%PDF-1.4
%
1 0 obj
<< /CP2 3 0 R /FICL:Enfocus 4 0 R /Metadata 5 0 R /OpenAction [ 6 0 R /XYZ null null null ] /PageLayout /SinglePage /PageMode /UseNone /Pages 7 0 R /Type /Catalog >>
endobj
2 0 obj
<< /Author (Cheng-Kun Yang; Ji-Jia Wu; Kai-Syun Chen; Yung-Yu Chuang; Yen-Yu Lin) /Producer (pikepdf 5.1.3) /Subject (IEEE Conference on Computer Vision and Pattern Recognition) /Title (An MIL-Derived Transformer for Weakly Supervised Point Cloud Segmentation) >>
endobj
3 0 obj
<< /PrivateData 8 0 R /PrivateDataFilter 9 0 R >>
endobj
4 0 obj
<< /CertifiedWF 10 0 R >>
endobj
5 0 obj
<< /Subtype /XML /Type /Metadata /Length 1215 >>
stream
An MIL-Derived Transformer for Weakly Supervised Point Cloud SegmentationCheng-Kun Yang Ji-Jia Wu Kai-Syun Chen Yung-Yu Chuang Yen-Yu LinIEEE Conference on Computer Vision and Pattern Recognition
endstream
endobj
6 0 obj
<< /Contents 11 0 R /MediaBox [ 0 0 612 792 ] /Parent 7 0 R /Resources << /ExtGState 12 0 R /Font 13 0 R /ProcSet [ /PDF /Text ] /XObject << /TXbdxF74oYvgrH4SH-tNFQ 14 0 R /uijd4MtnezeaR4tYoaKB6Q 15 0 R >> >> /Rotate 0 /Type /Page >>
endobj
7 0 obj
<< /Count 10 /Kids [ 6 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 ] /Type /Pages >>
endobj
8 0 obj
<< /16014 25 0 R /16015 26 0 R >>
endobj
9 0 obj
[ /FlateDecode ]
endobj
10 0 obj
<< /FileMapping 27 0 R /Sessions 28 0 R /Stamp [ 132 150 124 233 112 243 76 75 23 9 87 38 0 17 106 117 19 212 77 206 ] /Version 1 >>
endobj
11 0 obj
<< /Length 19430 >>
stream
q
q
q
q 0.1 0 0 0.1 0 0 cm
/R7 gs
0 g
q
10 0 0 10 0 0 cm BT
/R8 14.3462 Tf
1 0 0 1 51.4691 675.067 Tm
[(An)-250.007(MIL-Deri)10.009(v)9.99625(ed)-250.004(T)73.9916(ransf)24.9898(ormer)-249.998(f)24.9923(or)-249.995(W)65.0011(eakly)-250(Super)10.0031(vised)-249.99(P)20.0061(oint)-249.991(Cloud)-250.012(Segmentation)]TJ
/R10 11.9552 Tf
30.884 -37.8582 Td
[(Cheng-K)15(un)-250.012(Y)99.9837(ang)]TJ
/R12 7.9701 Tf
83.2918 4.33906 Td
[(1)-0.298663]TJ
/R10 11.9552 Tf
19.6762 -4.33906 Td
[(Ji-Jia)-249.995(W)50.0056(u)]TJ
/R12 7.9701 Tf
44.891 4.33906 Td
[(2)-0.300194]TJ
/R10 11.9552 Tf
19.6758 -4.33906 Td
[(Kai-Syun)-249.983(Chen)]TJ
/R12 7.9701 Tf
74.05 4.33906 Td
[(2)-0.300194]TJ
/R10 11.9552 Tf
19.6762 -4.33906 Td
[(Y)110.996(ung-Y)111.012(u)-250.002(Chuang)]TJ
/R12 7.9701 Tf
82.682 4.33906 Td
[(1)-0.300194]TJ
/R10 11.9552 Tf
19.6758 -4.33906 Td
[(Y)99.9826(en-Y)110.994(u)-250.002(Lin)]TJ
/R12 7.9701 Tf
55.5801 4.33906 Td
[(2)-0.300194]TJ
/R14 7.9701 Tf
4.23398 0 Td
[(;)-0.0980226]TJ
/R12 7.9701 Tf
2.35195 0 Td
[(3)-0.300194]TJ
-444.191 -13.948 Td
[(1)-0.300194]TJ
/R10 11.9552 Tf
4.73203 -4.33789 Td
[(National)-250(T)80.0137(aiw)10(an)-250.009(Uni)24.9946(v)14.9862(ersity)]TJ
/R12 7.9701 Tf
146.868 4.33789 Td
[(2)-0.300194]TJ
/R10 11.9552 Tf
4.7332 -4.33789 Td
[(National)-250(Y)99.9847(ang)-249.987(Ming)-250(Chiao)-249.987(T)44.9881(ung)-249.989(Uni)24.9957(v)14.9851(ersity)]TJ
/R12 7.9701 Tf
224.922 4.33789 Td
[(3)-0.300194]TJ
/R10 11.9552 Tf
4.7332 -4.33789 Td
[(Academia)-249.993(Sinica)]TJ
/R8 11.9552 Tf
-303.94 -41.0473 Td
(Abstract)Tj
/R16 9.9626 Tf
-83.9277 -20.8949 Td
[(W)91.9865(e)-271.998(addr)36.9951(ess)-271.006(weakly)-271.982(supervised)-270.987(point)-272.004(cloud)-272.016(se)39.9958(gmentation)]TJ
-11.9551 -11.9551 Td
