R 190 0 R 191 0 R 192 0 R 193 0 R 194 0 R 195 0 R 196 0 R 197 0 R 198 0 R 199 0 R 200 0 R 201 0 R 202 0 R 203 0 R 204 0 R 205 0 R 206 0 R ] /Contents [ 207 0 R 208 0 R ] /MediaBox [ 0 0 612 792 ] /Parent 9 0 R /Resources 209 0 R /Type /Page >>
endobj
52 0 obj
<< /Annots [ 210 0 R 211 0 R 212 0 R 213 0 R 214 0 R 215 0 R 216 0 R 217 0 R 218 0 R 219 0 R 220 0 R 221 0 R 222 0 R 223 0 R 224 0 R 225 0 R ] /Contents [ 226 0 R 227 0 R ] /MediaBox [ 0 0 612 792 ] /Parent 10 0 R /Resources 228 0 R /Type /Page >>
endobj
53 0 obj
<< /Annots [ 229 0 R ] /Contents [ 230 0 R 231 0 R ] /MediaBox [ 0 0 612 792 ] /Parent 10 0 R /Resources 232 0 R /Type /Page >>
endobj
54 0 obj
<< /Annots [ 233 0 R 234 0 R 235 0 R 236 0 R 237 0 R 238 0 R 239 0 R 240 0 R 241 0 R 242 0 R 243 0 R 244 0 R 245 0 R 246 0 R 247 0 R 248 0 R 249 0 R 250 0 R 251 0 R 252 0 R 253 0 R 254 0 R 255 0 R 256 0 R 257 0 R 258 0 R 259 0 R 260 0 R 261 0 R 262 0 R 263 0 R 264 0 R 265 0 R 266 0 R 267 0 R 268 0 R 269 0 R 270 0 R 271 0 R 272 0 R 273 0 R 274 0 R 275 0 R 276 0 R 277 0 R 278 0 R 279 0 R 280 0 R 281 0 R 282 0 R 283 0 R 284 0 R 285 0 R 286 0 R ] /Contents [ 287 0 R 288 0 R ] /MediaBox [ 0 0 612 792 ] /Parent 10 0 R /Resources 289 0 R /Type /Page >>
endobj
55 0 obj
<< /Annots [ 290 0 R 291 0 R 292 0 R 293 0 R 294 0 R 295 0 R 296 0 R 297 0 R 298 0 R 299 0 R 300 0 R 301 0 R 302 0 R 303 0 R 304 0 R 305 0 R 306 0 R 307 0 R 308 0 R 309 0 R 310 0 R 311 0 R 312 0 R 313 0 R 314 0 R 315 0 R 316 0 R 317 0 R 318 0 R 319 0 R 320 0 R 321 0 R 322 0 R 323 0 R 324 0 R ] /Contents [ 325 0 R 326 0 R ] /MediaBox [ 0 0 612 792 ] /Parent 10 0 R /Resources 327 0 R /Type /Page >>
endobj
56 0 obj
<< /Limits [ (Doc-Start) (cite.2015vqa) ] /Names [ (Doc-Start) 328 0 R (Hfootnote.1) 329 0 R (cite.1995boosting) 330 0 R (cite.2000gradient) 331 0 R (cite.2006some) 332 0 R (cite.2015vqa) 333 0 R ] >>
endobj
57 0 obj
<< /Limits [ (cite.2016xgboost) (cite.2017hat) ] /Names [ (cite.2016xgboost) 334 0 R (cite.2016yin) 335 0 R (cite.2017analysis) 336 0 R (cite.2017clevr) 337 0 R (cite.2017exploring) 338 0 R (cite.2017hat) 339 0 R ] >>
endobj
58 0 obj
<< /Limits [ (cite.2017mfh) (cite.2018bottomup) ] /Names [ (cite.2017mfh) 340 0 R (cite.2017n2nmn) 341 0 R (cite.2017rn) 342 0 R (cite.2018BAN) 343 0 R (cite.2018annotation) 344 0 R (cite.2018bottomup) 345 0 R ] >>
endobj
59 0 obj
<< /Limits [ (cite.2018film) (cite.2019SCR) ] /Names [ (cite.2018film) 346 0 R (cite.2018overcoming) 347 0 R (cite.2018stacknmn) 348 0 R (cite.2018vqacp) 349 0 R (cite.2018vqax) 350 0 R (cite.2019SCR) 351 0 R ] >>
