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

Amazon Rekognitionを用いてフォロワーの男女比を出す

Kazuki Ohashi
February 02, 2018

Amazon Rekognitionを用いてフォロワーの男女比を出す

Amazon RekogntionとFace APIを比べながら、フォロワーの男女比を出してみました。

Kazuki Ohashi

February 02, 2018
Tweet

More Decks by Kazuki Ohashi

Other Decks in Technology

Transcript

  1. "NB[PO3FLPHOJUJPOΛ༻͍ͯ
    ϑΥϩϫʔͷஉঁൺΛग़͢

    WEBΤϯδχΞษڧձ #05
    େڮ Ұथ @kzkohashi

    View full-size slide

  2. େڮҰथ!L[LPIBTIJ
    ΠϯϑϧΤϯαʔϚʔέςΟϯά
    αʔόαΠυ ϑϩϯτத৺
    झຯ͸ΧϨʔϥΠε

    View full-size slide

  3. ϑΥϩϫʔͷ৘ใ
    ݟͨ͜ͱ͋Γ·͔͢ʁ

    View full-size slide

  4. 4/4ͷϑΥϩϫʔ৘ใ
    *OTUBHSBNΠϯαΠτ 5XJUUFS"OBMZUJDT

    View full-size slide

  5. ଞਓͷ΋஌Γ͍ͨʂ

    View full-size slide

  6. Ͳ͏΍Ε͹Ͱ͖ΔͩΖ͏ʁ
    ϑΥϩϫʔ ϑΥϩϫʔ

    View full-size slide

  7. إೝࣝΛ࢖ͬͯஉঁྨਪʁ
    w จࣈϕʔεͰͷஉঁྨਪ͸೉ͦ͠͏
    w 4/4ʹΑͬͯ͸จࣈ͕গͳ͍৔߹͕͋Δ
    w ϓϩϑΟʔϧը૾͔Βஉ͔ঁ͔ͳΜͱ͘ͳ͘

    Θ͔Γͦ͏

    View full-size slide

  8. બఆख๏ͷௐࠪ
    ֎෦"1* ࣗલ
    'BDF"1*
    $MPVE7JTPO
    7JTVBM3FDPHOJUJPO
    0QFO#3
    ࣌ؒͱ࢖͑Δ͓ۚΛߟ͑ͯ֎෦"1*͔Βબఆʂ
    "NB[PO3FLPHOJUJPO

    View full-size slide

  9. ͜Μͳײ͡
    'BDF"1* "NB[PO3FDPHOJUJPO

    View full-size slide

  10. બఆൺ΂Δ
    w ຕͷ4/4ͷϓϩϑը૾Λ༻ҙ
    w உੑຕ ঁੑຕ ͦͷଞຕ
    w 'BDF"1*ͱ"NB[PO3FLPHOJUJPOͰൺֱ
    ͋Δํͷૉ੖Β͍͠ϒϩάͰࣝผ཰͕͔ͳΓΑ͔ͬͨʂ

    View full-size slide

  11. ൺ΂ͨ݁Ռ ਖ਼౴਺

    छྨ உੑ ঁੑ ͦͷଞ
    Amazon
    Rekognition
    112(68%) 174(75%) 512(92%)
    Face API 75(46%) 121(52%) 522(94%)
    ୯७ͳਖ਼౴਺͸"NB[PO3FLPHOJUJPO͕ଟ͍

    View full-size slide

  12. ൺ΂ͨ݁Ռʢޡ౴਺ʣ
    छྨ ঁੑ
    Amazon
    Rekognition
    13
    Face API 5
    छྨ உੑ
    Amazon
    Rekognition
    5
    Face API 3
    ౴͕͑ঁੑͷը૾
    ౴͕͑உੑͷը૾
    ޡ౴਺͸'BDF"1*ͷ΄͏͕গͳͦ͏

    View full-size slide

  13. 'BDF"1* "NB[PO3FLPHOJUJPO
    ޱͳͲ΋ݟ͑ͳ͍ͱ'BDF"1*͸ݕग़ͮ͠Β͍ʁ
    ݕग़ࣦഊ
    ݕग़੒ޭ

    View full-size slide

  14. 'BDF"1* "NB[PO3FLPHOJUJPO
    4/4ը૾͸ਅਖ਼໘Λ͋·Γ޲͍ͯͳ͍ͨΊݕग़཰ʹ͕ࠩग़ͨʁ
    ݕग़ࣦഊ
    ݕग़੒ޭ