[(by)-448.003(pr)44.9839(oposing)-448.005(a)-447.999(ne)15.0183(w)-447.99(model,)-496.997(MIL-derived)-447.984(tr)14.9901(ansformer)111.001(,)-497.993(to)]TJ
11.9551 TL
T*[(mine)-277.002(additional)-276.993(supervisory)-277.019(signals.)-390.989(F)45.017(ir)10.0106(st,)-284.016(the)-277.005(tr)14.9901(ansformer)]TJ
T*[(model)-235.004(is)-235.986(derived)-235.01(based)-235.002(on)-235.99(multiple)-234.983(instance)-234.993(learning)-236.002(\(MIL\))]TJ
T*[(to)-505.002(e)19.9924(xplor)36.9932(e)-505.018(pair)19.9807(-wise)-504.995(cloud-le)14.995(vel)-505.995(supervision,)-568.996(wher)36.9938(e)-505.019(two)]TJ
11.9563 TL
T*[(clouds)-292.986(of)-292.998(the)-293.982(s)0.986355(ame)-293.986(cate)39.997(gory)-292.998(yield)-292.999(a)-293.005(positive)-292.981(ba)10.0081(g)-293.985(while)-292.981(two)]TJ
11.9547 TL
T*[(of)-247.986(dif)18.0166(fer)36.9834(ent)-247.99(classes)-248.015(pr)44.9839(oduce)-247.992(a)-247.993(ne)39.9946(gative)-247.982(ba)10.0081(g)14.9828(.)-309.005(It)-247.993(le)14.9803(ver)15.0147(a)10.0032(g)10.0032(es)-247.995(not)]TJ
T*[(only)-298.016(individual)-298.019(cloud)-298.01(annotations)-298.005(b)20.0016(ut)-298.016(also)-298.007(pair)19.9795(-wise)-298.009(cloud)]TJ
T*[(semantics)-262.004(for)-262.016(model)-262.019(optimization.)-345.988(Second,)-265(Adaptive)-262.01(global)]TJ
T*[(weighted)-365.984(pooling)-366.017(\(AdaGWP\))-366.001(is)-366.004(inte)40.008(gr)14.9901(ated)-366.005(into)-365.993(our)-366(tr)14.9901(ans-)]TJ
T*[(former)-264.012(model)-263(to)-263.983(r)37.0196(eplace)-263.011(max)-263.991(pooling)-263.015(and)-264.02(aver)15.0196(a)10.0032(g)10.0032(e)-264(pooling)15.0171(.)]TJ
11.9563 TL
T*[(It)-206.981(intr)44.9986(oduces)-207.012(learnable)-206.992(weights)-206.005(to)-207.012(r)37.0196(e-scale)-206.983(lo)10.0106(gits)-207(in)-207.012(the)-207.017(class)]TJ
11.9551 TL
T*[(activation)-194.986(maps.)-292.008(It)-194.983(is)-195.013(mor)36.9877(e)-196.011(r)45.0182(ob)20.0065(ust)-195.008(to)-195.014(noise)-195.007(while)-196.016(disco)10.0179(vering)]TJ
T*[(mor)36.9883(e)-335.987(complete)-334.988(for)36.9926(e)39.9884(gr)44.9839(ound)-336.003(points)-335.007(under)-336.005(weak)-336.013(supervision.)]TJ
T*[(Thir)37.0067(d,)-392.006(we)-364.008(perform)-363.99(point)-363.008(subsampling)-363.998(and)-364.003(enfor)36.9938(ce)-363.988(featur)37.0012(e)]TJ
T*[(equivariance)-415(between)-415.992(the)-414.982(original)-416(and)-415.014(subsampled)-416.004(point)]TJ
11.9559 TL
T*[(clouds)-435.002(for)-435.007(r)37.0196(e)39.9884(gularization.)-864.998(The)-435.011(pr)44.9839(oposed)-435.003(method)-435.006(is)-435.013(end-)]TJ
11.9551 TL
T*[(to-end)-374.001(tr)14.9914(ainable)-373.985(and)-374.001(is)-374.003(g)10.0032(ener)15.0196(al)-374.004(because)-374.009(it)-374.001(can)-374.011(work)-373.996(with)]TJ
T*[(dif)18.0166(fer)36.9834(ent)-297.001(bac)20.0022(kbones)-296.997(with)-297.015(diver)10.0081(se)-297.008(types)-297(of)-296.997(weak)-297(supervision)]TJ
T*[(signals,)-202(including)-190.985(spar)9.98118(sely)-190.983(annotated)-190.006(points)-190.993(and)-191.012(cloud-le)14.995(vel)]TJ
T*[(labels.)-470.991(The)-304.012(e)19.9918(xperiments)-304(show)-303.018(that)-304.003(it)-304.013(ac)15.0171(hie)14.9852(ves)-303.991(state-of-the-)]TJ
T*[(art)-393.994(performance)-393.983(on)-394.983(the)-394.005(S3DIS)-393.998(and)-393.998(ScanNet)-395.005(benc)15.0196(hmarks.)]TJ
