CVPR2019参加速報 本会議 3日目 / CVPR2019 Personal Memo: Day 3

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 Main Conference the 3rd day
2. Deep LearningΛ࢖͍ͬͯΔ ͷ͕ͩɺLosslessͰը૾Λѹ ॖ͢Δख๏ɻ ֤ղ૾౓ͰEntropyූ߸ԽΛ ࢖ͬͨEncoding ˠղ૾౓ͷ௿͍ը૾ͱͯ͠࢒ ͕ࠩͰΔ(͜͜ͷৄࡉ͕Θ͔ Βͳ͔ͬͨ) ˠ(܁Γฦ͠) ˠখ͞ͳը૾͸࢒͕ࠩ0ʹͳ ΔΑ͏ʹEntropyූ߸Խ σίʔυ͸ٯॱͰɻ Poster൪߸1൪ʹ ϙελʔηογϣϯ ։࢝௚ޙʹ͍͘ͱ શ͚ۙͮ͘ͳ͍ͷͰ ஫ҙɻ •
3. 1࣍ݩʹslice͢ΔWasserstein Distance ˠn࣍ݩʹslice͢ΔWD(͜͜·Ͱैདྷख๏) n͕Exponentialʹେ͖͘ͳΔ໰୊͕͋ͬ ͨɻ ͦͷͨΊɺn࣍ݩͷmaxΛڑ཭ͱ͢ΔWD ΛఏҊɻڑ཭ͷެཧΛຬͨ͢ɻ ·ͨɺn͸ఆ਺Ͱྑ͍ɻ
4. Ϩγϐʹ͓͚Δ ׬੒ը૾⁶खॱจॻͷ ຒΊࠐΈɻ ςΩετ͔Β΋ը૾Λੜ੒ ͠ɺGANͰຊ෺ͬΆ͘͢Δɻ ڞ༗જࡏ্ۭؒͷಛ௃ྔ͚ͩ Ͱͳ͘ɺը૾ͷྨࣅ౓΋ߟྀ ʹ͍Εͯݕࡧ͢Δ͜ͱͰੑೳ ͕޲্ͱͷ͜ͱɻ (ٯʹݴ͑͹ɺಛ௃ྔͷൺֱͩ ͚Ͱ͸͋·ΓޮՌ͕ͳ͍ʁ)
5. ࡢ೥ͷStudent Best PaperͰ͋Δ Total Captureͷଓฤɻ ࡢ೥͸إ΍ମɺखͷ࢟੎౳ͷਪఆ ͸ผʑʹ΍ͬͯɺ઀ଓ෦Ͱແཧ΍ ΓҐஔ߹Θ͍͚ͤͯͨ͠Ͳɺ͜Ε Λ΋͏গ͠౷߹తͳਪఆʹͨ͠΋ ͷɻओʹɺख΍إ΋౷߹ͨ͠Part Orientation FieldsͱɺMesh TrackingʹΑΔମද໘ͷಈ͖ͷਪ ఆ͔ΒͳΔɻ ֶश༻σʔλ΋૿͑ͨ໛༷ɻ Tired
6. ը૾͔Β࢟੎Λਪఆ͢Δω οτϫʔΫΛ࢖͍ͭͭɺը ૾Λճసͤͯ͞ɺ͝ʹΐ͝ ʹΐͱ͢Δ͜ͱͰ࢟੎Ҏ֎ ͷಛ௃Λநग़͢Δωοτϫ ʔΫΛఏҊɻ
7. Flickerͷhash tagΛ࢖ͬͯɺͱʹ͔ ͨ͘͘͞ΜσʔλΛ࢖ͬͯpre- training͢Δ͜ͱΛ໨తͱͨ͠ख ๏ɻhash tagͷ͏ͪɺಈ࡞ͱؔ܎ ͳ͍΋ͷ΋ଟ͍ͷͰɺطଘͷσʔ ληοτͷΞϊςʔγϣϯͱྨࣅ ͨ͠λάͷΈΛऑڭࢣσʔλͱ͠ ͯ࠾༻ɻҎ্ͷॲཧΛશͯࣗಈ Ͱɺ1ϲֶ݄͔͚ۙͯ͘शͨ͠ͱͷ ͜ͱɻ 0 -_- ͍ͭͯ ͱͯ͠
