CrowdPose: Efficient Crowded Scenes Pose Estimation and A New Benchmark (CVPR'19)

CrowdPose Efﬁcient Crowded Scenes Pose Estimation and A New Benchmark
(CVPR2019)

)VNBO1PTF&TUJNBUJPO %1PTF&TUJNBUJPO 1PTF5SBDLJOH %1PTF&TUJNBUJPO 4IBQF3FDPOTUSVDUJPO ը૾தͷ୯Ұɾෳ਺ਓ਺ͷؔઅ఺ͷ̎࣍ݩ࠲ඪΛਪఆ ಈըதͷ୯Ұɾෳ਺ਓ਺Λ௥੻͠ͳ͕Βؔઅ఺ͷ̎࣍ݩ࠲ඪΛਪఆ ը૾தͷ୯Ұɾෳ਺ਓ਺ͷؔઅ఺ͷ࣍ݩ࠲ඪΛਪఆ ୯Ұɾෳ਺ࢹ఺ͷը૾΍ಈը͔Βਓ෺ද໘ͷܗঢ়Λີʹਪఆɾ෮ݩ ຊ೔ͷτϐοΫ

%.VMUJ1PTF&TUJNBUJPO ೖྗ ग़ྗ ᶃτοϓμ΢ϯܕ ᶄϘτϜΞοϓܕ ਓͷݕ஌ ؔઅ఺ݕ஌ ؔઅ఺ݕ஌ ؔઅ఺ͷϚονϯά ෛՙਪఆ
ফඅऀߦಈਪఆ εϙʔπղੳ ෆ৹ऀݕ஌ ࠎ֨ਪఆ ׆༻ ը૾தͷ୯Ұɾෳ਺ਓ਺ͷؔઅ఺ͷ̎࣍ݩ࠲ඪΛਪఆ

छྨ τοϓμ΢ϯܕ ϘτϜΞοϓܕ ྲྀΕ ϝϦοτ ੾Γ෼͚ͨλεΫʹ͓͍ͯผ෼໺ͷطଘݚڀΛྲྀ༻Ͱ͖Δ ෺ମݕ஌ɾ̍ਓͷࠎ֨ਪఆ ਓ਺͕૿͑ͯ΋ਪ࿦଎౓͕མͪʹ͍͘ ̍ճͷॱ఻ൖͰશͯͷީิ఺Λग़ྗͰ͖Δ σϝϦοτ
ਓ෺ݕ஌ͷਫ਼౓͕ѱ͍ͱશମͷਫ਼౓ʹେ͖͘ѱӨڹΛٴ΅͢ ਓ਺͕૿͑Δͱܭࢉ͕͔͔࣌ؒΔ Ϛονϯάͷܭࢉίετ͕ߴ͍ ྫ $SPXE1PTF "MQIB1PTF 0QFO1PTF ਓͷݕ஌ ؔઅ఺ݕ஌ ؔઅ఺ݕ஌ ؔઅ఺ͷϚονϯά %.VMUJ1PTF&TUJNBUJPO ը૾தͷ୯Ұɾෳ਺ਓ਺ͷؔઅ఺ͷ̎࣍ݩ࠲ඪΛਪఆ

طଘख๏ͷ໰୊఺ ਓ͕ࠞࡶ͢Ε͹͢Δ΄Ͳ ਫ਼౓͕ஶ͘͠௿Լ͢Δ $SPXE*OEFY = 1 n n ∑ i=1
Nb i Na i n : ͋Δը૾ʹࣸͬͨਓ਺ Na i : Nb i : ͋Δਓͷ##ʹࣸͬͨຊਓͷδϣΠϯτ਺ ͋Δਓͷ##ʹࣸͬͨຊਓҎ֎ͷδϣΠϯτ਺

ఏҊख๏ͷྲྀΕ ᶃ෺ମݕ஌ :0-0W ᶄδϣΠϯτީิͷग़ྗ ᶅδϣΠϯτϚονϯά

ఏҊख๏ͷྲྀΕ ᶃ෺ମݕ஌ :0-0W ᶄδϣΠϯτީิͷग़ྗ ᶅδϣΠϯτϚονϯά ϙΠϯτɿ͋Δਓͷ#PVOEJOH#PYʹೖ͍ͬͯΔશͯ ຊਓҎ֎ؚ ͷδϣΠϯτީิΛ ॏΈΛม͑ͯ ग़ྗ͢Δ