11.9559 TL
T*[(The)-352.985(sour)36.9963(ce)-353.008(code)-352.999(will)-353.005(be)-352.999(available)-353.006(at)]TJ
ET
Q
0.925 0 0.545 rg
q
10 0 0 10 0 0 cm BT
/R18 8.9664 Tf
1 0 0 1 203.87 298.306 Tm
[(https)-38.9938(:)-39.0183(/)-40.0203(/)-42.9827(github)-39.0183(.)]TJ
-153.758 -11.9551 Td
[(com/)-3.98351(jimmy15923/)-3.9944(wspss_mil_transformer)]TJ
ET
Q
0 g
q
10 0 0 10 0 0 cm BT
/R16 8.9664 Tf
1 0 0 1 244.18 286.351 Tm
(.)Tj
/R8 11.9552 Tf
-194.068 -30.4199 Td
[(1.)-249.99(Intr)18.0146(oduction)]TJ
/R10 9.9626 Tf
11.9551 -19.5141 Td
[(Point)-358.019(clouds)-359.014(capture)-358.017(geometric)-357.998(characteristics)-358.006(and)-358.994(sur)20.004(-)]TJ
-11.9551 -11.9551 Td
[(f)9.99527(ace)-264.988(conte)14.9865(xt,)-268.019(and)-265.008(hence)-265.017(serv)14.9816(e)-265.02(as)-265.013(an)-265.013(essential)-264.99(representation)]TJ
11.9551 TL
T*[(for)-279.013(man)14.9908(y)-279.007(3D)-279.004(vision)-279.016(applications)-279.01(such)-279.98(as)-279.01(scene)-279.015(understand-)]TJ
T*[(ing)-247.981([)]TJ
ET
Q
0 1 0 rg
q
10 0 0 10 0 0 cm BT
/R10 9.9626 Tf
1 0 0 1 68.6324 200.552 Tm
(6)Tj
ET
Q
0 g
q
10 0 0 10 0 0 cm BT
/R10 9.9626 Tf
1 0 0 1 73.6137 200.552 Tm
(,)Tj
ET
Q
0 1 0 rg
q
10 0 0 10 0 0 cm BT
/R10 9.9626 Tf
1 0 0 1 77.3797 200.552 Tm
(22)Tj
ET
Q
0 g
q
10 0 0 10 0 0 cm BT
/R10 9.9626 Tf
1 0 0 1 87.3422 200.552 Tm
(,)Tj
ET
Q
0 1 0 rg
q
10 0 0 10 0 0 cm BT
/R10 9.9626 Tf
1 0 0 1 91.098 200.552 Tm
(28)Tj
ET
Q
0 g
q
10 0 0 10 0 0 cm BT
/R10 9.9626 Tf
1 0 0 1 101.061 200.552 Tm
[(],)-248(autonomous)-249.005(v)14.9828(ehicles)-247.986([)]TJ
ET
Q
0 1 0 rg
q
10 0 0 10 0 0 cm BT
/R10 9.9626 Tf
1 0 0 1 198.814 200.552 Tm
(4)Tj
ET
Q
0 g
q
10 0 0 10 0 0 cm BT
/R10 9.9626 Tf
1 0 0 1 203.795 200.552 Tm
(,)Tj
ET
Q
0 1 0 rg
q
10 0 0 10 0 0 cm BT
/R10 9.9626 Tf
1 0 0 1 207.551 200.552 Tm
(5)Tj
ET
Q
0 g
q
10 0 0 10 0 0 cm BT
/R10 9.9626 Tf
1 0 0 1 212.532 200.552 Tm
[(],)-249.017(and)-247.993(robotics)-248.99([)]TJ
ET
Q
0 1 0 rg
q
10 0 0 10 0 0 cm BT
/R10 9.9626 Tf
1 0 0 1 275.575 200.552 Tm
(9)Tj
ET
Q
0 g
q
10 0 0 10 0 0 cm BT
/R10 9.9626 Tf
1 0 0 1 280.557 200.552 Tm
(].)Tj
-230.445 -11.9559 Td
[(Point)-398.013(cloud)-397.014(se)15.0183(gmentation)-398.006(aims)-397.991(to)-397.02(identify)-397.983(points)-398.017(belong-)]TJ
11.9551 TL
T*[(ing)-326.987(to)-326.012(semantic)-327.019(cate)15.011(gories)-326.008(of)-327.003(interest.)-540.004(It)-325.98(of)25.0179(fers)-327.009(point-le)25.0056(v)14.9828(el)]TJ
T*[(recognition,)-315.988(thereby)-302.998(being)-302.99(an)-301.988(intrinsic)-303.011(component)-303.011(of)-303.006(point)]TJ
T*[(cloud)-209.006(analysis.)-297.018(Ho)24.9836(we)25.0142(v)14.9828(er)39.986(,)-216.981(learning)-209.987(a)-208.99(se)15.0183(gmentation)-209.019(model)-210.019(re-)]TJ
T*[(lies)-272.017(on)-273.003(training)-272.004(data)-271.99(with)-272.011(point-le)25.0056(v)14.9828(el)-273.011(annotations.)-376.008(The)-273.008(high)]TJ
11.9559 TL
T*[(annotation)-341.983(cost)-342.008(poses)-342.003(an)-343(obstacle)-342.004(to)-342.009(this)-341.997(task.)-586.983(T)79.9916(o)-342.016(address)]TJ
11.9551 TL
T*[(this)-371.991(issue,)-401.989(e)15.0128(xisting)-370.993(weakly)-372.005(supervised)-371.009(methods)-372.019(deri)24.9811(v)14.9828(e)-371.982(the)]TJ