    View full-size slide

  15. 'BDF"1* "NB[PO3FLPHOJUJPO
    ࣝผ੒ޭ͚ͨ͠Ͳ
    Ұਓ͚ͩݕग़
    ೋਓݕग़͚ͨ͠Ͳ
    ࣝผࣦഊ
    ݕग़ͪ͠Ό͑͹'BDF"1*ڧ͍ɾɾɾ

    View full-size slide

  16. 'BDF"1* "NB[PO3FLPHOJUJPO
    ͜͏͍͏͜ͱ΋͋Δʢসʣ
    ࣝผ੒ޭ͚ͨ͠Ͳ
    ͲͪΒ΋ঁੑͱޡೝࣝ

    View full-size slide

  17. બఆ·ͱΊ
    w 'BDF"1*ͷ΄͏͕ਫ਼౓͸ྑͦ͞͏
    w ͨͩɺޱͱ͔ӅΕͨΓ͍ͯ͠Δ4/4ը૾ͷݕग़཰͸ѱ͗͢Δ
    w ࣝผ཰ͱݕग़཰ͷόϥϯεͰ"NB[PO3FLPHOJUJPOʹܾఆ

    ʢ୯७ͳਖ਼౴਺ͳΒΑ͔ͬͨʣ
    "NB[PO3FLPHOJUJPO

    View full-size slide

  18. "NB[PO3FLPHOJUJPOͱ͸
    w "84͕ఏڙ͍ͯ͠Δɺը૾ࣝผαʔϏε
    w إೝ͚ࣝͩͰ͸ͳ͘ɺΦϒδΣΫτʢؠͱ͔ʣͷݕग़΍

    ൺֱͱ͔΋Ͱ͖Δ
    w ͍͍ͳͱࢥͬͨͷ͸ɺஉঁࣝผʹ৴པ౓͕͍͍ͭͯΔ
    ͍·͞Β

    View full-size slide

  19. (FOEFS\
    7BMVF.BMF
    $POpEFODF
    ^
    ৴པ౓
    ࣗ෼ͨͪͷαʔϏεʹԠͯ͡ᮢ஋ܾΊΕΔ

    View full-size slide

  20. ΞʔΩςΫνϟ
    ϓϩϑը૾
    μ΢ϯϩʔυ
    ϑΥϩϫʔ৘ใ
    σϑΥϧτը૾͔
    ൱͔

    Ξοϓϩʔυ
    ˞
    ˞͓ۚ΋͍ͬͨͳ͍ͷͰɺσϑΥϧτը૾ͷਓ͸ল͍ͯΔ
    IUUQL[LPIBTIJIBUFOBCMPHDPNFOUSZ

    ʢQZUIPOΛ࢖ͬͯ03#ͱ1FSDFQUVBM)BTIͰը૾ͷྨࣅ౓Λൺ΂ͯΈΔ

    ৴པ౓ͷᮢ஋


    உঁ৘ใ
    "NB[PO3FLPHOJUJPO
    KTPO

    View full-size slide

  21. ݁Ռ
    ౴͑
    ঁੑ
    உੑ
    ࠓճͷγεςϜ ˞

    ঁੑ
    உੑ
    ˞αϯϓϦϯά਺͸ϥϯμϜͰׂ̍ఔ౓

    View full-size slide

  22. ·ͱΊ
    w ը૾ࣝผ"1*͸৭ʑ͋Δ
    w ༻్ʹ߹Θͤͯࣝผ཰ͱݕग़཰ͷόϥϯεͰબ୒
    w ϓϩϑը૾͚ͩͰ΋உঁൺͱΕͨ

    View full-size slide

  23. ͓·͚
    ͜ͷൺ཰ʹຊ౰ʹͳΔͷ͔ࢼͨ͠
    छྨ உੑ ঁੑ ͦͷଞ
    ࣗ෼Ͱ਺͑Δ 65%(97) 35%(52) 12
    Amazon
    Rekognition
    46%(36) 54%(42) 22
    Face API 51%(21) 49%(20) 121
    ͳΔ΄Ͳɾɾɾ

    View full-size slide