CVPR2019参加速報 本会議 2日目 / CVPR2019 Personal Memo: Day 2

チラシの裏チラシの裏チラシの裏チラシの裏チラシの裏チラシの裏

Transcript

1. CVPR2019 The Main Conference - the 2nd day -
2. - ࡢ೥·Ͱͱಉ༷ɺݸਓͷϝϞΛެ։͍ͯ͠ΔΑ͏ͳܗͷ΋ͷͰ͢ɻ͋͘·Ͱ΋ɺͪΒ͠ͷཪతͳ ѻ͍Ͱ͓ئ͍͠·͢ɻ - ࢲͷཧղͷޡΓ΋ଟʑ͋Δͱࢥ͍·͢ͷͰɺ͝༰͍ࣻͩ͘͞ɻ - ϝϞ: visualcommonsense.com (visual commonsense reasoning) Vision-and-language navigationͱ͍͏λεΫʹ͍ͭͯࢥͬͨ͜ͱɻ - ૝ఆ͍ͯ͠ΔϩϘοτ΁ͷࢦࣔ͸ࡉ͔͗͢ΔͷͰ͸ͳ͍͔ʁ࣮༻Λߟ͑ΔͳΒʮ͓෩ ࿊ʹ͋ΔYYYΛऔ͖ͬͯͯʯ͚ͩͰࡁ·͍ͤͨɻ஌Βͳ͍৔ॴͰϩϘοτΛOBWJHBUJPO ͢Δ࣮༻ੑ͸Ͳ͜ʹ͋Δͷ͔ʁ 
   ஌Βͳ͍֗ͳΒHPPHMF
   NBQͰ΋ྑ͍ 
   ͋Δ͍͸ݴޠͰࢦࣔ͢ΔOBWJHBUFS΋஌Βͳ͍৔ॴͳΒݴޠʹΑΔ΍ΓͱΓ΋ෆཁʁ 
   ࠷ऴతʹϩϘοτ͕ਓؒΛφϏήʔτ͢Δͷ͔ʁͰɺͲ͏φϏήʔτ͢΂͖͔Λࣗ෼ ͕ኲኾ͏͜ͱͰֶशɻͦΕͳΒΘ͔Δ͔΋ɻ
3. લఏ: ਖ਼ղϥϕϧͷ෇͚ํ͸ҟ ͳΔ͕ಉ͡࡞ۀΛ͍ͯ͠Δө ૾ʹର͢Δσʔληοτ͕͋ Δɻ ໨త: ্ख྆͘ํΛ࢖ͬͯਫ਼౓ Λ͍͋͛ͨɻ 2ͭͷσʔληοτΛࠞͥͨ batchΛ࡞Δɻશ෦ʹରͯ͠ಉ ҰͷBNΛ͔͚Δͱ܏޲͕ҧ͍ ্͗ͯ͢ख͍͔͘ͳ͍ɻ·ͬ ͨ͘ҧ͏BNΛ͔͚Δͱσʔλ Λ߹ΘͤͨԸܙ͕ͳ͍ɻ ޓ͍ʹɺҰఆ਺ΛࠞͥΔΑ͏ ʹ͢ΔɻԿݸࠞͥΔ͔͸ learning to learnతʹ(?)ࣗಈਪ ఆ͠ͳ͕Β΍Δɻ ࣮ݧ݁ՌΛΈΔͱɺҧ͏࡞ۀ ͷσʔληοτΛࠞͥͯ΋ྑ ͘ͳΒͳ͍͕ɺྉཧಉ࢜ͩͱ ྑ͘ͳΔɻ 㯪
4. ըૉʹର͢ΔఢରతPertubationͰ͸ͳ͘ɺಛ௃ ྔʹର͢ΔPertubationΛ࡞੒͢ΔͱΫϥε಺/Ϋ ϥεؒ෼ࢄൺ͕ྑ͘ͳΔɻ ଟ෼ɺͭԼͷൃදͱؔ࿈͍ͯͯ͠ɺಛ௃ۭؒ಺ ͰͷQFSUVCBUJPO͸ؤ݈ੑΛ͋͛Δྑ͍%BUB
   "VHVNFOUBUJPOʹͳ͍ͬͯΔͨΊͩͱࢥ͏ɻ ਂ͍ಛ௃ྔʹରͯ͜͠ΕΛ΍Δͷͱɺද૚Ͱಛ ௃෼෍ʹऩ·ΔൣғͰఢରతQFSUVCBUJPOΛੜ ੒͢Δͷ͸΄΅ಉ͡͸ͣɻ QFSUVCBUJPO
   ઁಈ