T*[(se)15.0189(gmentation)-404.005(model)-403.987(with)-403.989(dif)24.9848(ferent)-404.013(weak)-404.011(supervisory)-404.016(sig-)]TJ
T*[(nals,)-381.018(such)-354.008(as)-354.998(partially)-353.992(labeled)-354.989(points)-353.984([)]TJ
ET
Q
0 1 0 rg
q
10 0 0 10 0 0 cm BT
/R10 9.9626 Tf
1 0 0 1 206.316 92.9551 Tm
(26)Tj
ET
Q
0 g
q
10 0 0 10 0 0 cm BT
/R10 9.9626 Tf
1 0 0 1 216.278 92.9551 Tm
(,)Tj
ET
Q
0 1 0 rg
q
10 0 0 10 0 0 cm BT
/R10 9.9626 Tf
1 0 0 1 220.761 92.9551 Tm
(42)Tj
ET
Q
0 g
q
10 0 0 10 0 0 cm BT
/R10 9.9626 Tf
1 0 0 1 230.724 92.9551 Tm
(,)Tj
ET
Q
0 1 0 rg
q
10 0 0 10 0 0 cm BT
/R10 9.9626 Tf
1 0 0 1 235.197 92.9551 Tm
(46)Tj
ET
Q
0 g
q
10 0 0 10 0 0 cm BT
/R10 9.9626 Tf
1 0 0 1 245.159 92.9551 Tm
(,)Tj
ET
Q
0 1 0 rg
q
10 0 0 10 0 0 cm BT
/R10 9.9626 Tf
1 0 0 1 249.643 92.9551 Tm
(47)Tj
ET
Q
0 g
q
10 0 0 10 0 0 cm BT
/R10 9.9626 Tf
1 0 0 1 259.605 92.9551 Tm
[(],)-380.995(sub-)]TJ
-209.493 -11.9551 Td
[(cloud)-227.983(le)25.0185(v)14.9828(el)-227.998(annotations)-229.016([)]TJ
ET
Q
0 1 0 rg
q
10 0 0 10 0 0 cm BT
/R10 9.9626 Tf
1 0 0 1 147.297 81 Tm
(38)Tj
ET
Q
0 g
q
10 0 0 10 0 0 cm BT
/R10 9.9626 Tf
1 0 0 1 157.26 81 Tm
[(])-228.005(or)-228(scene)-228.004(le)25.0179(v)14.9828(el)-227.999(annotations)-229.016([)]TJ
ET
Q
0 1 0 rg
q
10 0 0 10 0 0 cm BT
/R10 9.9626 Tf
1 0 0 1 270.595 81 Tm
(31)Tj
ET
Q
0 g
q
10 0 0 10 0 0 cm BT
/R10 9.9626 Tf
1 0 0 1 280.557 81 Tm
(].)Tj
40.2602 501.215 Td
[(T)79.9916(o)-477.994(compensate)-478.016(for)-478.981(the)-477.989(lack)-477.994(of)-477.999(complete)-478.994(annotations,)]TJ
-11.9551 -11.9551 Td
[(weakly)-330.011(supervised)-329.996(point)-329.994(cloud)-330.004(se)15.0171(gmentation)-330.999(methods)-329.989([)]TJ
ET
Q
0 1 0 rg
q
10 0 0 10 0 0 cm BT
/R10 9.9626 Tf
1 0 0 1 532.662 570.26 Tm
(26)Tj
ET
Q
0 g
q
10 0 0 10 0 0 cm BT
/R10 9.9626 Tf
1 0 0 1 542.624 570.26 Tm
(,)Tj
ET
Q
0 1 0 rg
q
10 0 0 10 0 0 cm BT
/R10 9.9626 Tf
1 0 0 1 308.862 558.305 Tm
(31)Tj
ET
Q
0 g
q
10 0 0 10 0 0 cm BT
/R10 9.9626 Tf
1 0 0 1 318.825 558.305 Tm
(,)Tj
ET
Q
0 1 0 rg
q
10 0 0 10 0 0 cm BT
/R10 9.9626 Tf
1 0 0 1 322.77 558.305 Tm
(38)Tj
ET
Q
0 g
q
10 0 0 10 0 0 cm BT
/R10 9.9626 Tf
1 0 0 1 332.732 558.305 Tm
(,)Tj
ET
Q
0 1 0 rg
q
10 0 0 10 0 0 cm BT
/R10 9.9626 Tf
1 0 0 1 336.688 558.305 Tm
(42)Tj
ET
Q
0 g
q
10 0 0 10 0 0 cm BT
/R10 9.9626 Tf
1 0 0 1 346.65 558.305 Tm
(,)Tj
ET
Q
0 1 0 rg
q
10 0 0 10 0 0 cm BT
/R10 9.9626 Tf
1 0 0 1 350.595 558.305 Tm
(46)Tj
ET
Q
0 g
q
10 0 0 10 0 0 cm BT
/R10 9.9626 Tf
1 0 0 1 360.558 558.305 Tm
(,)Tj
ET
Q
0 1 0 rg
q
10 0 0 10 0 0 cm BT
/R10 9.9626 Tf
1 0 0 1 364.513 558.305 Tm
(47)Tj
ET
Q
0 g
q
10 0 0 10 0 0 cm BT
/R10 9.9626 Tf
1 0 0 1 374.475 558.305 Tm
[(])-273.996(mak)10.0106(e)-275.018(the)-275.003(most)-274.003(of)-275.013(weakly)-273.984(labeled)-275.003(data)-275.008(by)]TJ
-65.6133 -11.9551 Td
[(dif)24.986(ferent)-202.008(techniques)-202.018(such)-201.993(as)-200.986(graph)-201.996(propag)4.99426(ation,)-211.979(permutation)]TJ
11.9559 TL
T*[(consistenc)15.0073(y)65.0137(,)-446.002(and)-406.986(object)-407.015(proposals.)-779.992(Despite)-407.01(ef)25.0056(fecti)25.0154(v)14.9828(eness,)]TJ
11.9551 TL
T*[(these)-223.992(methods)-224.004(use)-223.014(only)]TJ
/R16 9.9626 Tf
94.1488 0 Td
[(intr)15.0024(a-cloud)]TJ
/R10 9.9626 Tf
46.9121 0 Td
[(information:)-297(The)-223.997(super)20.0065(-)]TJ