8. લఏ: ಉ͡खॱͷ࡞ۀΛͨ͠ಈը͕ෳ਺ ͋Δɻ ໨త: ࡞ۀͷதͷಈ࡞ΧςΰϦΛɺͦͷ ಛ௃ྔͱͱ΋ʹࣗಈͰಛఆɻ ํ๏: ࡞ۀͷ։͔࢝Βऴྃ·Ͱͷ࣌ࠁtΛ 0ʙ1Ͱද͢ɻ ֤࣌ࠁͷը૾ྻΛೖྗͱͯ͠ɺtΛਪఆ ͢Δself-supervised learningΛߦ͏ɻ (ಈ࡞ͷॱ൪ʹ܏޲͕͋ΔͳΒಛ௃ྔͱt ͕݁ͼͭ͘͸ͣɺͱ͍͏Ծఆ) ͜͏ͯ͠ಘΒΕͨಛ௃ྔΛΫϥελϦϯ ά͢Δͱɺಈ࡞ΫϥεʹͳΓಘΔɻ
9. Grassman Manifolds্ͷ఺ʹରͯ͠ɺࣜ(3)ͷ Α͏ʹೋͭͷΧʔωϧΛ༻͍ͨڑ཭(?)Λ spectral clusteringʹΑΓ࠷খԽɻ ΧʔωϧΛೋͭ༻͍Δ͜ͱͷϝϦοτ͸ྑ͘ Θ͔Βͳ͔ͬͨɻGrassman Manifolds্Ͱ Spectral Clusteringɺͱ͍͏ͷ΋ੲMulti-layer Spectral Clusteringͱ͍͏࿦จͰಡΜͩΑ͏ʹ ࢥ͏͠ɺͦ͜ʹৄ͘͠ॻ͍ͯ͋ͬͨ͸ͣͳͷ ͰɺͪΌΜͱཧղ͠Α͏ͱࢥͬͨΒɺ͔ͦ͜ ΒಡΈ௚ͯ͠΋ྑ͍͔΋ɻ
10. ͜ΕɺD2AE(CVPR2018)Λ VAE࢖ͬͯ៉ྷʹ࣮૷ͨ͠ ͚ͩͰ͸???
11. Monocular DepthͰɺ࣌ܥྻ؍ଌ஋Λ஝ੵ ͢Δ͜ͱͰਫ਼౓Λ͋͛Δख๏ɻ
12. Event Camera (ޫྔͷ૿ݮͷΈΛଊ ͑ΔΧϝϥ)ʹ͓͚Δػցֶश༻ͷଛ ࣦؔ਺ͷྑ͠ѱ͠Λௐ΂ͨαʔϕ Πɻ ͜͜ʹͨ͘͞Μྻڍ͞Ε͍ͯΔɻ Β Ұ
13. Monocular DepthΛ͢Δ࣌ ʹྨࣅ͢ΔλεΫ(optical flowͱ͔৭ʑ)Λ࢖ͬͨmulti- taskʹ͠Α͏ɺͱ͍͏ͱ͜ Ζ·Ͱ͸ڈ೥ʹͲ͔͜Ͱݟ ͨؾ͕͢Δɻ ͪΐͬͱͲ͜ʹ৽نੑ͕͋ Δͷ͔·Ͱ͸཈͑ΒΕ͍ͯ ͳ͍ɻ
14. Unsupervised Domain AdaptationΛ͢Δͱ͖ʹɺGANΛ࢖ͬͯυϝΠϯΛ૿΍ͯ͠ؤ݈ੑ Λ ͋͛Δख๏ɻී௨ͷreconstruction loss + adv. loss, color consistency loss + adv. loss, fullͷ 3छྨͰυϝΠϯΛ3ͭ૿΍͢(Domain Diversification... ͍ͩͿώϡʔϦεςΟοΫ͕ͩ). ͦͷ্ͰɺMulti-source domain adaptationΛߦ͏ɻ෺ମݕग़ͰධՁɻ
15. ࣸਅʹରͯ͠ɺࣗવݴޠͰ༩͑ͨΫΤ ϦʹΑͬͯɺͲͷ෦෼͕ݴٴ͞Ε͔ͨ Λattentionͱͯ͠औΓग़͢ख๏ɻ