5. ࣅͨΑ͏ͳൃද͕ଟ͗ͯ͢ৄࡉΛ๨Εͯ͠ ·ͬͨ... Cycle GANΛ࢖͍ͬͯΔͷ͸֬ ͔ɻ
6. 2্ͭͷൃදʹࣅ͍ͯΔɻ ఢରతpertubationΛɺ࣮ࡍͷσʔλ ෼෍͕࡞ΔmanifoldͷதͰ࡞Δͱؤ ݈ੑ͕͕͋Δɺͱ͍͏ใࠂɻ ͜Εɺ΋͏Data Augumentationͷε λϯμʔυͷҰͭͰྑ͍͔΋ɻ
7. υϝΠϯ͚ͩҟͳͬͯଞ͕Ұக͢Δ ϖΞ͕ඞཁͦ͏?Few-shotͱ͸͍ ͑ɺͦΜͳσʔλΛϦΞϧͰͱΔͷ ͸೉͍͠ͷͰ͸...ͱࢥͬͨɻ ͪΐͬͱࢥ͍ͬͯͨͷͱҧͬͨͷ Ͱɺࣸਅͱͬͨޙɺଟ͘͸ಡΈࠐΜ Ͱ͍ͳ͍ɻ
8. ݕग़͞ΕͨBB಺ʹɺੲͳ͕Βͷ SFHJPO proposalΛͨ͘͞Μੜ ੒ɻ Ͳ͏ͯ͠sparse annotationͰOK ʹͳΔͷ͔ɺલఏͳͲ΋ؚΊͯ Α͘Θ͔Βͳ͔ͬͨɻ
9. private branchͱcommon branch ʹ෼͚Δܗͷtwo-streamܕͷ disentangling. ೋͭͷbranchͷಛ ௃͕௚ߦ͢ΔΑ͏ͳlossΛೖΕ Δɻͬͯɺ͜ΕɺԼهͷ࿦จʹࠅ ࣅɻ https://arxiv.org/pdf/ 1804.09347.pdf
10. DomainຖʹBatch NormΛ͔͚Δ ͱྑ͍Αɺͱ͍͏ൃදɻ ͜ͷPDFͷ࠷ॳͰ঺հͨ͠ख๏ ͷԼҐޓ׵???
11. ͜Ε΋ࣸਅΛऔͬͨ͋ͱ Ͱɺࢥͬͯͨͷͱҧͬͨͷ ͰૣΊʹ࣍ʹҠಈ ΍͸Γ ೔ܦͭͱৄࡉΛΘ͢Ε͕ ͪ
12. Video Clipʹରͯ͠ɺ͋ΔtaskΛ͢Δͱ͖ʹͲͷ෺ମΛ࢖͏͔Λݕग़ɻ MS COCOͷσʔλʹରͯ͠ɺ௥ՃͰΞϊςʔγϣϯΛ͍ͯ͠ΔΒ͍͠ɻ ҙ֎ʹ΋ɺ2ਓͷΞϊςʔλؒͷਖ਼ղͷҰக౓͸90%Λ௒͑Δͱͷ͜ͱɻ serve wineʹରͯ͠glassͷྖҬΛબ୒͍ͯ͠ΔͷͰɺಓ۩Λࢦఆ͢ΔΑ ͏ʹ͍ͯ͠ΔͬΆ͍ʁ(wine bottleͰ͸ͳ͍...ͳͥ?) Ͱ
13. Video SummeryΛ࡞ΔࡍͷධՁࢦ ඪ͕ɺϥϯμϜੜ੒ʹରͯ͠΋ߴ ͘ͳͬͯ͠·ͬͨΓɺ৭ʑ໰୊͕ ͋ΔͷͰ͸ͳ͍͔ɺͱ͍͏ൃදɻ
14. summeryʹͳ͍ͬͯΔಈըͷू߹Λ realͱͯ͠ɺGANϕʔεͷֶशͰ summeryͰ͸ͳ͍ಈըͷม׵Λ͢Δ ωοτϫʔΫΛֶशɻ ैདྷͱҧͬͯɺݩಈըͱsummery͕ pairʹͳ͍ͬͯͳͯ͘ྑ͍ɻ
15. ਓ෺࢟੎ͷεέϧτϯϞσϧΛdirected graphͱͯ͠Graph NNͰಈ࡞ࣝผɻ ͜Ε͚ͩͰ৽نੑ͕͋Δͷ͔एׯٙ໰ͩ ͕...