endobj
60 0 obj
<< /Limits [ (cite.2019don) (cite.2019murel) ] /Names [ (cite.2019don) 352 0 R (cite.2019gqa) 353 0 R (cite.2019hint) 354 0 R (cite.2019interpretable) 355 0 R (cite.2019lcgn) 356 0 R (cite.2019murel) 357 0 R ] >>
endobj
61 0 obj
<< /Limits [ (cite.2019nscl) (cite.2020cf-vqa) ] /Names [ (cite.2019nscl) 358 0 R (cite.2019right) 359 0 R (cite.2019rubi) 360 0 R (cite.2020DLP) 361 0 R (cite.2020bbn) 362 0 R (cite.2020cf-vqa) 363 0 R ] >>
endobj
62 0 obj
<< /Limits [ (cite.2020counterfactual) (cite.2020negative) ] /Names [ (cite.2020counterfactual) 364 0 R (cite.2020decoupling) 365 0 R (cite.2020end) 366 0 R (cite.2020lff) 367 0 R (cite.2020mfe) 368 0 R (cite.2020negative) 369 0 R ] >>
endobj
63 0 obj
<< /Limits [ (cite.2020reducing) (equation.3.1) ] /Names [ (cite.2020reducing) 370 0 R (cite.bai2020attention) 371 0 R (cite.gao2019multi) 372 0 R (cite.schapire1990strength) 373 0 R (cite.teney2020value) 374 0 R (equation.3.1) 375 0 R ] >>
endobj
64 0 obj
<< /Limits [ (equation.3.2) (equation.4.14) ] /Names [ (equation.3.2) 376 0 R (equation.4.10) 377 0 R (equation.4.11) 378 0 R (equation.4.12) 379 0 R (equation.4.13) 380 0 R (equation.4.14) 381 0 R ] >>
endobj
65 0 obj
<< /Limits [ (equation.4.15) (equation.4.6) ] /Names [ (equation.4.15) 382 0 R (equation.4.16) 383 0 R (equation.4.3) 384 0 R (equation.4.4) 385 0 R (equation.4.5) 386 0 R (equation.4.6) 387 0 R ] >>
endobj
66 0 obj
<< /Limits [ (equation.4.7) (equation.5.19) ] /Names [ (equation.4.7) 388 0 R (equation.4.8) 389 0 R (equation.4.9) 390 0 R (equation.5.17) 391 0 R (equation.5.18) 392 0 R (equation.5.19) 393 0 R ] >>
endobj
67 0 obj
<< /Limits [ (equation.5.20) (page.1) ] /Names [ (equation.5.20) 394 0 R (figure.1) 395 0 R (figure.2) 396 0 R (figure.3) 397 0 R (figure.4) 398 0 R (page.1) 399 0 R ] >>
endobj
68 0 obj
<< /Limits [ (page.10) (page.6) ] /Names [ (page.10) 400 0 R (page.2) 401 0 R (page.3) 402 0 R (page.4) 403 0 R (page.5) 404 0 R (page.6) 405 0 R ] >>
endobj
69 0 obj
<< /Limits [ (page.7) (section.1) ] /Names [ (page.7) 406 0 R (page.8) 407 0 R (page.9) 408 0 R (section*.1) 409 0 R (section*.2) 410 0 R (section.1) 411 0 R ] >>
endobj
70 0 obj
<< /Limits [ (section.2) (subfigure.1.1) ] /Names [ (section.2) 412 0 R (section.3) 413 0 R (section.4) 414 0 R (section.5) 415 0 R (section.6) 416 0 R (subfigure.1.1) 417 0 R ] >>
endobj
71 0 obj
<< /Limits [ (subfigure.1.2) (subsection.2.1) ] /Names [ (subfigure.1.2) 418 0 R (subfigure.2.1) 419 0 R (subfigure.2.2) 420 0 R (subfigure.2.3) 421 0 R (subfigure.2.4) 422 0 R (subsection.2.1) 423 0 R ] >>