δϣΠϯτɿ෦Ґ w -FFET4QPSUT1PTF -41 ɿ  ຕͷ୯Ұਓ෺ը૾͔Β੒Δσʔληοτ w .1**)VNBO1PTFɿ  ໿ສਓ෺ͷؔઅ఺࠲ඪ͕Ξϊςʔγϣϯ͞Εͨ 
໿ສઍຕͷෳ਺ਓ෺ը૾͔Β੒Δσʔληοτ w .4$0$0ɿ  ໿ສਓ෺ͷؔઅ఺࠲ඪ͕Ξϊςʔγϣϯ͞Εͨ  ໿ສຕͷෳ਺ਓ෺ը૾͔Β੒Δσʔληοτ w "*$IBMMFOHFSɿ  ໿ສਓ෺ͷؔઅ఺࠲ඪ͕Ξϊςʔγϣϯ͞Εͨ  ສຕͷෳ਺ਓ෺ը૾͔Β੒Δσʔληοτ ༗໊ͳσʔληοτɹ ࠓճͷ਺஋࣮ݧͰ࢖༻

δϣΠϯτީิͷਪఆ 1. DeepPose: Human Pose Estimation via Deep Neural Networks
(CVPR’14) ɾ1PTF&TUJNBUJPOʹ%-ΛॳΊͯ׆༻ͨ͠࿦จ ɾը૾Λೖྗͱ͠δϣΠϯτͷ࠲ඪΛ ༧ଌ͢Δճؼ໰୊ͱͯ͠ఆࣜԽ 2. Eﬃcient Object Localization Using Convolutional Networks (CVPR’15) ɾը૾Λೖྗͱͨ͠ δϣΠϯτத৺ͷਖ਼ن෼෍Λ ਖ਼ղϥϕϧͱͯ͠༻͍ͯ ͦͷ஋Λ༧ଌ͢Δճؼ໰୊ͱͯ͠ఆࣜԽ ˞෼෍ ͷύϥϝλ Λग़ྗ͢ΔͷͰ͸ແ͍ DG7"& ఺ʹΑΔ༧ଌ ώʔτϚοϓʹΑΔ༧ଌ ࠓͷओྲྀ

ީิͷਪఆ ώʔτϚοϓʹΑΔ༧ଌ ࠞࡶ͍ͯ͠Δঢ়گͰ͸ɺ͋ΔਓΛݕ஌ͨ͠##ʹଞͷਓͷδϣΠϯτ͕ ࣸΓࠐΉՄೳੑ͕͋Δ δϣΠϯτΛݕ஌͍ͨ͠ ଞਓͷδϣΠϯτ͸ଞਓͷϚονϯάʹ ༗༻ͳ৘ใͱͳΔ͜ͱ͕͋Δ ର৅ऀͷδϣΠϯτ 5BSHFU+PJOUT ର৅ऀͷҎ֎ͷδϣΠϯτ
*OUFSGFSFODF+PJOUT δϣΠϯτΛݕ஌͍ͨ͠ ର৅ऀͰ͸ແ͍৘ใΛอ࣋ͨ͠·· δϣΠϯτຖʹ̍ͭͷώʔτϚοϓͰ࣮ݱ͢Δɻ

֤δϣΠϯτຖʹώʔτϚοϓΛग़ྗ͢ΔωοτϫʔΫΛ࡞੒ɾֶश͢Δ L൪໨ͷδϣΠϯτͷJ൪໨ͷத৺࠲ඪ த৺ ඪ४ภࠩ ͷਖ਼ن෼෍ L൪໨ͷδϣΠϯτͷJ൪໨ͷਓͷਖ਼ղώʔτϚοϓ
J൪໨ͷਓͷը૾ʹөΓࠐΜͩL൪໨ͷδϣΠϯτͷଞͷਓͷը૾ L൪໨ͷδϣΠϯτͷJ൪໨ͷਓͷਖ਼ղώʔτϚοϓ L൪໨ͷδϣΠϯτͷJ൪໨ͷਓͷग़ྗώʔτϚοϓ ωοτϫʔΫग़ྗ pk i : G(pk i |σ) pk i σ Tk i := G(pk i |σ) Ωk i : Ck i := ∑ p∈Ωk i G(p|σ) Pi Lossi = 1 K K ∑ k=1 MSE [Pk i , Tk i + μCk i ] ࠓճͷ਺஋࣮ݧͰ͸ ͱ͍ͯ͠Δɻ μ = 0.5 ͱ͓͘ͱطଘݚڀʹରԠ͢Δɻ μ = 0 ɾ5BSHFU+PJOUT ɾ*OUFSGFSFODF+PJOUT μ = 0.5 ީิͷਪఆ ώʔτϚοϓʹΑΔ༧ଌ

