[DL輪読会]Domain Adaptive Faster R-CNN for Object Detection in the Wild

>100 Views

June 20, 18

スライド概要

2018/06/15
Deep Learning JP:
http://deeplearning.jp/seminar-2/

シェア

埋め込む »CMSなどでJSが使えない場合

関連スライド

各ページのテキスト
1.

“Domain Adaptive Faster R-CNN for Object Detection in theWild (CVPR 2018) ଞ” Kento Doi, Iwasaki Lab 1

2.

ൃද࿦จ • Domain Adaptive Faster R-CNN for Object Detection in theWild. • Y. Tsai et al. • CVPR 2018 • Learning to Adapt Structured Output Space for Semantic Segmentation. • Y. Chen et al. • CVPR 2018 • ͲͪΒ΋CVPR accept, υϝΠϯదԠʹؔ͢Δ‫ڀݚ‬ ※ ग़య‫ݩ‬ͷ‫͍ͳ͕ࡌه‬ը૾͸ൃද࿦จ͔ΒͷҾ༻ 2

3.

બఆཧ༝ • ࣮ར༻ʹ͓͍ͯυϝΠϯదԠ͸େ੾ ྫ) • γϛϡϨʔγϣϯը૾Ͱࣝผ‫ث‬Λֶशˠ‫ੈ࣮ݱ‬քʹద༻ • ͋Δࠃͷಓ࿏γʔϯը૾Ͱֶशͨࣝ͠ผ‫ث‬ˠଞͷࠃʹద༻ • ಓ࿏γʔϯը૾ͷఱ‫ؾ‬΍࣌ؒͷมԽʹରԠ • ࣗ෼ͷઐ໳ͷ஍‫؍ٿ‬ଌ෼໺ʹ͓͍ͯ΋υϝΠϯదԠ͕ॏཁͩͱߟ͑ͯ ͍Δ͔Β(‫؍‬ଌ஍Ҭɺηϯαʔ͕ҟͳΔ৔߹) • CVPRͷ࿦จΛ࠷ۙಡΜͰ͍Δ 3

4.

υϝΠϯదԠ(Domain Adaptation)ͱ͸ • 2ͭͷσʔληοτͷυϝΠϯόΠΞεʹରԠͰ͖ΔΑ͏ʹࣝผ‫ث‬Λֶ श͢Δ͜ͱ υϝΠϯదԠʹ͍ͭͯ͸ԼͷࢿྉʹΑ͘·ͱ·͍ͬͯΔɻ [DLྠಡձ] “Asymmetric Tri-training for Unsupervised Domain Adaptation (ICML2017)” and Neural Domain Adaptation (https://www.slideshare.net/YuusukeIwasawa/dl-asymmetric-tritrainingfor-unsupervised-domain-adaptation-icml2017-and-neural-domain-adaptation) 4

5.

υϝΠϯదԠ(Domain Adaptation)ͱ͸ • 2ͭͷσʔληοτͷυϝΠϯόΠΞεʹରԠͰ͖ΔΑ͏ʹࣝผ‫ث‬Λֶ श͢Δ͜ͱ ͭ·Γɺยํͷ υϝΠϯͷΈਖ਼ ղΛ࣋ͭͱ͍͏ ໰୊ઃఆ υϝΠϯదԠʹ͍ͭͯ͸ԼͷࢿྉʹΑ͘·ͱ·͍ͬͯΔɻ [DLྠಡձ] “Asymmetric Tri-training for Unsupervised Domain Adaptation (ICML2017)” and Neural Domain Adaptation (https://www.slideshare.net/YuusukeIwasawa/dl-asymmetric-tritrainingfor-unsupervised-domain-adaptation-icml2017-and-neural-domain-adaptation) 5

6.

υϝΠϯదԠ(Domain Adaptation)ͱ͸ • 2ͭͷσʔληοτͷυϝΠϯόΠΞεʹରԠͰ͖ΔΑ͏ʹࣝผ‫ث‬Λֶ श͢Δ͜ͱ Deep Visual Domain Adaptation: A Survey [M. Wang 2018] 6

7.