-141.061 -11.9551 Td
[(visory)-218.998(signals)-218.99(are)-218.015(grabbed)-219.012(from)-219.012(point)-218.993(clouds)-218.998(independently)65.0137(.)]TJ
T*[(Inspired)-195.009(by)-193.997(image)-195.012(co-se)15.0171(gmentation)-194.982([)]TJ
ET
Q
0 1 0 rg
q
10 0 0 10 0 0 cm BT
/R10 9.9626 Tf
1 0 0 1 452.602 498.529 Tm
(13)Tj
ET
Q
0 g
q
10 0 0 10 0 0 cm BT
/R10 9.9626 Tf
1 0 0 1 462.565 498.529 Tm
(,)Tj
ET
Q
0 1 0 rg
q
10 0 0 10 0 0 cm BT
/R10 9.9626 Tf
1 0 0 1 465.424 498.529 Tm
(45)Tj
ET
Q
0 g
q
10 0 0 10 0 0 cm BT
/R10 9.9626 Tf
1 0 0 1 475.387 498.529 Tm
[(])-194.992(and)-194.982(cross-image)]TJ
-166.525 -11.9547 Td
[(pattern)-304.015(mining)-303.986([)]TJ
ET
Q
0 1 0 rg
q
10 0 0 10 0 0 cm BT
/R10 9.9626 Tf
1 0 0 1 374.137 486.574 Tm
(32)Tj
ET
Q
0 g
q
10 0 0 10 0 0 cm BT
/R10 9.9626 Tf
1 0 0 1 384.1 486.574 Tm
[(],)-317.986(we)-303.991(aim)-304.013(to)-304.018(e)15.0122(xplore)]TJ
/R16 9.9626 Tf
85.1363 0 Td
[(inter)20.0163(-cloud)]TJ
/R10 9.9626 Tf
47.1039 0 Td
(seman-)Tj
-207.478 -11.9551 Td
[(tics)-286.014(to)-286.018(supervise)-286.009(se)15.0196(gmentation)-285.987(model)-286.006(training.)-418.999(T)79.9916(o)-285.987(this)-286.006(end,)]TJ
11.9559 TL
T*[(we)-228.984(generalize)-230(the)-229.013(transformer)-229.006(model)-228.996([)]TJ
ET
Q
0 1 0 rg
q
10 0 0 10 0 0 cm BT
/R10 9.9626 Tf
1 0 0 1 459.706 462.663 Tm
(34)Tj
ET
Q
0 g
q
10 0 0 10 0 0 cm BT
/R10 9.9626 Tf
1 0 0 1 469.668 462.663 Tm
[(])-230.006(to)-229.011(w)10.0032(ork)-229.016(on)-229.991(paired)]TJ
-160.806 -11.9551 Td
[(point)-387.984(clouds)-387.989(and)-388.989(formulate)-387.994(the)-388.004(problem)-387.989(as)-388.014(a)]TJ
/R16 9.9626 Tf
188.657 0 Td
[(multiple)-388.014(in-)]TJ
-188.657 -11.9551 Td
[(stance)-261(learning)-261.998(\(MIL\))]TJ
/R10 9.9626 Tf
90.8219 0 Td
([)Tj
ET
Q
0 1 0 rg
q
10 0 0 10 0 0 cm BT
/R10 9.9626 Tf
1 0 0 1 403.002 438.753 Tm
(27)Tj
ET
Q
0 g
q
10 0 0 10 0 0 cm BT
/R10 9.9626 Tf
1 0 0 1 412.964 438.753 Tm
[(])-261.02(task.)-344.987(It)-261.01(follo)24.9958(ws)-261.986(that)-261(our)-261.01(method)]TJ
-104.102 -11.9551 Td
[(can)-244.981(use)-245.011(both)-244.996(intra-cloud)-244.991(and)-245.013(inter)20.0065(-cloud)-245.003(information)-245.013(to)-245.008(bet-)]TJ
11.9547 TL
T*[(ter)-249.997(accomplish)-250.002(weakly)-249.985(supervised)-250.007(se)15.0171(gmentation.)]TJ
11.9551 -15.6402 Td
[(Speci\002cally)65.0137(,)-567(we)-503.017(de)25.0154(v)14.9828(elop)-503.98(an)-503.014(MIL-deri)24.9909(v)14.9828(ed)-503.994(transformer)]TJ
-11.9551 -11.9551 Td
[(where)-369.002(MIL)-368.005(addresses)-368.985(the)-368.987(uncertainty)-369.012(of)-368.997(weak)-368.017(labels.)-667.006(As)]TJ
T*[(sho)24.9934(wn)-336.995(in)-336.99(Figure)]TJ
ET
Q
1 0 0 rg
q
10 0 0 10 0 0 cm BT
/R10 9.9626 Tf
1 0 0 1 378.451 375.293 Tm
(1)Tj
ET
Q
0 g
q
10 0 0 10 0 0 cm BT
/R10 9.9626 Tf
1 0 0 1 383.432 375.293 Tm
[(,)-358.996(we)-337.005(apply)-336.983(the)-336.013(transformer)-336.988(to)-336.993(tw)10.0081(o)-336.998(point)]TJ
-74.5699 -11.9551 Td
[(clouds)-397.007(of)-396.99(the)-398.005(same)-396.983(cate)15.0122(gory)64.9794(.)-751.987(One)-396.988(cloud)-397.012(is)-397.017(treated)-396.988(as)-398.012(an)]TJ
11.9563 TL
T*[(anchor)-358.016(with)-357.997(each)-357.987(of)-358.996(its)-358(points)-357.984(being)-358(a)]TJ
