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

作業動画と手順書を対象としたマルチモーダル理解

Taichi.Nishimura
October 30, 2023
960

 作業動画と手順書を対象としたマルチモーダル理解

IBIS2023の企画セッション"Vision and Languageの最前線"にて発表した資料です

Taichi.Nishimura

October 30, 2023
Tweet

Transcript

  1. ੢ଜଠҰ 2 ݚڀ಺༰ ࡞ۀಈըͱݴޠ৘ใΛର৅ͱͨ͠ϚϧνϞʔμϧॲཧ (େֶ) ܦྺ 2019 ~ 2020೥: ژ౎େֶେֶӃ

    ৘ใֶݚڀՊ म࢜՝ఔमྃ (৿ݚڀࣨ) 2020 ~ 2023೥: ಉେֶ ത࢜՝ఔमྃ, ത࢜ (৘ใֶ) 2023೥ ~ ݱࡏ: LINEϠϑʔגࣜձࣾ Image and video AI νʔϜ ྉཧಈը͔ΒͷϨγϐੜ੒ ACMMM21, MTAP22, ACM TOMM ࣮ݧಈըͷVision&Languageσʔληοτ ICCVW21, JNLP22 (༏ल৆) Add the butter into a pan Crack the egg and stir … खॱ.1 Invert 4 times to mix खॱ. 2 Add 10μl of Alkaline Protease Solution
  2. ߨԋͷείʔϓ: ࡞ۀಈըͱ෇ਵ͢Δݴޠ৘ใ 3 ࡞ۀಈը (ྉཧ, DIY, ..etc)ͱݴޠ (खॱॻ, Ի੠ॻ͖ى͜͠, …etc)

    ͔Βֶश͢ΔVideo-and-Language … खॱॻ (ਓखͰ࡞੒, ϊΠζগ) Ի੠ॻ͖ى͜͠ (ࣗಈ, ϊΠζଟ) 1. Crack the eggs and whisk 2. Add butter into the pan 3. Pour the eggs into the pan 4. … [0~10s]: music [10~12]: hi, I am Mina … [30~50]: crack the eggs and carefully… ࡞ۀͱແؔ܎ ֶश
  3. ͳͥ࡞ۀಈըΛऔΓѻ͏ͷ͔: ֶज़తଆ໘ 5 (1) ଟ༷ͳ෺ମɾಈ࡞ͷৄࡉͳೝࣝ खݩͷࡉ͔͍ಈ࡞ͷೝ͕ࣝඞཁ ࡞ۀಈը͔Βֶशͨ͠ಛ௃͸Ұൠͷಈ࡞ೝࣝʹద༻Մ (2) ௕͍ಈըશମͷཧղ ௕͍ಈըͷཧղ͸ۙ೥ͷίϯϐϡʔλϏδϣϯྖҬͷؔ৺ͷ1ͭ

    ਓ͕ૢ࡞Λߦ͏աఔ: Ϟϊ͕੒Ռ෺ʹ޲͚ͯࠞ͟Γ߹͍ͬͯ͘ (1) ৄࡉಈ࡞ೝࣝ [Rohrbach+ CVPR12] Material list eggs butter Key events cheese step 3 step 2 step 1 added cracked stirred added stirred action(s) : state State transition add the butter into a pan crack the eggs and stir add the egg mixture and cheese and stir Procedural text (2) ಈըશମͷཧղ [Nishimura+ ACMMM21]
  4. ࠓ೔ͷ಺༰ 6 1. Web্ͷ࡞ۀಈըΛର৅ͱͨ͠ݚڀ 2. Ұਓশࢹ఺ͷ࡞ۀಈըΛର৅ͱͨ͠ݚڀ 3. զʑͷݚڀʹ͍ͭͯ: Պֶ࣮ݧಈըͷϚϧνϞʔμϧཧղ ॱ൪:

    σʔληοτ, ࣄલֶशϞσϧ, Ԡ༻՝୊ͷॱʹͦΕͧΕ঺հ Web্ͷಈը Ұਓশಈը ແؔ܎ͳϑϨʔϜ ର৅ΛζʔϜ ଟ༷ͳࢹ఺ ࡞ۀର৅Ҏ֎ ͷ෺ମ͕ଟ͘өΔ
  5. ୯ҰͷυϝΠϯʹয఺ (ͱΓΘ͚ྉཧ): YouCook2 [Zhou+ AAAI18] 9 ྉཧΛબͿཧ༝ 1. ෺ମͱಈ࡞͕ଟ༷ 2.