υϝΠϯదԠ(Domain Adaptation)ͷϙΠϯτ • ιʔευϝΠϯ͸ਖ਼ղσʔλ͋ΓɺλʔήοτυϝΠϯ͸ͳ͠ • ಛ௃্ۭؒͰιʔευϝΠϯͱλʔήοτυϝΠϯͷ෼෍͕Ұக͢Ε ͹ྑ͍ ➡ DNNΛ༻͍֤ͯυϝΠϯͷ෼෍Λ͚ۙͮΔख๏ ➡ ఢରతֶशΛߦ͏ख๏͕༗ޮ • "EWFSTBSJBM%JTDSJNJOBUJWF%PNBJO"EBQUBUJPO<&5[FOHFUBM$713> • "%*355"1130"$)506/461&37*4&%%0."*/"%"15"5*0/<3VJ4IVFUBM*$-3> • 0QFO4FU%PNBJO"EBQUBUJPOCZ#BDLQSPQBHBUJPO<,4BJUPFUBMBS9JW> 7

8.

ઌߦ‫ڀݚ‬ "EWFSTBSJBM%JTDSJNJOBUJWF%PNBJO"EBQUBUJPO<&5[FOHFUBM$713> • ఢରతֶशΛ༻͍ͨυϝΠϯదԠ Source domainͰ෼ྨ‫ث‬Λֶश Source domainͷಛ௃நग़‫ث‬͸‫ݻ‬ఆ ఢରతֶशʹΑΓɺtarget domainͷ ಛ௃ϕΫτϧͷ෼෍Λsource domainʹ͚ۙͮΔ 8

9.

ઌߦ‫ڀݚ‬ "EWFSTBSJBM%JTDSJNJOBUJWF%PNBJO"EBQUBUJPO<&5[FOHFUBM$713> • ఢରతֶशΛ༻͍ͨυϝΠϯదԠ Source domainͰ෼ྨ‫ث‬Λֶश Source domainͷಛ௃நग़‫ث‬͸‫ݻ‬ఆ ఢରతֶशʹΑΓɺtarget domainͷ ಛ௃ϕΫτϧͷ෼෍Λsource domainʹ͚ۙͮΔ 9

10.

ઌߦ‫ڀݚ‬ "EWFSTBSJBM%JTDSJNJOBUJWF%PNBJO"EBQUBUJPO<&5[FOHFUBM$713> • ఢରతֶशΛ༻͍ͨυϝΠϯదԠ Source domainͰ෼ྨ‫ث‬Λֶश Source domainͷಛ௃நग़‫ث‬͸‫ݻ‬ఆ ఢରతֶशʹΑΓɺtarget domainͷ ಛ௃ϕΫτϧͷ෼෍Λsource domainʹ͚ۙͮΔ 10

11.

ઌߦ‫ڀݚ‬ "EWFSTBSJBM%JTDSJNJOBUJWF%PNBJO"EBQUBUJPO<&5[FOHFUBM$713> • චऀʹΑΔఢରతֶशΛ༻͍ͨυϝΠϯదԠͷҰൠԽ 11

12.

ઌߦ‫ڀݚ‬ "EWFSTBSJBM%JTDSJNJOBUJWF%PNBJO"EBQUBUJPO<&5[FOHFUBM$713> • චऀʹΑΔఢରతֶशΛ༻͍ͨυϝΠϯదԠͷҰൠԽ ఢରతֶशΛ ͍ͯ͠Δ͔Ͳ ͏͔ ੜ੒Ϟσϧ͔ ࣝผϞσϧ͔ ॏΈΛ‫ڞ‬༗͢ Δ͔൱͔ 12

13.

Domain Adaptive Faster R-CNN for Object Detection in the Wild 13

14.

Domain Adaptive Faster R-CNN for Object Detection in the Wild • CVPR2018ʹaccept – චऀ͸νϡʔϦοώ޻ՊେͷPh.D student • ఢରతֶशʹΑΔυϝΠϯదԠΛ෺ମ‫ݕ‬ग़λεΫͰߦͬͨ – end-to-endͰߦͬͨྫ͸ॳΊͯ – ಓ࿏γʔϯͷυϝΠϯదԠΛߦ͍‫ূݕ‬ • Caffe࣮૷͋Γ – https://github.com/yuhuayc/da-faster-rcnn 14

15.