16. FairnessΛ୲อ͢ΔͷʹɺMaximal Mean DiscrepancyͰ͸ͳ͘ɺSpurious lossͱNon-Spurious lossͱ͍͏΋ͷΛ ఏҊɻಛʹϝδϟʔͳैདྷख๏ͱͷൺ ֱ͕ͳ͍Α͏ͳ...
17. ڈ೥ͷCVPRͰ΋RLΛ࢖ͬͯMask R-CNNΛend- to-endʹ͢Δݚڀ͕͕͋ͬͨɺ͜Ε͸ɺෳ਺ͷBB ʹରͯ͠Semantic SegmentationΛ͢Δॱ൪ΛRL ͰܾΊΔ͜ͱͰਫ਼౓޲্Λ໨ࢦͨ͠΋ͷɻ
18. Instance SegmentationͱSemantic SegmentationͷؒͷΑ͏ͳ໰୊ΛఏҊɻ ͭ·Γɺಉ͡Semantic LabelͷྖҬͰ΋ɺ Πϯελϯε͕ྡΓ߹͍ͬͯΔ৔߹ʹɺͦͷڥ ք΋ݕग़͢ΔͬΆ͍ɻ
19. ੲͳ͕ΒͷώϡʔϦεςΟοΫͳ Region Proposal͔ΒUnsupervisedʹ ΑΓྑ͍Region ProposalΛੜ੒͢Δ ωοτϫʔΫΛֶशɻ
20. Ωϟϓγϣϯͷ֤entityʹର͠ ͯɺରԠ͢ΔྖҬͷBBΛΞϊς ʔγϣϯͨ͠৽͍͠σʔληο τΛެ։ɻ SupervisedʹτϨʔχϯάͨ͠ ϞσϧͰɺGrounding͕Ͱ͖Δ ͜ͱ΋֬ೝɻ
21. Entityͷ֓೦ϨϕϧʹԠͨ͡ hierarchicalͳWord embeddingͰɺ ը૾ͱςΩετͷΑΓΑ͍ڞ௨જࡏ ۭؒΛֶशɻ
22. Vision-Language NavigationͰ Imitation Learning. Ͳ͏ͯ͠Self- supervisedͳImitation Learning͕੒ Γཱͭͷ͔ྑ͘Θ͔Βͳ͔͕ͬͨɺ ਓ͕ଟ࣭ͯ͘໰Ͱ͖ͣɻ
23. τϤλͷंɻςεϥͷํ͕֨޷ྑ͔ͬ ͨɻࣸਅ͸ͳ͍͚Ͳɻ
24. ࡢ೥ͷࡈ౻͞Μͷख๏ΛɺMMDͰ ͸ͳ͘ɺMaximum Classifier DiscrepancyͰম͖௚ͨ͠Α͏ͳख ๏ʹݟ͑ΔɻIncremental?
25. SpectralNetωλ͕ٸʹ૿͑ͨҹ৅ɻ ͜Ε͸ɺΫϥελϦϯά͢Δͱ͖ʹɺಉ࣌ʹ GANͰfakeαϯϓϧ΋ੜ੒ͤ͞Δ͜ͱͰɺ ΫϥελϦϯάͰself-paced learningΛ࣮ݱ͢ Δํ๏ͱͷ͜ͱɻ ͪΐͬͱself-paced learningʹ͍ͭͯͷษڧ͕ ଍Γͳ͍͔Β͔ɺ׬શʹ͸ཧղͰ͖ͣɻ
26. Cycle GANͩͱɺgradientΛ อ࣋͢Δม਺͕Ұഋɻ Cycle consistencyͰ͸ͳ͘ɺ AEΛ2ͭʹͯ͠ɺA-Bม׵ͷ ෦෼ͱAEx2ΛಠཱʹֶशͰ͖ ΔΑ͏ʹ͢Δ͜ͱͰϝϞϦޮ ཰Λ͋͛Α͏ɺͱ͍͏࿩ͩͱ ཧղɻ
27. ͜Εɺ͔ͳΓ໘ന͔ͬͨ Overfitting͕࢝·Δͱω οτϫʔΫͷϊʔυؒͷ ڞىؔ܎ʹ໌֬ͳมԽ͕ ݱΕ͍ͯΔͷͰ͸ʁͱ͍ ͏ͷΛɺάϥϑͷύϥϝ λͰଌͬͨ΋ͷɻ ͜͜Ͱ Overfitting͕࢝ ·͍ͬͯΔΒ ͍͠ɻ t.TT?