/R16 9.9626 Tf
160.251 0 Td
(query)Tj
/R10 9.9626 Tf
26.2531 0 Td
[(in)-358.009(the)-358.009(trans-)]TJ
-186.504 -11.9547 Td
[(former)55.0129(.)-303.006(The)-231.015(other)-229.991(cloud)-231.003(serv)14.9828(es)-230(as)-231.02(a)-230.006(reference)-231.006(where)-230.006(each)-230.991(of)]TJ
11.9551 TL
T*[(its)-306.008(points)-305.993(forms)-306.01(a)]TJ
/R16 9.9626 Tf
73.6359 0 Td
[(k)9.99343(e)29.9901(y-value)]TJ
/R10 9.9626 Tf
40.818 0 Td
[(pair)55.0153(.)-477.994(The)-305.983(transformer)-306.013(encoder)]TJ
-114.454 -11.9551 Td
[(and)-271.989(decoder)-273.016(are)-272.006(applied)-272.013(to)-273.003(the)-271.984(reference)-272.999(and)-271.989(the)-271.984(anchor)-273.008(re-)]TJ
T*[(specti)24.9934(v)14.9828(ely)65.0161(.)-440.003(Through)-292.985(the)-293.003(cross-attention)-293.01([)]TJ
ET
Q
0 1 0 rg
q
10 0 0 10 0 0 cm BT
/R10 9.9626 Tf
1 0 0 1 472.309 303.562 Tm
(34)Tj
ET
Q
0 g
q
10 0 0 10 0 0 cm BT
/R10 9.9626 Tf
1 0 0 1 482.271 303.562 Tm
[(])-293.995(me)0.990031(chanism)-294.005(of)]TJ
-173.409 -11.9547 Td
[(the)-257.989(decoder)39.9909(,)-259.993(each)-258.004(query)-257.989(\(from)-257.986(the)-257.986(anchor\))-259.016(i)0.98513(s)-259.011(e)15.0122(xpressed)-257.981(as)-257.996(a)]TJ
11.9563 TL
T*[(weighted)-232.015(sum)-232.998(of)-231.998(the)-231.993(v)24.9811(alues)-233.01(\(from)-231.991(the)-231.991(reference\).)-304.018(The)-233.015(resul-)]TJ
11.9551 TL
T*[(tant)-247.983(feature)-248.985(v)14.9828(ectors)-247.99(of)-249.015(all)-247.988(points)-248.983(\()]TJ
/R16 9.9626 Tf
132.011 0 Td
(i.e)Tj
/R10 9.9626 Tf
9.68281 0 Td
[(.,)-249.012(queries\))-247.988(of)-249.017(the)-247.988(anchor)]TJ
-141.694 -11.9547 Td
[(yield)-244.018(a)]TJ
/R16 9.9626 Tf
29.2148 0 Td
[(positive)-244.008(ba)10.0081(g)]TJ
/R10 9.9626 Tf
50.7051 0 Td
[(in)-243.986(MIL)-243.989(for)-245.018(the)-243.989(common)-244.003(cate)15.0122(gory)-243.989(of)-245.018(the)]TJ
-79.9199 -11.9551 Td
[(tw)10.0081(o)-221.017(clouds.)-300.982(MIL)-221.01(via)-221.995(this)-220.997(positi)25.0154(v)14.9828(e)-220.988(bag)-221.997(encourages)-221.017(the)-221.992(model)]TJ
T*[(to)-249.985(attend)-249.983(to)-249.985(the)-249.99(fore)15.0122(ground)-250.017(points)-250.002(of)-249.997(the)-249.988(anchor)55.0153(.)]TJ
11.9551 -15.6398 Td
[(In)-382.992(contrast,)-415.987(we)-382.997(consider)-384(another)-382.985(reference)-382.98(point)-383.005(cloud)]TJ
-11.9551 -11.9551 Td
[(whose)-419.994(cate)15.0122(gory)-419.996(is)-419.016(dif)24.986(ferent)-420.013(from)-419.999(the)-419.999(anchor)55.0153(.)-818.996(This)-419.984(time,)]TJ
T*[(the)-350.993(feature)-351.988(v)14.9828(ectors)-350.993(of)-352.017(all)-350.99(queries)-352.01(of)-351(the)-352.01(anchor)-350.995(generate)-351.985(a)]TJ
/R16 9.9626 Tf
11.9559 TL
T*[(ne)39.9933(gative)-318.011(ba)10.0081(g)]TJ
/R10 9.9626 Tf
54.527 0 Td
[(because)-317.989(the)-316.996(feature)-317.991(v)14.9828(ector)-317.006(of)-317.986(each)-317.016(query)-318.016(is)-318.011(a)]TJ
-54.527 -11.9551 Td
[(weighted)-343.997(sum)-343.999(of)-344.019(v)24.9811(alues)-344.011(of)-344.021(this)-343.997(reference)-344.987(where)-343.987(the)-344.011(tar)17.997(get)]TJ
11.9551 TL
T*[(cate)15.0122(gory)-290.018(is)-290.018(not)-290.013(present.)-431.011(The)-289.986(model)-290.006(via)-289.981(this)-290.006(ne)15.0171(g)4.98446(ati)24.986(v)14.9828(e)-291.015(bag)-289.986(is)]TJ