࿦จ֓ཁ • ෺ମ‫ݕ‬ग़ಛ༗ͷυϝΠϯγϑτΛఆٛ – Image-level shift -> ը૾ελΠϧɺ໌Δ͞ͳͲ – Instance-level shift -> ෺ମͷαΠζɺ֎‫؍‬ͷҧ͍ • 2ͭͷυϝΠϯదԠͷίϯϙʔωϯτΛFaster-RCNNʹ௥Ճ – ελΠϧͷҰ؏ੑΛߟྀͨ͠ਖ਼ଇԽ΋ಋೖ • ϕϯνϚʔΫ(Faster R-CNN)ͱͷൺֱͰ༗ޮੑΛ֬ೝ 15

16.

ಓ࿏γʔϯͷυϝΠϯγϑτ KITTI Cityscape Foggy Cityscapes SIM10K 16

17.

Review: Faster R-CNN • 2ஈ֊ͷωοτϫʔΫͰ෺ମ‫ݕ‬ग़Λߦ͏ϑϨʔϜϫʔΫ – Region Proposal Network (RPN)ʹΑΓީิྖҬΛநग़ – நग़ͯ͠ީิྖҬͷ෺ମͷbounding boxͱΧςΰϦΛਪఆ 17

18.
[beta]
υϝΠϯదԠͷఆࣜԽ
• ෼෍ͷۙ͞͸H-μΠόʔδΣϯεʹΑΓද͞ΕΔ

dH (S, T ) = 2(1

min(errS (h(x)) + errT (h(x)))

h2H

<latexit sha1_base64="zk0n5gPcSAlpF46Nb6l5Ti4epq0=">AAAC1nichVHPSxtBFH5ZrdpYNa2XgpehwbJBDW+D0CIIYi8e1SQaMLLsrqMZsr/cnQR1iTcRPHnrwVMLpYh/hpf+A6X4J0iPFrwo+LLZ1qqob9mZb773vjffzJi+LUKJeJZSurpf9PT2vUz3vxoYHMq8frMUeo3A4mXLs72gYhoht4XLy1JIm1f8gBuOafNls/6pnV9u8iAUnluS2z5fdYwNV6wLy5BE6ZnNNb3qGLJmGXY011L/4eI4K7VybJoVVI1NsKojXD2qsapw2X/1LZUHgR7dqoipqVu5HBtjdzOlv5mcnsliHuNgD4GWgCwkMe9lvkMV1sADCxrgAAcXJGEbDAjpWwENEHziViEiLiAk4jyHFqRJ26AqThUGsXUaN2i1krAurds9w1ht0S42/QEpGYziTzzGC/yBJ3iOV4/2iuIebS/bNJsdLff1oYO3xctnVQ7NEmq3qic9S1iHj7FXQd79mGmfwuromzufL4pTi6PRe/yKv8n/FzzDUzqB2/xjfVvgi0eQpgfQ7l/3Q7BUyGuY1xYmszOzyVP0wQi8A5Xu+wPMwBzMQ5n2/QXXqZ5Ur1JRdpU9Zb9TqqQSzTDcCeXwBilcsn8=</latexit>

–

h(·) ͸υϝΠϯ෼ྨ‫ث‬,H ͸ h(·)ͷ฼ू߹, x ͸ಛ௃ϕΫτϧ
<latexit sha1_base64="T5eNG1t/rh/1LJqdf1tI18J36qA=">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</latexit>
<latexit

<latexit sha1_base64="oZG8X0I1YX9a5Kue3TapGpoTijo=">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</latexit>
<latexit