    ࡱӨ͠΍͘͢Web্Ͱσʔλ͕ଟ͍ YouCook2 [Zhou+ AAAI18]͕࠷΋Α͘࢖ΘΕ͍ͯΔ (ಈը਺: 2,000) ಈըʹ࣌ؒ۠ؒͷΞϊςʔγϣϯͱ۠ؒ͝ͱʹจ͕෇༩ (ਓख) ಈը͔Βͷจੜ੒΍ಈ࡞۠ؒೝࣝͷϕϯνϚʔΫͱͯ͠ར༻ ࣌ؒ۠ؒͷΞϊςʔγϣϯ จͷΞϊςʔγϣϯ (۠ؒ͝ͱ)
  6. ୯ҰͷυϝΠϯʹয఺ (ϝΠΫ): YouMakeup [Wang+ EMNLP19] 10 ϝΠΫΛબͿཧ༝: ྉཧΑΓ΋खݩͷৄࡉͳೝ͕ࣝඞཁ YouCook2 [Zhou+

    AAAI18]ͱಉ༷: ࣌ؒ۠ؒͷΞϊςʔγϣϯͱ۠ؒ͝ͱʹจ ಈը਺͸2,800ͱYouCook2ΑΓଟ͍
  7. ෳ਺ͷυϝΠϯ΁ͷ֦ு (1): COIN [Tang+ CVPR19] 11 ಈը਺͸11,827Ͱ12ͷυϝΠϯ (ྫ: ྉཧ, ిؾ޻ࣄ)ΛΧόʔ

    ࣌ؒ۠ؒͱจͷΞϊςʔγϣϯ ಈ࡞۠ؒೝࣝʹԠ༻ (ಈ࡞͕ߦΘΕ͍ͯΔ۠ؒ + ۠ؒ͝ͱͷΧςΰϦ)
  8. ෳ਺ͷυϝΠϯ΁ͷ֦ு (2): Howto100M [Miech+ ICCV19] 12 ݱঢ়࠷΋େن໛ͳ࡞ۀಈըσʔληοτ (1ԯӽ͑) ਓखͷΞϊςʔγϣϯ͸ͳ͘, ಈըʹ෇ਵ͢ΔԻ੠ॻ͖ى͜͠Λར༻

    ಈըͱݴޠͷڞ௨ຒΊࠐΈϞσϧͷֶश, θϩγϣοτಈըݕࡧͳͲͰߴੑೳ You just apply a heavy coat let it set… Ի੠ॻ͖ى͜͠ ಈը (Ұ෦) ಈը Τϯίʔμ ςΩετ Τϯίʔμ જࡏۭؒ
  9. ڧྗͳج൫Ϟσϧͷొ৔ (2) [Lin+ CVPR22] 14 MIL-NCE͸खॱͷݴޠ஌ࣝΛ༻͍͍ͯͳ͍ ✔︎ ࣄલֶश࣌ʹखॱॻ (wikiHow͔Βऩू)Λར༻ Ի੠ॻ͖ى͜͠ͱ࠷΋ߴ͍ྨࣅ౓ͷจ΁ม׵

    खॱॻ͸Ի੠ॻ͖ى͜͠ʹൺ΂ͯϊΠζ͸গͳ͍ ޙஈͷ՝୊ (ྫ: จੜ੒)ʹ͓͍ͯଞͷࣄલֶशख๏ΑΓ΋ߴ͍ੑೳΛൃش
  10. ڧྗͳج൫Ϟσϧͷొ৔ (3): Frozen in Time [Bain+ ICCV21] 15 ը૾, ಈըͱจͷରσʔλΛWeb͔Βऩूͯ͠ରরֶश

    (Howto100M΋༗) ࢹ֮Τϯίʔμ͸Vision TransformerΛ࣌ؒํ޲ʹ֦ுͨ͠Ϟσϧ ୯७ͳΞΠσΞͳ͕ΒMIL-NCEΛ྇կ͢Δθϩγϣοτಈըݕࡧੑೳ
  11. ಈը࣭໰Ԡ౴: Just Ask [Yang+ ICCV21] 18 ಈը ࣭໰ What design

    are they making? ࣭໰Ԡ౴Ϟσϧ Rose ౴͑ Howto100M͔Β(ಈը, ࣭໰, ౴͑)ͷࡾͭ૊Λࣗಈతʹߏங: HowToVQA69M ϊΠζ͕ଟ͘30%΄Ͳ͔͠ਖ਼͍͠૊Έ߹ΘͤͰ͸ͳ͍ ͜ͷσʔληοτͰֶश࣭ͨ͠໰Ԡ౴Ϟσϧ͸طଘͷಈըQA՝୊Ͱߴ͍ੑೳ
  12. ྉཧಈը͔ΒͷϨγϐੜ੒ [Nishimura+ arXiv22] 19 ೖྗ: ྉཧಈը (ࠨ) ग़ྗ: ॏཁͳΠϕϯτͱઆ໌จ (ӈ)