T*[(enforced)-345.006(to)-346.008(suppress)-345.016(irrele)25.0154(v)24.9811(ant)-346.011(points.)-596.018(W)39.9933(ith)-344.982(the)-346.011(proposed)]TJ
T*[(MIL-transformer)40.0154(,)-297.987(each)-287.999(pair)-287.986(of)-287.991(point)-288.981(clouds,)-296.985(forming)-288.981(either)]TJ
T*[(a)-383(positi)25.013(v)14.9828(e)-382.02(or)-382.992(a)-383(ne)15.0171(g)4.98446(ati)24.986(v)14.9828(e)-383(bag,)-416(produces)-382(e)15.0122(xtra)-382.99(signal)-383.01(to)-382.985(su-)]TJ
11.9563 TL
T*[(pervise)-239.987(model)-240.994(training.)-307.015(In)-240(addition,)-241.999(long-range)-241.004(dependenc)14.9877(y)]TJ
11.9551 TL
T*[(can)-250(be)-249.997(tak)10.0057(en)-249.997(into)-250.012(account)-249.988(via)-249.988(the)-249.988(transformer)54.9859(.)]TJ
11.9551 -15.6387 Td
[(Max)-304.015(or)-305.005(a)19.9918(v)14.9828(erage)-304(pooling)-305.008(is)-304.013(widely)-304.003(used)-304.998(to)-304.018(aggre)15.0171(g)4.98446(ate)-305.008(in-)]TJ
-25.6961 -29.8883 Td
(1)Tj
ET
Q
Q
Q
Q
q
1 0 0 1 0 0 cm
/uijd4MtnezeaR4tYoaKB6Q Do
Q
Q
q
1 0 0 1 0 0 cm
/TXbdxF74oYvgrH4SH-tNFQ Do
Q
endstream
endobj
12 0 obj
<< /R7 29 0 R >>
endobj
13 0 obj
<< /R10 30 0 R /R12 31 0 R /R14 32 0 R /R16 33 0 R /R18 34 0 R /R8 35 0 R >>
endobj
14 0 obj
<< /BBox [ 0 0 612 792 ] /Matrix [ 1 0 0 1 0 0 ] /Resources << /Font 36 0 R /ProcSet [ /PDF /Text /ImageB /ImageC /ImageI ] >> /Subtype /Form /Type /XObject /Length 114 /Filter /FlateDecode >>
stream
x-0D|CMk21T;p>=%=_;L!bcƞrk5UfB,NPH cP۹{b6//3
endstream
endobj
15 0 obj
<< /BBox [ 0 0 612 792 ] /Group << /CS /DeviceRGB /I true /S /Transparency /Type /Group >> /Resources << /ExtGState << /a0 << /CA 1 /ca 1 >> /a1 << /CA 0.5 /ca 0.5 >> >> /XObject << /x6 37 0 R >> >> /Subtype /Form /Type /XObject /Length 382 /Filter /FlateDecode >>
stream
xe[n!Y7P|=Q'y(Ji;#$v0U-V\oZWɽ:5db,6IR6@6Pc"3K+GQ؎3T<9j`UT@
v5ڳtHN;9+7[^S؇:[E"8+sW\qEeКu6Bhsj;)y;Tp0nؚgʦFǃS-="z7ۺ9*&Q~ZPuí}C9嬂8l?Jܪmx%e8w](Rlq鞋t×e:=p=y0`XeJCWxau~uof{`r
endstream
endobj
16 0 obj
<< /Contents 38 0 R /MediaBox [ 0 0 612 792 ] /Parent 7 0 R /Resources << /ExtGState 39 0 R /Font 40 0 R /ProcSet [ /PDF /ImageC /Text ] /XObject 41 0 R >> /Rotate 0 /Type /Page >>
endobj
17 0 obj
<< /Contents 42 0 R /MediaBox [ 0 0 612 792 ] /Parent 7 0 R /Resources << /ExtGState 39 0 R /Font 43 0 R /ProcSet [ /PDF /ImageC /Text ] /XObject 44 0 R >> /Rotate 0 /Type /Page >>
endobj
18 0 obj
<< /Contents 45 0 R /MediaBox [ 0 0 612 792 ] /Parent 7 0 R /Resources << /ExtGState 12 0 R /Font 46 0 R /ProcSet [ /PDF /Text ] /XObject << /070XxmNY8OMo_qrlr0KQfw 47 0 R >> >> /Rotate 0 /Type /Page >>
endobj
19 0 obj
<< /Contents 48 0 R /MediaBox [ 0 0 612 792 ] /Parent 7 0 R /Resources << /ExtGState 12 0 R /Font 49 0 R /ProcSet [ /PDF /Text ] /XObject << /LmabTUgbw-xdlSTEM9BSjw 50 0 R >> >> /Rotate 0 /Type /Page >>
endobj
20 0 obj
<< /Contents 51 0 R /MediaBox [ 0 0 612 792 ] /Parent 7 0 R /Resources << /ExtGState 12 0 R /Font 52 0 R /ProcSet [ /PDF /Text ] /XObject << /60-GLI2Gs-1-2REegz_f5g 53 0 R >> >> /Rotate 0 /Type /Page >>
endobj
21 0 obj
<< /Contents 54 0 R /Group 55 0 R /MediaBox [ 0 0 612 792 ] /Parent 7 0 R /Resources << /ExtGState 56 0 R /Font 57 0 R /ProcSet [ /PDF /Text ] /XObject 58 0 R >> /Rotate 0 /Type /Page >>
endobj
22 0 obj