<latexit sha1_base64="rpG2rJy3HIUJn9eXjgKVUCFJ2TY=">AAACa3ichVHLSsNAFD2Nr1ofrXYj6kIsFd2UGxEUV0U3Ln3VFqpIEkcNpklIpoVa/AGXblzUjYKI+Blu/AEX/QQRVxXcuPA2DYiKesNkzpy5586ZO7prmb4kakSUjs6u7p5ob6yvf2Awnhga3vKdsmeInOFYjlfQNV9Ypi1y0pSWKLie0Eq6JfL60XJrP18Rnm869qasumKnpB3Y5r5paJKpwuH0trHnyJndRIoyFMTET6CGIIUwVp3EDbaxBwcGyihBwIZkbEGDz18RKgguczuoMecxMoN9gRPEWFvmLMEZGrNH/D/gVTFkbV63avqB2uBTLB4eKyeQpke6pSY90B090fuvtWpBjZaXKs96Wyvc3fjpyMbbv6oSzxKHn6o/PUvsYyHwarJ3N2BatzDa+srxeXNjcT1dm6Iremb/l9Sge76BXXk1rtfEeh0xfgD1e7t/gq3ZjEoZdW0ulV0KnyKKMUximvs9jyxWsIpc0Ocz1HEReVGSyqgy3k5VIqEmiS+hpD8AlA6MhQ==</latexit>

–

<latexit sha1_base64="rpG2rJy3HIUJn9eXjgKVUCFJ2TY=">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</latexit>

err ͸υϝΠϯ෼ྨͷଛࣦΛද͢
<latexit sha1_base64="yKs4214AnmzLb+usGMMbMK6uBI4=">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</latexit>

• H-μΠόʔδΣϯεΛ࠷খԽ͢ΔΑ͏ʹಛ௃நग़‫ث‬Λֶश͢Δ
min dH (S, T ) , max min{errS (h(x)) + errT (h(x))}
<latexit sha1_base64="x5IFo/EZdyrzyu+zAJRAM58Tfh8=">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</latexit>

f

f

f 2H

18

19.
[beta]
υϝΠϯదԠͷఆࣜԽ
• σʔλ෼෍Λ P (C, B, I) Ͱఆٛ
<latexit sha1_base64="FsCLDgx8yVZr53UyYh5dUSrt2q0=">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</latexit>

– C: ෺ମͷΧςΰϦ C 2 {1, ..., K}
<latexit sha1_base64="+Id8Gd4NNG73rYLiuN8eedJF5xU=">AAACeHichVHLSsNAFD2N7/qquhHcBIsvkHAjguKq6EZw46sqWClJHOtgmoQkLdTSH/AHXLgQBVHxM9z4Ay78BHGpIIgLb9OAqKh3mJkzZ+65c2bG9GwZhEQPCaWpuaW1rb0j2dnV3dOb6uvfCNySb4ms5dquv2UagbClI7KhDG2x5fnCKJq22DQPFur7m2XhB9J11sOKJ3aKRsGRe9IyQqbyqYEFNScdNVfVJ1VN0ybVpVwtn0qTRlGoP4EegzTiWHZTl8hhFy4slFCEgIOQsQ0DAbdt6CB4zO2gypzPSEb7AjUkWVviLMEZBrMHPBZ4tR2zDq/rNYNIbfEpNneflSpG6J6u6Znu6IYe6f3XWtWoRt1LhWezoRVevvdocO31X1WR5xD7n6o/PYfYw2zkVbJ3L2Lqt7Aa+vLh8fPa3OpIdZTO6Yn9n9ED3fINnPKLdbEiVk+Q5A/Qvz/3T7Axpemk6SvT6cx8/BXtGMIwxvm9Z5DBIpaR5XMrOMUVrhNviqqMKRONVCURawbwJZSpD2Xkj2k=</latexit>

– B: ෺ମͷҐஔ

– I : ը૾ͷಛ௃ද‫ݱ‬
– υϝΠϯγϑτ͕͋Δ࣌ PS (C, B, I) 6= PT (C, B, I)
<latexit sha1_base64="/6xH3SmOopi1FJ7IjfIugqXLbbw=">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</latexit>

• ͜ͷΑ͏ͳఆࣜԽͰυϝΠϯγϑτΛ֬཰࿦తʹղੳ͢Δ

19

20.
[beta]
υϝΠϯదԠͷఆࣜԽ
• σʔλ෼෍Λ P (C, B, I) Ͱఆٛ

B

C

B

<latexit sha1_base64="FsCLDgx8yVZr53UyYh5dUSrt2q0=">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</latexit>