16. ࣝผ݁Ռʹରͯ͠ɺͳͥͦ͏൑அͨ͠ͷ͔ ΋౴͑ΒΕΔΑ͏ʹͨ͠Ϟσϧɻ આ໌͸ ςΩετ+BBͰ༩͑ΒΕΔɻ આ໌ͷੜ੒෦෼͸ࣄલʹڭࢣ͋ΓͰֶ शɻ ʔ
17. ಉ͘͡આ໌Λੜ੒͢Δݚ ڀɻ # ྆ํͱ΋౦େݪాݚ ࣸਅͰ͸ͪΐ͏ͲӅΕͯ͠ ·͍ͬͯΔ͕ɺઆ໌Λ Because ࣝผ݁Ռͷௗͷಛ ௃ͷઆ໌ as ը૾1, not as ը ૾2ɺΈ͍ͨͳܗͰ 1. ಉ͡ಛ௃Λ΋ͭଞͷௗ 2. ͦͷಛ௃Λ࣋ͨͳ͍͕ࣅ ͨௗ Λදࣔ͢Δɻ
18. Hand-crafted ͳWarpingؔ਺Λೖ ΕΔ͜ͱͰdisentanglingΛ͢Δख ๏ɻwarpingؔ਺Ͱ͸Realͳը૾ Λੜ੒Ͱ͖ͳ͍ͷͰɺ݁ՌͷΫΦ ϦςΟ͸ͪΐͬͱѱ͍Α͏ͳؾ΋ ͢Δɻ
19. Video frameͷॱংΛsort͢Δself-supervised learningख๏͸CVPR2017Ͱൃද͞Ε͍ͯΔ ͕ɺframeͰ͸ͳ͘ɺvideo clipͷsortΛ͢Δख ๏ɻ 2017೥ͷࠒΑΓGPUͷmemoryαΠζ͕૿͑ ͯྑ͔ͬͨɻ
20. UDAΛ͢Δͱ͖ͷDiscriminator ͱͯ͠1࣍ݩͷSliced Wasserstein DistanceΛ༻͍ Δɻ ΧςΰϦؒʹ೚ҙͷॱংؔ܎Λ ࡞ΓɺΧςΰϦؒͷWDΛ(ॱং ʹԠͯ͡)େ͖͘͠ͳ͕Βɺυ ϝΠϯؒͷڑ཭Λখ͘͢͞Δ͜ ͱͰΑΓΑ͍UDAΛ࣮ݱɻ ޙड़͢Δ͕ɺ͜Ε͸͔ͳΓ͍͍ ײ͕ͩ͡SOTAͰ͸ͳ͍ɻ
21. ಈըͷதʹɺ΄͔ͷಈը͔ΒಘΒΕ ͨಈ෺ମͷը૾ྻΛࣗવʹຒΊࠐΉ ͜ͱͰɺ෺ମݕग़ͷ܇࿅σʔλΛਫ ૿͠ɻ
22. ࠷ऴϑϨʔϜ͔Βͷܦա࣌ࠁΛ conditionͱͯ͠ར༻͢Δ͜ͱͰɺ೚ ҙͷ࣌ࠁͷanticipationΛ͢Δख๏ɻ Epic KitchenͰ60ඵઌ·Ͱ༧ଌɻ ಉ͡࡞ۀ͕ଓ͔͘Ͳ͏͔͸౰ͯΒΕ ͍ͯΔ͕ɺ֬৴౓ͷΑ͏ͳ΋ͷ΋ಉ ࣌ʹग़ྗ͍ͨ͠ͱײͨ͡ɻ ͳ͓ɺ͋·Γԕ͘ͷະདྷ·Ͱཉுͬ ֶͯशͤ͞Δͱਫ਼౓͕Լ͕Δɺͱ͍ ͏ͷ͸΍ͬͯΈͨΒ͍͠ɻ ࠷ॳɺEpic KitchenͷAction Anticipation͸͜͏͍͏λεΫͩͱࢥ ͬͯͨɻ1ඵઌΛ༧ଌ͢ΔΜ͡Όͳ͘ ͬͯ...ɻ