<< /Contents 59 0 R /Group 55 0 R /MediaBox [ 0 0 612 792 ] /Parent 7 0 R /Resources << /ExtGState 56 0 R /Font 60 0 R /ProcSet [ /PDF /Text ] /XObject 61 0 R >> /Rotate 0 /Type /Page >>
endobj
23 0 obj
<< /Contents 62 0 R /MediaBox [ 0 0 612 792 ] /Parent 7 0 R /Resources << /ExtGState 12 0 R /Font 63 0 R /ProcSet [ /PDF /Text ] /XObject << /9HdbcRehQ-9I5A5_ewNkZg 64 0 R >> >> /Rotate 0 /Type /Page >>
endobj
24 0 obj
<< /Contents 65 0 R /MediaBox [ 0 0 612 792 ] /Parent 7 0 R /Resources << /ExtGState 12 0 R /Font 66 0 R /ProcSet [ /PDF /Text ] /XObject << /cxtfv93hA5ubJN9EeeTang 67 0 R >> >> /Rotate 0 /Type /Page >>
endobj
25 0 obj
<< /Filter /FlateDecode /Length 1503 >>
stream
xWsD3aVRsvfvI2!I+[,]$9((((((((((((((((((((((((((((((((x;9 3s{ntsL3z??Ͽq0o;;;hZ$@tyM].%DY~U8b[kWI'ֳBU^[K4}]egM0|8UY#=+yQuAD3Zq6K4 A`p&'*`W{CPJӳZMD4p{ovkF(9G
vnԴX
xr{4
FcC(Q.'@ 3f6dI3!ܢ~ZjC3R_CUq!uG{4UpwR?uw&YLM<6[Lkڬ
::h=IƧ߭7lHcrW0w{ekpn(@ܖ80q0-ݒbUoʢn}`kncƱ!kblh5F,Tؖ*bxuGVkh-5UYgMFej-Z7\Jm#wMǍn/(,YoPB{ۘ_:2*eLFJ)sU:_%zVgHE0#r)#'Nlq~W!_xTzYXQ%Ҩ }KnO\H-RozyV"DPh; ZwbP)6mبw5CBM):zz8`Euэ0+]E"X!C6Q̀4*[yaS6e`HR_}RG5wd_,*^c7 \Y_+g07-S4y8;q StQB]H?1DVHf!a>ۋG@zM} HSV6og}TqCsb:mHa|.-bJ0=u=kjA
=zWtz&j"ݐpgOUC?}s!0+ap)qom5YnpҙdB4ɢJ`_vP,dYB#Fht/=.!ҔgwYMqגCbV2j0{G,3oZsS=NF}Y>+dlrpAu6εVCa&dLz>BCY0CV!ti
`١"IPDq7e c.]D1^8sǞ8)NW]/43ܾ\SW:A 2 %2JRe/uSŕdT0\ƃXA#7+C"vҐ#-M
endstream
endobj
26 0 obj
<< /Filter /FlateDecode /Length 14126 >>
stream
x} xEֶ(((nW1
"P]f& dAGD`T
qEYEPq\Ad\q}TsԕoyνowW9u+ݷqxi3g;v-gʇ]lzo~6mHčTosiS&F#cC*=k*Eیls^$(::VZfd0nQ?O͂䧾uk*[ь[*+[+-?X=vltXq"Zjj*S۳gdq}T9"De^sms^}Y,U'CjcUczG8s"49|\\ɏV=$1jl4/7$;u@=Cr^ŵѱձ)>Uc˫}+'U
ʪ3-3,/ػ+u)uU%v.L/sX4V(Z5v\*QQ
4 5N^
[7!mg^*zXt]0pJ3Y3^xk"WzjoVu{Wʛ}-w1>.K~wl+VQ/3_~dцO֝{~>^5w/n|Zغ_?uȶk~x{wm³3}nާ5O9{7/mڦ{r?~ᆵ[;7#[-Vߺ?}~ny㒋ﯾ{{ZuY8WE1O?%w(ywv?^-ci|{Gn%1 F9%c&T\^d!j+-F(I,ŕe,}_SK^5p{DS"ɸH7`CJjzҳ&Fk'[7LTm!,gFcB2^HE|s>kzaW]ScD= ,hYEq(gC/_#SE'V%0t&"$>TQ=6)wF$5ȸWEJ~A%[}|h=`nIRxLH
0)C^R*yOќBq@C(H"[A iHHJU$s$AH)3 t$ !FҐu/sd sg, ,"t
$@JaDw\Ϗx@2XY.6x".<*a%R%0za
!9B`q!8^|d@9ņJ@P(q<p7});@Rb#(| AcKCp([&@RHfKGme09RȨCbPhaE{JH,AMBs^{Kʊ+`c'8dZdx]c[ۣ#,h/:S`L]MvHmX1X0?uK$&2ˣ
;53#<
]ܔj#$ݨiƂ7l7nEBl7?v̴%obA/vdumLoMIZ5)
qeS
=--/+qEq]ճqTpjb 3FQ$nq2t9wRLKͿ:v`ǎYNU䍣~cI*MLz8~~&w4yk(o塓oPstwqǷ={,qȱyvszQߴ_E-Z]?d'ޜww\vDǷT>v|o5z+;|ޅLqmGx絊~qMF^tvvNYSOz6UnSa9rЂ7kމ0yW͛v+4ݲ#=r?>l9vٲo?;eL[Nxbv?y=ljOvN9#M:dW}+pAuθk{_r+]˻y?hוwg2EZvFKFuX}m)4~ߟ:tLiZ4 vj@='yg]yE),h=QQ@0##Q2U斕b2~rnf)tZ9^,S&x
~4/R[+^XWPcƭޔ ۄȩlyȔʒNnaſV#CF'`RC>sDr}(
k#lv¼g6TZø)uAjiUwM'6OΜFtI6}qn>>a[w]2G,pyGU]WMb{5;uqgɼ3Z<{_ZMmF3UC?|Uv[a!K#;gC\2ܶeL_ܪu]|hрoX{5gO8v'*luwjEO>{-l:gl6ť/ryMyاM'F)m7p4d[$gLI\#Ljt R&]"}
L]aTʤVk/hwܽ1OP sC&I?g9Y gZ J4Y 2Dla Kp$cpbO q\4gNβ&)QUS^Op~d
ohW3=]SH)d2PpΨp<Ω5t,Zr{Kpǜ4
¯RdE{Y}Ts pҁ1cu|Yq$
aVW(=XB[z9665ظykF H/H,A=k;90AԵ YH/,AII@w %wx(aԇ"Pb6iP`(NKV7`t%6Af %|F%2N
]5 p]/ ֳCA}/&NU,6,^ #2#eTpU{msV
PhNrhC
rr6G(t4C@}'z?Ƞq3AJM~
hPuĽT.gIHj`PIHUNsF_XփdA~3:`&<+sg37,HPp]j2HP4Ӑs8e@5 1Q\ _pz0-)ARA|s
uEg T`Bu`ؑ#jeSo\G,H.@#I.D#ja
nHD,Kz}{@)HCە!k[!ņ2 eAa}C'!qb{c)H
U@b;u$`ҐmAӦ 9x-CkHRE$!7B5XN[&E
\\,$Ґ<xJ4pyfK)huƺ0d<\IDwoh5:p
ɁMdc60(70\Rh$Ġ00X=)D9 ̢J "k kDCh`wQK9GQԀ7K: 49ARƞiHUX$Ƞ0l)H
.r
d=@NsaQcעudPjR@L<=!5 ,I*]7l'!61]
"0Ә_7)fZ+lD`#z
NA{*Ġ_M$x$rA4&~Rz hx7`1!
g%tc[Fl:dGf{6s<!NeBX*E-oL}>m%m?t8٭:z%ǾNcV}l./Bf_vƿҺɿ|>o|pFuozӹ^s?ѵ^5~a+]rϫ.ѳ~ܯsw]Rq`맞{^mNiq̎qǎ n_vRuVYQ3W>Xt)?r^~}FT͚<.t+F}<ˊ͖t`k-:bc+7;fbnS|䘣W-)L1j\<Sp =LC&kOs*