論文紹介 Hardness-Aware Deep Metric Learning [CVPR 2019]

Transcript

1. )BSEOFTT"XBSF
   %FFQ
   .FUSJD
   -FBSOJOH
   $713
   
   0SBM 8FO[IBP ;IFOH
   ;IBPEPOH $IFO
   +JXFO -V
   +JF ;IPV

   %FQBSUNFOU
   PG
   "VUPNBUJPO
   5TJOHIVB
   6OJWFSTJUZ
   $IJOB FUD 1

2. ֓ཁ 2 ɾ/FHBUJWF
   TBNQMFͷ೉қ౓Λௐ੔͢ΔϑϨʔϜϫʔΫ )%.- )BSEOFTT"XBSF
   %FFQ
   .FUSJD
   -FBSOJOH ΛఏҊ  /FHBUJWF
   TBNQMFͷ೉қ౓Λજࡏ্ۭؒͷઢܗิؒʹΑΓௐ੔ ‎ ֶशঢ়گʹద੾ͳ೉͠͞ͷOFHBUJWFαϯϓϧΛੜ੒͢Δ

3. എܠ • /FHBUJWF
   TBNQMJOH͸ॏཁͳ໰୊ • ఏҊ͞Ε͍ͯΔख๏ͷଟ͘͸ɼֶशΛଅਐ͢Δ ೉͠ ͍ /FHBUJWFΛͲ͏બ୒͢Δ͔ʹয఺Λ͍͋ͯͯͨ ‑ Ұ෦ͷTBNQMFΛऔΓଓ͚Δ͜ͱʹͳΓɼજࡏۭؒͷେ

   ہతͳܗΛଊ͑Δ͜ͱ͕Ͱ͖͍ͯͳ͍ PWFSGJUUJOH 3

4. 4 ఏҊख๏֓આ ᶃ )BSEBXBSF
   GFBUVSF
   TZOUIFTJT  ΞϯΧʔʹ͚ۙͮͨOFHBUJWF ! Λੜ੒ ᶄ )BSEOFTTBOE-BCFM1SFTFSWJOH
   GFBUVSF
   TZOUIFTJT

    ੜ੒ͨ͠OFHBUJWF
   ! Λ ͷϥϕϧͱಉ͡ʹͳΔΑ͏ʹඍௐ੔ ᶃ ᶄ ! " = "

5. 5 .BOJGPME $MBTT
   " ఏҊख๏֓આ ᶃ)BSEBXBSF
   GFBUVSF
   TZOUIFTJT .BOJGPME $MBTT
   #

6. 6 .BOJGPME $MBTT
   " : → GFBUVSF
   TQBDF͔Β FNCFEEJOH
   TQBDF NFUSJD
   TQBDF ʹࣹӨ ఏҊख๏֓આ

   ᶃ)BSEBXBSF
   GFBUVSF
   TZOUIFTJT .BOJGPME $MBTT
   #

7. 7 .BOJGPME $MBTT
   " & ! = + " ! −

   " ∈ [0,1] ҎԼͷઢܗิؒʹΑΓ ʹ͚ۙͮͨΑΓ೉͍͠ ̂ Λੜ੒ ఏҊख๏֓આ ᶃ)BSEBXBSF
   GFBUVSF
   TZOUIFTJT .BOJGPME $MBTT
   #

8. 8 Hard-aware feature .BOJGPME $MBTT
   " l% !ͱ!͕ಉϥϕϧz͸ อূ͞Ε͍ͯͳ͍ ˣ !ͱಉϥϕϧʹ

   ͳΔΑ͏ͳ( !ΛϚοϓ ఏҊख๏֓આ ᶄ)BSEOFTTBOE-BCFM1SFTFSWJOH
   GFBUVSF
   TZOUIFTJT : → .BOJGPME $MBTT
   #

9. ఏҊϑϨʔϜϫʔΫ )%.- 9 : → : → .FUSJD
   OFUXPSL "VHNFOUFS HLP(Hardness-and-Label-Preserving)

   Generator Network "VHNFOUFS )-1
   (FOFSBUPS
   /FUXPSL

10. "VHNFOUFS 10 : → : → .FUSJD
    OFUXPSL "VHNFOUPS & !