    طଘϞσϧ͸ΠϕϯτͱจΛฒྻʹ༧ଌ (Πϕϯτؒͷґଘؔ܎ແࢹ) ✔︎ લʹ༧ଌͨ͠ΠϕϯτͱจΛ΋ͱʹ࣍ͷखॱΛ༧ଌ Ours: multimodal recurrent prediction Event Candidates T … … 1 Add the butter into a pan 3 Crack the eggs and stir Add the egg mixture and cheese and stir 2 Input Video Events and sentences Event selector Sentence generator Multimodal memory mixing Our model DVC Event Extractor
  13. ਖ਼ղ طଘख๏ ఏҊख๏ … ਖ਼ղ طଘख๏ ఏҊख๏ (a) (b) (c)

    (d) (e) (f) (g) (a) (b) (c) (d) (e) (f) (g) (a) (b) (c) (d) (e) (f) (a) Cut the pork into slices (b) Cover the pork in plastic wrap and pound (c) Sprinkle salt and pepper on top of the meat (d) Melt butter in the pan (e) Mix eggs milk salt and pepper together (f) Dip the pork in the egg mixture and the bread crumbs (g) Fry the pork in the pan (a) Add oil and salt to a bowl (b) Mix the chicken in the flour and mix (c) Fry the chicken in a pan (d) Fry the chicken in a pan (e) Heat the oil in a pan (f) Fry the chicken in a pan (g) Fry the chicken in the pan (a) Cut the pork in half and remove the pork (b) Season the pork with salt and pepper (c) Season the pork with salt and pepper (d) Heat some butter in a pan (e) Coat the pork in the break crumbs (f) Fry the pork in a pan … … … … 20
  14. ಈըݕࡧͱจੜ੒ͷಉֶ࣌श: HiREST [Zala+ CVPR23] 21 Hierarchical Video-Moment Retrieval and Step

    Captioning (HiREST) (1) ಈըݕࡧ, (2) ಈըͷ࣌ؒ۠ؒݕࡧ, (3) खॱ͝ͱʹ෼ׂ, (4) จੜ੒ ਓखͷΞϊςʔγϣϯͷͨΊσʔλ਺গ: 1,100ఔ౓
  15. Ұਓশࢹ఺ͷར఺ 25 ར఺ 1. ະฤूͷ࡞ۀߦಈͷཧղ (Webಈը͸ฤू͞ΕΔ) 2. ਓؒͷࢹ఺: AR΍ϩϘοτ΁ͷ׆༻͕ظ଴ େن໛ͳσʔληοτͷߏங͸೉͔͕ͬͨ͠,

    EPIC-KITCHEN, Ego4DͳͲొ৔ ৽ͨͳ௅ઓ: ෳࡶͰࡉ΍͔ͳγʔϯͷཧղ, खͱ෺ମͷΠϯλϥΫγϣϯ Ego4D [Grauman+ CVPR22] EPIC-KITCHEN [Damen+ ECCV18]
  16. EPIC-KITCHENSͷ౷ܭ͓Αͼ՝୊ 31 ࢀՃऀ਺ ಈը਺ ಈ࡞Ϋϥε਺ ಈ࡞۠ؒ਺ ෺ମۣܗ਺ 32 432 149

    39,596 454,255 EPIC-KITCHENSͷ౷ܭ৘ใ EPIC-KITCHENS Challenges (Workshop at CVPR22) ྫ: action detection ಈը͕༩͑ΒΕͯ, ։࢝/ऴ͓ྃ࣌ؒΑͼಈ࡞ϥϕϧΛ༧ଌ
  17. Ego4D: ੈքن໛ͷҰਓশࢹ఺ಈըσʔληοτ [Grauman+ CVPR22] 32 Meta AIͱ13ͷେֶ૊৫͔ΒͳΔେن໛ͳҰਓশಈըσʔληοτ ଟ༷ͳυϝΠϯ (԰಺/԰֎, ೔ৗ,

    ࢓ࣄ..etc)Λऩ࿥: ࡞ۀಈը΋ଟ͍ ಈը਺: 3670࣌ؒ ࢀՃऀ਺: 931 ஍఺: 74 ࠃ: 9 γφϦΦ: 136 φϨʔγϣϯ: 385ສ ϕϯνϚʔΫ: 5
  18. Ego4D: ੈքن໛ͷҰਓশࢹ఺ಈըσʔληοτ [Grauman+ CVPR22] 33 Meta AIͱ13ͷେֶ૊৫͔ΒͳΔେن໛ͳҰਓশಈըσʔληοτ ଟ༷ͳυϝΠϯ (԰಺/԰֎, ೔ৗ,

