Upgrade to Pro — share decks privately, control downloads, hide ads and more …

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

648cff4d467bb490796a225da8f33797?s=47 hyodo
June 10, 2019

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

研究室のゼミで"Deep Metric Learning"というタイトルで発表した資料の一部になります。ご指摘や議論等お待ちしております。
Twitter @onysuke

648cff4d467bb490796a225da8f33797?s=128

hyodo

June 10, 2019
Tweet

Transcript

  1. )BSEOFTT"XBSF%FFQ.FUSJD-FBSOJOH $7130SBM 8FO[IBP ;IFOH ;IBPEPOH $IFO +JXFO -V +JF ;IPV

    %FQBSUNFOUPG"VUPNBUJPO 5TJOHIVB6OJWFSTJUZ $IJOB FUD 1
  2. ֓ཁ 2 ɾ/FHBUJWFTBNQMFͷ೉қ౓Λௐ੔͢ΔϑϨʔϜϫʔΫ )%.- )BSEOFTT"XBSF%FFQ.FUSJD-FBSOJOH ΛఏҊ  /FHBUJWFTBNQMFͷ೉қ౓Λજࡏ্ۭؒͷઢܗิؒʹΑΓௐ੔ ‎ ֶशঢ়گʹద੾ͳ೉͠͞ͷOFHBUJWFαϯϓϧΛੜ੒͢Δ

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

    ہతͳܗΛଊ͑Δ͜ͱ͕Ͱ͖͍ͯͳ͍ PWFSGJUUJOH 3
  4. 4 ఏҊख๏֓આ ᶃ )BSEBXBSFGFBUVSFTZOUIFTJT  ΞϯΧʔʹ͚ۙͮͨOFHBUJWF ! Λੜ੒ ᶄ )BSEOFTTBOE-BCFM1SFTFSWJOHGFBUVSFTZOUIFTJT

     ੜ੒ͨ͠OFHBUJWF ! Λ ͷϥϕϧͱಉ͡ʹͳΔΑ͏ʹඍௐ੔ ᶃ ᶄ ! " = "
  5. 5 .BOJGPME $MBTT" ఏҊख๏֓આ ᶃ)BSEBXBSFGFBUVSFTZOUIFTJT .BOJGPME $MBTT#

  6. 6 .BOJGPME $MBTT" : → GFBUVSFTQBDF͔Β FNCFEEJOHTQBDF NFUSJDTQBDF ʹࣹӨ ఏҊख๏֓આ

    ᶃ)BSEBXBSFGFBUVSFTZOUIFTJT .BOJGPME $MBTT#
  7. 7 .BOJGPME $MBTT" & ! = + " ! −

    " ∈ [0,1] ҎԼͷઢܗิؒʹΑΓ ʹ͚ۙͮͨΑΓ೉͍͠ ̂ Λੜ੒ ఏҊख๏֓આ ᶃ)BSEBXBSFGFBUVSFTZOUIFTJT .BOJGPME $MBTT#
  8. 8 Hard-aware feature .BOJGPME $MBTT" l% !ͱ!͕ಉϥϕϧz͸ อূ͞Ε͍ͯͳ͍ ˣ !ͱಉϥϕϧʹ

    ͳΔΑ͏ͳ( !ΛϚοϓ ఏҊख๏֓આ ᶄ)BSEOFTTBOE-BCFM1SFTFSWJOHGFBUVSFTZOUIFTJT : → .BOJGPME $MBTT#
  9. ఏҊϑϨʔϜϫʔΫ )%.- 9 : → : → .FUSJDOFUXPSL "VHNFOUFS HLP(Hardness-and-Label-Preserving)

    Generator Network "VHNFOUFS )-1(FOFSBUPS/FUXPSL
  10. "VHNFOUFS 10 : → : → .FUSJDOFUXPSL "VHNFOUPS & !

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

    %&'( , # + 1 − # $ %&'( &] # − , # , , # > & # , , # ≤ & … (4) '() ʜͭલͷFQPDIͷ"WFSBHFNFUSJDMPTT FY5SJQMFUMPTT 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: → PCKFDUJWFGVODUJPO )-1(FOFSBUPS /FUXPSL &OD %FD ͱͯ͠ͷ੍໿߲ ݩͷϥϕϧ Λอূ͢Δ
  13. .FUSJDOFUXPSL PCKFDUJWFGVODUJPO .FUSJDOFUXPSL 13 : → : → .FUSJDOFUXPSL "VHNFOUFS

    HLP(Hardness-and-Label-Preserving) Generator Network EFGHIJ = ! K L!"#E + 1 − ! K L!"# MNO = ! K L!"#() + 1 − ! K L!"# (; ) NFUSJDMPTT FY5SJQMFUMPTT /QBJSMPTT ݩͷσʔλର ੜ੒ͨ͠σʔλର ৴པͰ͖Δ 㱺 ੜ੒ͨ͠σʔλର ৴པͰ͖ͳ͍ 㱺 ݩͷσʔλର HFOFSBUPS ͕ ͷNFUSJDMPTTʹॏ͖Λ͓͘
  14. $6#σʔληοτ ௗͷը૾ छྨ ܭ ຕ 5SBJO ຕ छྨ 5FTU 

    ຕ छྨ 5SBJOͱ5FTUʹಉ͡Ϋϥεͷը૾͸ଘࡏ͠ͳ͍ 㱺 ;FSPTIPUTFUUJOH 14
  15. ࣮ݧઃఆ DMVTUFSJOHSFUSJFWBMUBTL 15 $MVTUFSJOHUBTL ධՁࢦඪ /.* ਖ਼نԽ૬ޓ৘ใྔ ' 3FDBMM!, 5FTU

    5SBJO Clustering task Retrieval task 3FUSJFWBMUBTL ֤UFTUը૾ RVFSZ  ʹରͯ͠ ,ίۙ๣ͷ఺Λநग़͠ɼ ಉ͡Ϋϥε͕ଐ͍ͯ͠Ε͹ TDPSFFMTFTDPSF
  16. .FUSJDMPTTͷछྨʹΑΒͣ )%.-Ͱ͸ࣝผతͳಛ௃ྔ͕ಘΒΕͨ 16

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

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