– C: ෺ମͷΧςΰϦ C 2 {1, ..., K}
<latexit sha1_base64="+Id8Gd4NNG73rYLiuN8eedJF5xU=">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</latexit>

– B: ෺ମͷҐஔ

I

– I : ը૾ͷಛ௃ද‫ݱ‬

– υϝΠϯγϑτ͕͋Δ࣌ PS (C, B, I) 6= PT (C, B, I)
<latexit sha1_base64="/6xH3SmOopi1FJ7IjfIugqXLbbw=">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</latexit>

• ͜ͷΑ͏ͳఆࣜԽͰυϝΠϯγϑτΛ֬཰࿦తʹղੳ͢Δ

20

21.

Image-level Adaptation • ϕΠζͷఆཧΑΓP (C, B, I) = P (C, B|I)P (I) <latexit sha1_base64="W1r/mjxiAw9mGoeP4epHF6dlOY8=">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</latexit> • PS (C, B|I) = PT (C, B|I) ͱԾఆ͢Δ(covariate shift assumption) <latexit sha1_base64="is2ODp6hlGslvVImQfxSMBXDfnk=">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</latexit> – ͭ·Γɺ‫ݕ‬ग़ωοτϫʔΫ͸υϝΠϯෆมͱԾఆ • PS (I) = PT (I) ʹͳΔΑ͏ʹಛ௃நग़‫ث‬Λֶश͢Ε͹ྑ͍ <latexit sha1_base64="n/c7hUwyWyPioxQWmgpV4VQI2g4=">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</latexit> • ࣮૷Ͱ͸ҎԼͷΑ͏ʹఢରతֶशΛ͢Δ 21

22.

Instance-level Adaptation • ಉ༷ʹ P (C, B, I) = P (C|B, I)P (B, I) <latexit sha1_base64="rhXLBJFDhM3o9jnkw0+VrfTTFL8=">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</latexit> • PS (C|B, I) = PT (C|B, I) ͱԾఆ͢Δ(covariate shift assumption) <latexit sha1_base64="PiHpPQLI3yLcokz8hFFBVWrf70M=">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</latexit> • PS (B, I) = PT (B, I)ʹͳΔΑ͏ʹಛ௃நग़‫ث‬Λֶश͢Ε͹ྑ͍ <latexit sha1_base64="ZCsQhn8DE0ASJmIi19v7E3DzyPU=">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</latexit> • ࣮૷Ͱ͸ҎԼͷΑ͏ͳఢରతֶशΛ͢Δ 22

23.

Consistency Regularization • ͔͠͠ɺP (B, I) ΛదԠͤ͞Δͷ͸೉͍͠ <latexit sha1_base64="LKiQUU+c+HER3epJ5ng4N9KtztQ=">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</latexit> • P (B, I) = P (B|I)P (I) – P (I) Ͱ‫׬‬શʹपลԽ͢Δͷ͸ࠔ೉ <latexit sha1_base64="r16VObvu1oiJxbeRTbyCQCNf+l0=">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</latexit> <latexit sha1_base64="921V3wPsUP2lInCPE4cyUpFc8KQ=">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</latexit> – Bounding boxͷΞϊςʔγϣϯ͕ιʔευϝΠϯ͔͠ͳ͍ ➡ ϕΠζͷఆཧΑΓP (D|B, I)P (B|I) = P (B|D, I)P (D|I) <latexit sha1_base64="v1oafSPo0Ikizj3T1ofMhixFX1o=">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</latexit> • ্ࣜΑΓ P (D|B, I) = P (D|I) ͱͳΔΑ͏ʹֶश͢Ε͹ <latexit sha1_base64="tPqT7Y40SMK67wx1ZReWwrSZRa0=">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</latexit> P (B|I) = P (B|D, I) ͕ಘΒΕΔ <latexit sha1_base64="jfSDYjsv00RGosd8NSdF4ILmfWU=">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</latexit> ͭ·ΓɺP (D|B, I) = P (D|I) ͷਖ਼ଇԽΛֶशͰೖΕͯ͋͛Ε͹ྑ͍ 23 <latexit sha1_base64="tPqT7Y40SMK67wx1ZReWwrSZRa0=">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</latexit>

24.