    = + " ! − , "∈ 0,1 … (1) " = + + 1 − # , ! , , ! > # 1 , , ! ≤ # , ∈ 0,1 … (2) ; " ∈ $! $ ," , 1 ͱͯ͠  , ! = ! − ' % ! = + [ , ! + 1 − #] "! $ ," , , ! > # … (3) ' ! = * + [ ! " #!"# , ! + 1 − ! " #!"# $] ! − , ! , , ! > $ ! , , ! ≤ $ … (4) % = 0ͷͱ͖' ! = ͱͳͬͯ͠·͏ʜ  ʹ  Λ୅ೖ͢Δͱ = ! # $%&'ͱͯ͠

11. "VHNFOUFS
    ֶशঢ়گʹԠͨ͡೉қ౓ͷOFHBUJWFΛੜ੒ ; % # = ' + [ # $

    %&'( , # + 1 − # $ %&'( &] # − , # , , # > & # , , # ≤ & … (4) '() ʜͭલͷFQPDIͷ"WFSBHF
    NFUSJD
    MPTT
    FY
    5SJQMFU
    MPTT 11 @AB খ େ # $ %&'( 0 1 % ! = + $! $ ," (! − ) % ! = ! % !ͷ೉қ౓ easy hard MPTTͷେ͖͞ ֶशঢ়گ ʹԠͯ͡ੜ੒͢ΔOFHBUJWFͷ೉қ౓Λௐ੔

12. )-1
    (FOFSBUPS
    /FUXPSL 12 : → : → "VHNFOUPS HLP(Hardness-and-Label-Preserving) Generator Network

    9:; = <:=>; + λ?>@A = − B C + λ?>@A () , ) l% #ͱ#͕ಉϥϕϧz͸อূ͞Ε͍ͯͳ͍ ⇒ #ͱಉϥϕϧʹͳΔΑ͏ͳE #ΛϚοϓ HFOFSBUPS
    : → PCKFDUJWF
    GVODUJPO )-1
    (FOFSBUPS
    /FUXPSL 
    &OD
    
    %FD ͱͯ͠ͷ੍໿߲ ݩͷϥϕϧ Λอূ͢Δ

13. .FUSJD
    OFUXPSL PCKFDUJWF
    GVODUJPO .FUSJD
    OFUXPSL 13 : → : → .FUSJD
    OFUXPSL "VHNFOUFS

    HLP(Hardness-and-Label-Preserving) Generator Network EFGHIJ = ! K L!"#E + 1 − ! K L!"# MNO = ! K L!"#() + 1 − ! K L!"# (; ) 
    NFUSJD
    MPTT
    FY
    5SJQMFU
    MPTT
    /QBJS
    MPTT ݩͷσʔλର ੜ੒ͨ͠σʔλର ৴པͰ͖Δ 㱺 ੜ੒ͨ͠σʔλର ৴པͰ͖ͳ͍ 㱺 ݩͷσʔλର HFOFSBUPS ͕ ͷNFUSJD
    MPTTʹॏ͖Λ͓͘

14. $6#σʔληοτ ௗͷը૾ छྨ
    ܭ ຕ 5SBJO
    
     ຕ छྨ 5FTU 

    ຕ छྨ 5SBJOͱ5FTUʹಉ͡Ϋϥεͷը૾͸ଘࡏ͠ͳ͍ 㱺 ;FSPTIPU
    TFUUJOH 14

15. ࣮ݧઃఆ DMVTUFSJOH
    
    SFUSJFWBM
    UBTL 15 $MVTUFSJOH
    UBTL ධՁࢦඪ /.* ਖ਼نԽ૬ޓ৘ใྔ
    '
    3FDBMM!, 5FTU

    5SBJO Clustering task Retrieval task 3FUSJFWBM
    UBTL ֤UFTUը૾ RVFSZ
    ʹରͯ͠ ,ίۙ๣ͷ఺Λநग़͠ɼ ಉ͡Ϋϥε͕ଐ͍ͯ͠Ε͹ TDPSF
    FMTF
    TDPSF

16. .FUSJD
    MPTTͷछྨʹΑΒͣ )%.-Ͱ͸ࣝผతͳಛ௃ྔ͕ಘΒΕͨ 16

17. !"#$ ֶ͕शʹ͓͍ͯॏཁͳཁૉͰ͋Δ 17 HFJQO ͳ͠ͰϕʔεϥΠϯΛ্ճΔ 㱺 *+,- ͚ͩͰݱ࣮తͳಛ௃දݱͷϚοϐϯά͕ՄೳͰ͋ͬͨͱߟ͑ΒΕΔ

18. Ϋϥε಺ͷมԽ എܠ
    ࢹ఺
    র໌ FUD ΫϥεؒͷΘ͔ͣͳҧ͍ ௗͷ໛༷ 18 ʹରॲ