23. إͷdepthը૾͸Domain InvariantͰ͋Δɺͱ ͍͏ԾઆͷԼɺdepthΛ༧ଌ͢Δmain branch ͱɺdomainຖͷdomain specific branchΛֶ शɻ֤domain specific branchͷಛ௃͕main branchͱ۠ผ͔ͭͳ͘ͳΔΑ͏ʹ͢Δ͜ͱͰ generalizationΛߦ͏ɻ depth͸domainඇґଘɺͱ͍͏ࣄલ஌͕ࣝॏ ཁʹݟ͑Δɻٯʹݴ͑͹ɺdepthηϯαʔͷछ ྨ΍إͷ֯౓ɺڑ཭ͳͲ͕ҧͬͨΒɺ͜ͷख ๏͸͏·͍͔͘ͳ͍͸ͣɻ
24. खͷಛ௃Λ෼཭͍ͯ͠Δ͸ͣͩ ͕ɺஶऀ΋͍ͳ͍͠ Inference modelsͷઆ໌΋ෆ໌ͳ ͷͰɺৄࡉΘ͔Βͣɻ
25. ೕ༮ࣇͷֶशաఔΛ໛฿͢ Δ͜ͱͰɺIncrementalͳ෺ ମࣝผΛ࣮ݱ͠Α͏ɺͱ͍ ͏νϟϨϯδɻ σʔληοτ͸ɺΧϝϥ͕ ਎ମಈ༳ͷΑ͏ͳಈ͖Λ͠ ͓ͯΓɺͦͷதͰ஫໨෺ମ ͕(೺࣋͞Εͯ৭ʑͳ֯౓ ʹ͞Ε͍ͯΔ͔ͷΑ͏ʹ)͘ Δ͘Δճ͍ͬͯΔɻ ϥϯμϜʹ৭ʑͳ෺ମ͕ॱ ൪ʹ೺࣋͞ΕΔɻ ;ͭ͏ʹ΍Δͱ Catastrophic Forgetting͕ ى͖ΔͷͰͦΕΛͲ͏ճආ ͠·͠ΐ͏ʁͱ͍͏໰୊ఏ ى(ͱ؆୯ͳbase lineΆ͍
26. खͰ೺࣋ͨ͠෦෼ͷԹ౓͕͔͋ͨͨ ͘ͳΔ͜ͱΛར༻ͯ͠ ͞·͟·ͳ෺ମ(ͷ3DϓϦϯλͰҹ ࡮ͨ͠ϨϓϦΧ)ΛඃݧऀʹѲͬͯ΋ ΒͬͯɺͲ͜Λ৮͍͔ͬͯͨΛճస ςʔϒϧʹ৐ͤͯ360౓ࡱӨͨ͠σ ʔληοτɻ ϩϘοτͷ೺࣋Λҙ͍ࣝͯ͠ΔΒ͠ ͍͕ɺखͷܗ͕ਓͱϩϘοτͰҧ͏ ͔ΒɺͲ͏࢖͑ΔΜͩΖ͏͔ɺͱͪ ΐͬͱࢥ͏͕ɺগͳ͘ͱ΋ਓ͕৮ͬ ͯྑ͍ͱ൑அͨ͠৔ॴͳͲ͸Θ͔Δ ͔ͳɻ
27. 1ϑϨʔϜ͚ͩਖ਼ղͷsegmentation͕෇͍͍ͯͯɺͦ ΕΛ௥੻͢ΔΑ͏ͳ໰୊ઃఆʹ͓͍ͯɺͦͷ1ϑϨʔ ϜΛͲ͜ʹ͢ΔͱҰ൪ਫ਼౓͕ߴ͘ͳΔ͔Λ൑ఆ͢Δ ख๏ͷఏҊɻجຊతʹ͸ը૾2ຕΛ͍ΕͯɺͲͪΒ͕ ྑ͍͔൑ఆ͢ΔωοτϫʔΫΛ࢖͍ɺόϒϧιʔτ Λͤ͞Δɻ CVPR͸Ұٳ͞Μ͔ͳʁͱࢥͬͨݩڟͷݚڀɻ
28. UDAͷSOTAɻଞͷख๏͕Adv. Training ΍DiscrepancyΛ࠷খԽ͢Δͷʹର͠ ͯɺਅ໘໨ʹυϝΠϯຖͷฏۉͱ෼ࢄ Λܭࢉͯ͠ Batch Normalizationͷ࣌ʹฏۉ0,෼ࢄ ͕EͱͳΔΑ͏ʹwhitening͢Δɺͭ· Γਅ໘໨ͳύϥϝλਪఆʹΑΔख๏ɻ ͜ΕʹՃ͑ͯMin-Entropy Consensus Lossͱ͍͏޻෉΋͍ͯ͠Δ͕ɺͦΕ͕ ͳͯ͘΋SOTAɻͳ͓ɺTarget Domain ୯ମͷࣝผ݁ՌΑΓ΋ਫ਼౓͕ߴ͍ɻ