    ࢓ࣄ..etc)Λऩ࿥: ࡞ۀಈը΋ଟ͍ ಈը਺: 3670࣌ؒ ࢀՃऀ਺: 931 ஍఺: 74 ࠃ: 9 γφϦΦ: 136 φϨʔγϣϯ: 385ສ ϕϯνϚʔΫ: 5
  19. Ego4DͷλεΫ 35 աڈ, ݱࡏ, ະདྷʹؔ͢Δ5ͭͷλεΫʹؔ࿈͢Δ14ͷϕϯνϚʔΫ (CVPR23 Workshop) աڈ: ΤϐιʔυهԱ ݱࡏ:

    खͱ෺ମ, Ի੠-ࢹ֮ͷ࿩ऀ෼཭, ࣾձΠϯλϥΫγϣϯ ະདྷ: কདྷ༧ଌ ϕϯνϚʔΫྫ ը૾͔Β2D/3Dग़ݱҐஔ ςΩετ͔Β࣌ࠁ ߦಈݕࡧ ෺ମ௥੻ ςΩετ͔Β෺ମۣܗ ϕϯνϚʔΫྫ ෺ମͷঢ়ଶมԽλΠϛϯάͷਪఆ ෺ମঢ়ଶมԽͷछྨͱ෼ྨ ࿩ऀಛఆͱॻ͖ى͜͠ ࢹઢ, ձ࿩ઌͷ༧ଌ ϕϯνϚʔΫྫ কདྷ༧ଌ
  20. ࣮ݧϓϩτίϧ 39 • ϓϩτίϧ: ࣮ݧΛ࠶ݱ͢ΔͨΊʹهड़͞Εͨखॱॻ • ಈ࡞ͱಈ࡞ର৅ͷ෺ମΛهड़ • ඞཁʹԠͯ͡ྔ, ࣌ؒ,

    ಈ࡞Λम০͢Δ෭ࢺ • ໰୊఺: खॱͷൈ͚࿙Ε΍௚ײతͰ͸ͳ͍هड़ ϛχϓϨοϓͷϓϩτίϧ  5SBOTGFSNMPGDVMUVSFPG PWFSOJHIUUP&QQFO  $FOUSJGVHFJUGPSNJOVUFT  5IPSPVHIMZSFTVTQFOEQFMMFU XJUIЖMPG$FMM3FTPMVUJPO 4PMVUJPO  ʜ  ʜ
  21. BioVL2σʔληοτ [Nishimura+ ICCVW21, JNLP22] 41 Ұਓশࢹ఺ͷ࣮ݧө૾͔ΒͳΔσʔληοτ (ө૾਺32) (a) ϓϩτίϧதͷ֤खॱͱө૾தͷΠϕϯτͷରԠؔ܎ (b)

    ϓϩτίϧதͷ෺ମͷۣܗΞϊςʔγϣϯ ࿦จͰ͸ө૾͔ΒϓϩτίϧΛੜ੒͢Δ՝୊ʹऔΓ૊Έ, ݱ࣌఺ͷ౸ୡ఺Λٞ࿦ Transfer 1-10 ml of culture overnight to Eppen Centrifuge for 5 minutes … Thoroughly resuspend pellet with 250 μl of Cell Resuspension Solution ࣌ؒ … Πϕϯτ1 खॱ1 खॱ2 Πϕϯτ2 खॱ3 Πϕϯτ3 Culture Eppen Cell Resuspension Solution Pellet (a) ࢹ֮ͱݴޠͷରԠؔ܎ͷΞϊςʔγϣϯ (b) ϓϩτίϧ಺ʹݱΕΔ෺ମͷۣܗΞϊςʔγϣϯ ϓϩτίϧ
  22. BioVL-MicroQR 45 ࣮ݧಓ۩ʹMicro QRίʔυ (σϯιʔ)ΛషΓ෇͚ͯࡱӨ ߹ܭ24ಈը (4छྨ 6ಈը), ߹ܭ2.1࣌ؒ ՝୊:

    ಈ͘෺ମͷQRίʔυͷೝࣝ཰͸ߴ͘͸ͳ͍ લ͔Βಡ·ΕΔ͜ͱΛ૝ఆ͍ͯͯࣼ͠ΊʹͳͬͨΓ͢Δͱೝࣝʹࣦഊ