28. Adversarial attack ʹΑͬͯޡࣝผ͕ ى͖Δͷ͕ɺ΄ͱ ΜͲൃՐ͠ͳ͍χ ϡʔϩϯ͕ൃՐ͠ ͯ͠·͏͜ͱͰ͋ Δͱ͍͏Ծઆͷ΋ ͱɺΧςΰϦ͝ͱ ʹɺྑ͘࢖ΘΕΔ NN্ͷܦ࿏Ҏ֎ Λແࢹͯࣝ͠ผ͢ ΔΑ͏ʹ͢Δ͜ͱ Ͱɺdefense͢Δ (ޡࣝผΛ๷͙)ɻ
29. Interclass domain discrepancyΛ૿΍͠ ͯɺ Intraclass domain discrepancyΛԼ͛ Δɻ Target domain͸ϥϕϧ͕ͳ͍ͷͰɺ source domainͷ֤Ϋϥεͷಛ௃ۭؒ ʹ͓͚Δத৺ΛؚΉɺे෼ʹີ౓ͷߴ ͍ΫϥελΛclassͱΈͳ͢ɻ ͪΐͬͱΞυϗοΫͳײ͕͢͡Δɻ
30. BB͸͍͋ͬͯͯΧςΰϦΛؒҧ͍͑ͯΔΑ͏ͳ݁ՌΛ্ख͘ϦαΠΫϧ͢Δɻಉ ͡ը૾಺Ͱݕग़͞Εͨଞͷϥϕϧͱͷڞىස౓ͳͲΛ࢖͏ͬΆ͍ɻ ·ͨɺBBͷݕग़ʹՃ͑ͯɺ2ͭͷBBͷۙ͞΍ۣܗ಺ΛಉҰϥϕϧͱΈͳͨ͠ (ڧҾͳ)SegmentationλεΫΛՃ͑ͨmulti-task learningʹ͢Δɻ ͜Ε΋େ෼ΞυϗοΫ...
31. ௨ৗͷUDAʹՃ͑ͯ ୯ʹSource Domainͷࣝผ͚ͩΛߦ ͏ಛ௃+Dense૚x1ΛՃ͑Δ͜ͱ Ͱɺͪΐͬͱ͚ͩਫ਼౓͕޲্͢Δɻ SOTAΑΓେ෼௿͍Ͷɺͱฉ͍ͨ ΒɺͰ΋Ͳͷख๏ʹ΋଍͢͜ͱ͕Ͱ ͖ΔΑɺͱͷ͜ͱɻ
32. ֤ը૾ʹର͢ΔSegmentation݁ ՌΛ࣌ܥྻํ޲ʹ׈Β͔ʹͳΔ Α͏ɺVideoΛ༻͍ͯ௥Ճֶश͢ Δख๏ɻ
33. SpectralNet͕ൺֱख๏ʹͳ͔ͬͨͷͩ ͕ɺ΍͍ͬͯΔ͜ͱͷ൒෼͸ SpectralNetͰ͸ͳ͍ͩΖ͏͔...???
34. DESiREͱ͔ͷྲྀΕͳΜͩͱࢥ͏ͷ͚ͩΕͲɺͲ͕͜Ұ൪ͷ৽نੑͳͷ͔ ಡΈऔΕΔ΄Ͳɺ͜ͷ෼໺ʹৄ͘͠ͳ͔ͬͨɻ
35. GANͱ͸ผͷํ๏Ͱإͷ߹੒Λ͠Α͏ͱ͢Δख๏ɻ ैདྷख๏Ͱ͋ΔIMLE͕Θ͔Βͳ͍ͷͰɺͪΐͬͱԿΛ͍ͯ͠Δ͔Θ͔Βͳ͔ͬͨɻ
36. AdaINΛఏҊɻ૚ຖʹཚ਺Λม͑ͨͱ͖ʹมԽ͢Δ΋ͷ͕ҧ͏ͷͰɺσϞ͕Impressive. https://towardsdatascience.com/explained-a-style-based-generator-architecture-for-gans- generating-and-tuning-realistic-6cb2be0f431