29. ͜ΕɺҰൃ͚ܳͩͲ໘ന͔ͬͨɻ semantic segmentationʹՃ͑ͯɺ෺ମத৺ ͕ͦͷըૉ͔ΒΈͯͲͪΒ(&ڑ཭)ʹ͋Δ͔ɺ ͱ͍͏ํ޲ͷ৘ใ΋ֶशɻ͜ΕʹΑΓɺಉҰ ෺ମͷsegmentͰ͋ͬͯ΋instanceͷڥքΛ ܭࢉՄೳʹ͍ͯ͠Δɻ ํ޲ͷ৘ใΛOptical FlowΈ͍ͨʹՄࢹԽ͠ ͨ΋ͷ ʔ
30. Motion SegmentationͰRegion ProposalΛunsupervisedʹ܇࿅͢ Δख๏ɻ Medical Imagingͱ͔Ͱ໾ʹཱͪ ͦ͏ɻ
31. Ӵ੕ࣸਅͳͲͷൺֱ࣌ʹɺΑ͋͘ ΔมԽ(೔র౳ͷࡱӨ৚݅มԽ)ͱ ϨΞͳมԽ(ݐ෺͕૿͑ͨɺͳͲ) Λ۠ผ͢Δख๏ɻ 2ຕͷը૾ʹڞ௨͢ΔbranchΛ࢖ ͬͨreconstruction݁Ռͱɺೖྗ ը૾ͷޡࠩʹج͍ͮͯϨΞ͔Ͳ͏ ͔Λ൑ఆ(ݫີʹ͸Backprop.ͨ͠ ͱ͖ͷgradientͷେ͖͞ɺͱ͍ͬ ͍͕ͯͨɺࠓʹͳͬͯɺޡࠩͰ൑ ఆ͢ΔͷͱԿ͕ҧ͏ͷ͔Θ͔Βͳ ͘ͳ͍ͬͯΔͳ͏)
32. ըૉຖʹԻݯ෼཭Λ͢ Δख๏ͳͲͰ͸ɺैདྷ ख๏Ͱ͸ɺԻݯ͕ө૾ ͷϑϨʔϜ಺ʹ͋Δ͜ ͱ͕Ծఆ͞Ε͍ͯͨɻ ͜ͷԾఆΛ͸ͣͨ͢Ί ʹɺڞ༗જࡏಛ௃Λֶ शͨ͋͠ͱͰɺಈըຖ ʹΫϥελϦϯάʹΑ ͬͯରԠؔ܎ΛٻΊ ΔɻԻ͔Βಘͨಛ௃ͩ ͚͔ΒͳΔΫϥελ ͸ɺը૾಺ʹ͍ࣸͬͯ ͳ͍Իݯͱ൑ఆͰ͖ Δɻ
33. Deep NNʹAdaBoostΛద༻ ͢Δͱɺաֶशͯ݁͠Ռ͕ѱ ͘ͳΔɻ͜ͷݪҼΛɺαϯϓ ϧຖͷWeightίϯτϩʔϧ ʹ͋Δʢͭ·Γɺޡࣝผ͢ Δαϯϓϧ͕গͳ͗͢Δͱɺ ޙଓͷClassifier͕ͦΕΒʹ overfit͢Δʣͱߟ͑ɺ category-wiseͳॏΈͷߋ৽ ʹΑΔAdaBoost-likeͳΞϯ αϯϒϧख๏ΛఏҊɻ AdaBoostʹৄ͔ͬͨ͠Βൃ දऀʹײಈ͞Εͨͷ͕ͩɺ ͜Ε΋(δΣωϨʔγϣϯ)Ϊ ϟοϓ๖͑ͷҰछͩΖ͏͔ʁ
34. ֶशͷޮ཰ԽͷͨΊʹConvolutionͷνϟϯ ωϧΛάϧʔϓʹΘ͚Δʢάϧʔϓؒͷࢬ͕ ফ͑ΔͷͰύϥϝλ͕ݮΔʣɺͱ͍͏ख๏ ʹ͓͍ͯɺGroup෼͚΋ֶशͰಘΑ͏ɺͱ͍ ͏ख๏ɻ
35. ֶशʹ༻͍Δ΂͖Ͱͳ͍όΠΞεΛֶशͨ͠ ৔߹ʹϖφϧςΟΛ༩͑Δख๏ɻ ࢒೦͗͢Δͷ͸όΠΞε͕ط஌Ͱ͋Δͱ͍͏ ԾఆɻόΠΞεΛ͍ͬͯ͠Ε͹DAͰ͖Δ͠ɺ DAͰ͖ͨΒόΠΞε΋Θ͔ΔͷͰɺܲͱཛͷ ؔ܎ʹ͋Δ͸ͣͳͷʹɺҰํΛط஌ͱͨ͠Β ͩΊͰ͠ΐ...ࠪಡऀ࢓ࣄ͠ɺ͛;Μɻ