ఏҊख๏ͷྲྀΕ ᶃ෺ମݕ஌ :0-0W ᶄδϣΠϯτީิͷग़ྗ ᶅδϣΠϯτϚονϯά ϙΠϯτɿݸʑਓ୯ҐͰ͸ແ͘શମͰϚονϯά໰୊Λղ͘͜ͱʹΑΓɺ ಛʹࠞࡶঢ়گʹ͓͍ͯ શମͷ੔߹ੑΛߟྀͨ͠ࠎ֨ਪఆΛ࣮ݱ

άϥϑͷߏ੒ ϊʔυ h1 h2 vhead 1 ℋ
= {hi : ∀i ∈ {1…M}} hi : J൪໨ͷਓͷϊʔυ M : ෺ମݕ஌Ͱਓͱݕ஌ͨ͠#PVOEJOH#PYͷ਺ɹ = {vk j : GPSk ∈ {1,…, K}, j ∈ {1,…, Nk}} Nk : vk j : ෦ҐLͷδϣΠϯτϊʔυͷ਺ ෦ҐLͷδϣΠϯτϊʔυK൪໨ɹ || = ∑ k Nk vhead 2 vbody 1 vbody 2 vbody 3 1FSTPO/PEF4FU +PJOU/PEF4FU

άϥϑͷߏ੒ Τοδ ℰ = {ek i,j : ∀i, j, k}
ek i,j : J൪໨ͷਓͱݕ஌ͨ͠#PVOEJOH#PYͷதʹ /PEF ؚ͕·Ε͍ͯΔ࣌ࢬΛҾ͘ vk j 1FSTPO+PJOU&EHF = ((ℋ, ), ℰ) 1FSTPO+PJOU(SBQI wk i,j : ͷॏΈ ώʔτϚοϓͷग़ྗΛݩʹܭࢉʁ 5IFSFTQPOTFTDPSFPGUIBUDBOEJEBUFKPJOU ek i,j vhead 1 vhead 2 vbody 1 vbody 2 vbody 3 h1 h2

ྫ άϥϑͷߏ੒ vhead 1 vhead 3 vR−knee
2 vR−knee 1 vR−knee 3 vhead 1 vR−knee 1 vhead 3 vhead 2 vhead 4 vhead 2 vR−knee 2 vR−knee 3 vhead 4 h1 h2

άϥϑͷߏ੒ ϊʔυͷϚʔδ p(k) 1 − p(k) 2 2 ≤ min
{uk 1 , uk 2 } δ(k) pk i : uk i : δk : ෦ҐLͷϊʔυJ (BVTTJBOSFTQPOTFTJ[FPGUXPKPJOUTPOIFBUNBQT EFUFSNJOFECZUIF(BVTTJBOSFTQPOTFEFWJBUJPO UIFQBSBNFUFSGPSDPOUSPMMJOHEFWJBUJPO ӈͷ৚݅Λຬͨ̎ͭ͢ͷϊʔυ͸Ϛʔδͯ̍ͭ͠ͷϊʔυͱݟͳ͢ ਓ͕ॏͳ͍ͬͯΔঢ়ଶ ಛʹಉ͡෦Ґ͕͍ۙঢ়ଶ ͷ࣌ʹ ෳ਺ਓͷ͋Δ෦ҐͷώʔτϚοϓͷ࠷΋େ͖ͳ஋ͷ ΤϦΞ͕͍ۙ͜͠ͱ͕ى͜Γ͏Δ ࠨਤ Ϛονϯάޙɺಉ͡෦Ґ͕ ෳ਺ͷਓؒͷδϣΠϯτͱਪఆ͞Εͯ͠·͏ ةݥੑ͕͋Δ