Consistency Regularization • P (D|B, I) = P (D|I) ͷਖ਼ଇԽ <latexit sha1_base64="tPqT7Y40SMK67wx1ZReWwrSZRa0=">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</latexit> • ͜Ε͸ɺը૾શମͷυϝΠϯͱΦϒδΣΫτʹண໨ͨ͠ͱ͖ͷυϝΠ ϯʹҰ؏ੑΛ࣋ͨͤΔͱ͍͏͜ͱʹ౳͍͠ • ࣮૷Ͱ͸࣍ͷΑ͏ʹͳΔ Image-level domain Instance-level domain 24

25.

ωοτϫʔΫ֓‫؍‬ 25

26.

ωοτϫʔΫ֓‫؍‬ ީิྖҬͷ நग़ ީิྖҬͷ෼ ྨͱlocalization Ldet = Lrpn + Lroi <latexit sha1_base64="j34hNYJRGY51Jqs2HGUOeUkjvOM=">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</latexit> 26

27.

ωοτϫʔΫ֓‫؍‬ Limg <latexit sha1_base64="vlrpTlsCJkj3/dNPowGLUUB3zFg=">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</latexit> Ldet = Lrpn + Lroi <latexit sha1_base64="j34hNYJRGY51Jqs2HGUOeUkjvOM=">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</latexit> ‫఻ٯࠩޡ‬೻࣌ ʹූ߸Λ൓స Discriminator (Domain classifier) ͤ͞Δlayer 27

28.

ωοτϫʔΫ֓‫؍‬ Lins <latexit sha1_base64="DL2N+p6sQn6vhxOhqD9lyo0pt6A=">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</latexit> ‫఻ٯࠩޡ‬೻࣌ ʹූ߸Λ൓స Ldet = Lrpn + Lroi Discriminator (Domain classifier) Limg <latexit sha1_base64="vlrpTlsCJkj3/dNPowGLUUB3zFg=">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</latexit> ͤ͞Δlayer <latexit sha1_base64="j34hNYJRGY51Jqs2HGUOeUkjvOM=">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</latexit> 28

29.

ωοτϫʔΫ֓‫؍‬ Lins <latexit sha1_base64="DL2N+p6sQn6vhxOhqD9lyo0pt6A=">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</latexit> ྆υϝΠϯͷ Ұ؏ੑͷਖ਼ଇԽ Lcst <latexit sha1_base64="cm/dIZ+pJprXu9qq40uE4ZFwb4Q=">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</latexit> Limg <latexit sha1_base64="vlrpTlsCJkj3/dNPowGLUUB3zFg=">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</latexit> Ldet = Lrpn + Lroi <latexit sha1_base64="j34hNYJRGY51Jqs2HGUOeUkjvOM=">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</latexit> 29

30.

ωοτϫʔΫ֓‫؍‬ Lins <latexit sha1_base64="DL2N+p6sQn6vhxOhqD9lyo0pt6A=">AAACdnichVHLLgRBFD3T3uM12EgkIiaD1eS2SIiVsLGw8BokyKS7FSr6le6aETrzA37AQixIEPEZNn7AwieIJQkLC3d6OhEEt1JVp07dc+tUlenbMlREDymtrr6hsam5Jd3a1t7RmenqXgm9UmCJguXZXrBmGqGwpSsKSipbrPmBMBzTFqvm3kx1f7UsglB67rI68MWmY+y4cltahmKqmOnacAy1axl2NFcpRtINK8VMlvIUx8BPoCcgiyTmvcwVNrAFDxZKcCDgQjG2YSDktg4dBJ+5TUTMBYxkvC9QQZq1Jc4SnGEwu8fjDq/WE9bldbVmGKstPsXmHrByADm6p2t6pju6oUd6/7VWFNeoejng2axphV/sPOpdev1X5fCssPup+tOzwjYmYq+SvfsxU72FVdOXD4+flyYXc9EQndMT+z+jB7rlG7jlF+tiQSyeIM0foH9/7p9gZTSvU15fGMtOTSdf0Yw+DGKE33scU5jFPAp87j5OcYmr1JvWr+W04Vqqlko0PfgSGn0AUmmRbA==</latexit> Lcst <latexit sha1_base64="cm/dIZ+pJprXu9qq40uE4ZFwb4Q=">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</latexit> Limg <latexit sha1_base64="vlrpTlsCJkj3/dNPowGLUUB3zFg=">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</latexit> Ldet = Lrpn + Lroi <latexit sha1_base64="j34hNYJRGY51Jqs2HGUOeUkjvOM=">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</latexit> ࠷খԽ͢Δଛࣦ͸࣍ͷΑ͏ʹͳΔ L = Ldet + (Limg + Lins + Lcst ) <latexit sha1_base64="aV4BryQ2cHSnt2Ae0uetWJULrIc=">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</latexit> ࣮‫Ͱݧ‬͸0.1ʹͨ͠ 30