endobj
72 0 obj
<< /Limits [ (subsection.2.2) (subsection.4.3) ] /Names [ (subsection.2.2) 424 0 R (subsection.3.1) 425 0 R (subsection.3.2) 426 0 R (subsection.4.1) 427 0 R (subsection.4.2) 428 0 R (subsection.4.3) 429 0 R ] >>
endobj
73 0 obj
<< /Limits [ (subsection.5.1) (table.1) ] /Names [ (subsection.5.1) 430 0 R (subsection.5.2) 431 0 R (subsection.5.3) 432 0 R (subsection.5.4) 433 0 R (subsection.5.5) 434 0 R (table.1) 435 0 R ] >>
endobj
74 0 obj
<< /Limits [ (table.2) (table.4) ] /Names [ (table.2) 436 0 R (table.3) 437 0 R (table.4) 438 0 R ] >>
endobj
75 0 obj
<< /BaseFont /JZKXYQ+CMR8 /FirstChar 49 /FontDescriptor 439 0 R /LastChar 53 /Subtype /Type1 /ToUnicode 440 0 R /Type /Font /Widths 441 0 R >>
endobj
76 0 obj
<< /BaseFont /EAAUWX+CMMI8 /FirstChar 59 /FontDescriptor 442 0 R /LastChar 59 /Subtype /Type1 /ToUnicode 443 0 R /Type /Font /Widths 444 0 R >>
endobj
77 0 obj
<< /BaseFont /FRNIHB+CMSY8 /FirstChar 3 /FontDescriptor 445 0 R /LastChar 3 /Subtype /Type1 /ToUnicode 446 0 R /Type /Font /Widths 447 0 R >>
endobj
78 0 obj
<< /BaseFont /GUOWTK+CMSY6 /FirstChar 3 /FontDescriptor 448 0 R /LastChar 3 /Subtype /Type1 /ToUnicode 449 0 R /Type /Font /Widths 450 0 R >>
endobj
79 0 obj
<< /BaseFont /YYJESM+NimbusRomNo9L-Medi /Encoding 451 0 R /FirstChar 2 /FontDescriptor 452 0 R /LastChar 173 /Subtype /Type1 /ToUnicode 453 0 R /Type /Font /Widths 454 0 R >>
endobj
80 0 obj
<< /BaseFont /UAFJQS+NimbusRomNo9L-Regu /Encoding 451 0 R /FirstChar 2 /FontDescriptor 455 0 R /LastChar 150 /Subtype /Type1 /ToUnicode 456 0 R /Type /Font /Widths 457 0 R >>
endobj
81 0 obj
<< /BaseFont /GWKGYY+NimbusMonL-Regu /Encoding 451 0 R /FirstChar 44 /FontDescriptor 458 0 R /LastChar 122 /Subtype /Type1 /ToUnicode 459 0 R /Type /Font /Widths 460 0 R >>
endobj
82 0 obj
<< /BaseFont /XNRGJT+NimbusRomNo9L-ReguItal /Encoding 451 0 R /FirstChar 2 /FontDescriptor 461 0 R /LastChar 121 /Subtype /Type1 /ToUnicode 462 0 R /Type /Font /Widths 463 0 R >>
endobj
83 0 obj
<< /BBox [ 0 34.5891 612 361.854 ] /Filter /FlateDecode /FormType 1 /PTEX.FileName (D:/\210\221\232\204论\226\207/iccv/iccv2021AuthorKit/LaTeX/figure/distribution_bias1.pdf) /PTEX.InfoDict 464 0 R /PTEX.PageNumber 1 /Resources << /Font << /F1 465 0 R >> /XObject << /Image1 466 0 R /Image2 467 0 R /Image3 468 0 R /Image4 469 0 R /Image5 470 0 R >> >> /Subtype /Form /Type /XObject /Length 1325 >>
stream
xYMo6W\`O`a ==,vCIMgi&@`ӶHJViZ^&_.{^Ť\J)>xf|ETZOZ*٫QKw8*2dTy0L{v
c_yϿC_Oҡ*G