ྫ άϥϑͷߏ੒ Ϛʔδલ vhead 1 vhead 3
vR−knee 2 vR−knee 1 vR−knee 3 vhead 1 vR−knee 1 vhead 3 vhead 2 vhead 4 vhead 2 vR−knee 2 vR−knee 3 vhead 4 h1 h2 ಛఆͷ৚݅Λຬͨ̎ͭ͢ͷϊʔυ͸Ϛʔδͯ̍ͭ͠ͷϊʔυͱݟͳ͢

ྫ άϥϑͷߏ੒ Ϛʔδޙ vhead 1 vhead 3
vR−knee 2 vR−knee 1 vR−knee 3 vhead 1 vR−knee 1 vhead 3 vhead 2 vhead 4 vhead 2 vR−knee 2 vR−knee 3 vhead 4 h1 h2 vhead 1 vhead 2 vR−knee 1 ຊ౰͸͜ͷϊʔυͷΠϯσοΫεΛ̎ʹऔΓସ͑Δ⬆ vhead 1 vhead 2 vR−knee 1 ಛఆͷ৚݅Λຬͨ̎ͭ͢ͷϊʔυ͸Ϛʔδͯ̍ͭ͠ͷϊʔυͱݟͳ͢

ϚονϯάఆࣜԽ ม਺ dk ij = { 1 0 ϢʔβJ͕δϣΠϯτLʹOPEFKΛ࣋ͭ ͦͷଞ
ఆࣜԽ max d = max d ∑ i,j,k w(k) i,j ⋅ d(k) i,j ∑ j d(k) i,j ≤ 1 (∀k ∈ {1,…, K}, ∀i ∈ {1,…, M}) ∑ i d(k) i,j ≤ 1 (∀k ∈ {1,…, K}, ∀j ∈ {1,…, Mk }) d(k) i,j ∈ {0,1} (∀i, j, k) subject to બ͹Εͨࢬͷ໬౓ͷ࿨ͷ࠷େԽ ਓͷ͋Δ෦Ґʹબ͹ΕΔϊʔυ͸̍ͭҎԼ ͋Δϊʔυ͸࠷େ̍ճ͔͠બ͹Εͳ͍ vhead 1 vhead 2 vbody 1 vbody 2 vbody 3 h1 h2 ͜ͷࢬΛ࠾༻͢Δ͔൱͔ͷม਺ dhead 11

ྫ άϥϑͷߏ੒ Ϛονϯάલ vhead 1 vhead 3
vR−knee 2 vR−knee 1 vR−knee 3 vhead 1 vR−knee 1 vhead 3 vhead 2 vhead 4 vhead 2 vR−knee 2 vR−knee 3 vhead 4 h1 h2 vhead 1 vhead 2 vR−knee 1 ຊ౰͸͜ͷϊʔυͷΠϯσοΫεΛ̎ʹऔΓସ͑Δ⬆ vhead 1 vhead 2 vR−knee 1

ྫ άϥϑͷߏ੒ Ϛονϯάޙ vR−knee 2 vhead 1
vR−knee 1 vhead 3 vhead 2 vhead 4 vhead 2 vR−knee 2 vR−knee 3 vhead 4 h1 h2 vhead 1 vhead 2 vR−knee 1 ຊ౰͸͜ͷϊʔυͷΠϯσοΫεΛ̎ʹऔΓସ͑Δ⬆ vhead 1 vhead 2 vR−knee 1

= ((ℋ, ), ℰ) 1FSTPO+PJOU(SBQI ͸ҎԼͷΑ͏ʹLݸͷ ෦෼άϥϑʹ෼ղͰ͖Δɻ k = ((ℋ,
(k)), ℰ(k)) k = {v(k) j : ∀j ∈ {1,…, Nk}} ℰk = {e(k) i,j : ∀i ∈ {1…M}, j ∈ {1…Nk}} vhead 1 vhead 2 vbody 1 vbody 2 vbody 3 h1 h2 vhead 1 vhead 2 h1 h2 vbody 1 vbody 2 vbody 3 h1 h2 ෦Ґຖʹ෼ղͰ͖Δʂ Ϛονϯάʹ͍ͭͯ