31.

࣮‫ݧ‬ • 3छྨͷυϝΠϯదԠΛ‫ূݕ‬ – ςϨϏήʔϜ(SIM10K)→‫ੈ࣮ݱ‬ք(Cityscapes) – ੖ఱ(Cityscapes)→ѱఱީ(ໄ)(Foggy Cityscapes) – ҟͳΔΧϝϥ΁ͷదԠ(Cityscapes, KITTI) • ֶशʹ࢖͏υϝΠϯదԠͷଛࣦΛม͑ͯ݁ՌΛൺֱ 31

32.

ήʔϜˠ‫ੈ࣮ݱ‬ք • ݁Ռ͸ंΫϥεͷΈͰൺֱ • શͯͷυϝΠϯదԠͷଛࣦΛ࢖ͬͨํ͕ྑ͍݁Ռʹ 32

33.

੖ఱˠѱఱީ(ໄ)ɺҟͳΔΧϝϥؒͷυϝΠϯదԠ 33

34.

֤υϝΠϯదԠͷޮՌ • Image-level Adaptationͷํ͕݁Ռ΁ͷ‫د‬༩͕େ͖͍ • චऀ͸Image-level adptationͷํ͕RPNʹ௚઀ޮ͔͘ΒͰ͸ͳ͍͔ͱ ड़΂͍ͯΔ 34

35.

·ͱΊ • ෺ମ‫ݕ‬ग़ͷυϝΠϯదԠʹ͓͍ͯ2ͭͷυϝΠϯγϑτΛఆٛ – Image-level shift – Instance-level shift • 2छྨͷυϝΠϯదԠͷͨΊͷίϯϙʔωϯτΛFaster R-CNNʹ௥Ճ • ͞ΒʹυϝΠϯͷҰ؏ੑʹ͍ͭͯͷਖ਼ଇԽͷͨΊͷίϯϙʔωϯτ΋ ௥Ճ • 3छྨͷυϝΠϯγϑτʹ͓͍ͯɺख๏ͷ༗ޮੑΛ‫ূݕ‬ 35

36.

Learning to Adapt Structured Output Space for Semantic Segmentation 36

37.

Learning to Adapt Structured Output Space for Semantic Segmentation • CVPR2018 accept – ஶऀ͸NECͷ‫ڀݚ‬һ • Visual Domain Adaptation ChallengeͰ3Ґʹͳͬͨख๏ • PyTorch࣮૷͋Γ – https://github.com/wasidennis/AdaptSegNet 37

38.

࿦จ֓ཁ • ಓ࿏γʔϯͷυϝΠϯదԠ(Semantic Segmentation) • ෼ྨ݁Ռͷۭؒతͳ഑ஔʹ஫໨ • Multi level domain adaptation – ෳ਺ͷಛ௃ϚοϓͷυϝΠϯదԠ • ࣮‫ ͱݧ‬ablation studyͷ࣮ࢪ 38

39.

໰୊ઃఆɾ΍Γ͍ͨ͜ͱ • ηάϝϯςʔγϣϯλεΫʹ͓͚ΔυϝΠϯదԠ 39

40.

ઌߦ‫ڀݚ‬: FCNs in theWild: Pixel-level Adversarial and Constraint-based Adaptation • ηϚϯςΟοΫηάϝϯςʔγϣ ϯͷϞσϧʹॳΊͯυϝΠϯద ԠΛద༻ͨ͠ɻ • Global domain alignment – ը૾ͷυϝΠϯΛదԠͤ͞ Δ • Category specific adaptation – ෼ྨϚοϓͷ֤Ϋϥεͷׂ ߹ʹ੍໿ FCNs in theWild: Pixel-level Adversarial and Constraint-based Adaptation [J. Hoffman 2016] 40

41.

ઌߦ‫ڀݚ‬: CYCADA: CYCLE-CONSISTENT ADVERSARIAL DOMAIN ADAPTATION • Cycle-GANͰιʔευϝΠϯͷ ը૾ΛλʔήοτυϝΠϯʹม ‫ͯ͠׵‬෼ྨ‫ث‬ͷֶशΛ͢Δ 41

42.

ωοτϫʔΫ֓‫؍‬ • ιʔευϝΠϯͱλʔήοτυϝΠϯͰωοτϫʔΫͷॏΈΛ‫ڞ‬༗ • υϝΠϯదԠ͸ఢରతֶशʹΑΓߦ͏ 42

43.

ωοτϫʔΫ֓‫؍‬ ηάϝϯςʔ γϣϯͷଛࣦ (Cross Entropy) ଛࣦؔ਺͸ӈͷࣜʹͳΔ 43

44.

ωοτϫʔΫ֓‫؍‬ ϥϕϧϚοϓ ͷ෼෍Λۙͮ ͚Δଛࣦ (Adversarial Loss) ଛࣦؔ਺͸ӈͷࣜʹͳΔ 44

45.

ωοτϫʔΫ֓‫؍‬ ԼҐͷ૚Ͱ΋ಉ༷ ͷଛࣦΛ‫͢ࢉܭ‬Δ (Multi-level Learning) 45

46.

ωοτϫʔΫ֓‫؍‬ ࠷খԽ͢Δଛࣦ͸ 46

47.

࣮‫ݧ‬ • ωοτϫʔΫߏ଄ – Discriminator • DCGANͷ΋ͷͱࣅͨߏ଄(5convolution, leaky ReLU) – Segmentation Network • DeepLab v2 (ResNet101ϕʔε) • conv4, 5ͷग़ྗΛ༻͍ͯ෼ྨ (Deep Labʹ͍ͭͯ͸ա‫ڈ‬ͷྠಡձࢿྉΛࢀর) https://www.slideshare.net/DeepLearningJP2016/dlencoderdecoder-with-atrous-separable-convolutionfor-semantic-image-segmentation 47

48.

࣮‫݁ݧ‬Ռ • GTA5͸ςϨϏήʔϜ͔Β࡞ΒΕͨσʔληοτ • ଞͷϕϯνϚʔΫΛ্ճ͍ͬͯΔ 48

49.

·ͱΊ • ηάϝϯςʔγϣϯλεΫʹ͓͚Δಓ࿏γʔϯͷυϝΠϯదԠΛߦͬ ͨ • ෼ྨ݁Ռͷ෼෍͕ۙͮ͘Α͏ʹఢରతֶशΛߦͬͨ • Multi-levelͰυϝΠϯదԠΛߦͬͨ • ࣮‫ʹݧ‬ΑΓଞͷϕϯνϚʔΫΛ্ճΔ͜ͱΛ֬ೝ 49

50.

ࢀߟࢿྉ • Adversarial Discriminative Domain Adaptation [E. Tzeng et al. CVPR 2017] • Domain Adaptive Faster R-CNN for Object Detection in the Wild [Y. Chen et al. CVPR2018] • Learning to Adapt Structured Output Space for Semantic Segmentation [Y. H. Tsai et al. CVPR2018] • υϝΠϯదԠͷαʔϕΠ࿦จ΋ࢀߟʹͳΔͱࢥ͍·͢ ➡ Deep Visual Domain Adaptation: A Survey [M. Wang et al. Neurocomputing 2018] 50