Ϛονϯάʹ͍ͭͯ max d = max d ∑ i,j,k w(k) i,j
⋅ d(k) i,j = K ∑ k=1 max d(k) ∑ i,j w(k) i,j ⋅ d(k) i,j = K ∑ k=1 max d(k) k = ((ℋ, ), ℰ) 1FSTPO+PJOU(SBQI ͸ҎԼͷΑ͏ʹLݸͷ ෦෼άϥϑʹ෼ղͰ͖Δɻ k = ((ℋ, (k)), ℰ(k)) k = {v(k) j : ∀j ∈ {1,…, Nk}} ℰk = {e(k) i,j : ∀i ∈ {1…M}, j ∈ {1…Nk}} ࠨͷࣄ࣮ΑΓ໨తؔ਺͸ԼهͷΑ͏ʹ෼ղͰ͖Δ ͭ·Γɺ֤෦Ґຖʹ໰୊Λಠཱʹղ͍ͯྑ͍ VQEBUFE,VIO.VOLSFTBMHPSJUINΛ༻͍ͯٻΊͨ )VOHBSJBO.BYJNVN.BUDIJOH"MHPSJUIN ಈઢ௥੻Ͱ࢖ΘΕΔΞϧΰϦζϜ

Ϛονϯάͷܭࢉྔʹ͍ͭͯ ఆٛ ೚ҙͷۭͰͳ͍෦෼άϥϑ ͕ OPEF਺Λ ͱද͢ ࠷େ ຊͷࢬ͔࣋ͨ͠ͳ͍࣌ʹ άϥϑ ͸
Ͱ͋Δͱ͍͏ɻ ୠ͠ɺ X |X| k|X| − l G (k, l) − sparse 0 ≤ l < 2k ೚ҙͷδϣΠϯτؚ͕·ΕΔਓͷ#PVOEJOH#PY͸͍͍ͤͥ̐ͭ ࠨਤ ΑͬͯҎԼͷෆ౳͕ࣜ੒ཱ L൪໨ͷδϣΠϯτʹؔ͢Δ෦෼άϥϑ ʹ͍ͭͯߟ͑Δͱ ͸ Ͱ͋Δ͜ͱ͕෼͔Δɻ k = ((ℋ, (k)), ℰ(k)) k (4,0) − sparse vR−knee 3 ℰ(k) ≤ 4 k − 0

Ϛονϯάͷܭࢉྔʹ͍ͭͯ 4QBSTFͳάϥϑͷׂ౰໰୊ -JOFBS"TTJHONFOU1SPCMFN ͷܭࢉྔ͸ ͳͷͰ $BSQBOFUPFUBM ෦෼άϥϑ ͷϚονϯά໰୊͸ Ͱղ͚Δɻ
(n2) k ((|ℋ| + |(k) |))2 લड़ͨ͠ϊʔυͷϚʔδʹΑΓ |(k) | ∼ |ℋ| ((|ℋ| + |(k) |))2 (|ℋ|)2 ࠓճͷϚονϯά໰୊ͱ/.4ͷܭࢉྔ͸౳͍͠ ਓͱݕ஌ͨ͠#PVOEJOH#PYͷ਺

σʔληοτʹ͍ͭͯ $SPXE1PTF ຊ࿦จͰߏங .4$0$0 ໿ສਓ෺ͷؔઅ఺࠲ඪ͕Ξϊςʔγϣϯ͞Εͨ໿ສຕͷෳ਺ਓ ෺ը૾͔Β੒Δσʔληοτ ਓͷΦΫϧʔδϣϯ͕͔ͳΓগͳ͍ ໿ສਓ෺ͷؔઅ఺࠲ඪ͕Ξϊςʔγϣϯ͞Εͨ໿ສຕͷෳ਺ਓ ෺ը૾͔Β੒Δσʔληοτ ਓͷΦΫϧʔδϣϯ͕͔ͳΓଟ͍
$SPXE*OEFY = 1 n n ∑ i=1 Nb i Na i n : ͋Δը૾ʹࣸͬͨਓ਺ Na i : Nb i : ͋Δਓͷ##ʹࣸͬͨຊਓͷδϣΠϯτ਺ ͋Δਓͷ##ʹࣸͬͨຊਓҎ֎ͷδϣΠϯτ਺

ධՁࢦඪʹ͍ͭͯ OKS = ∑ i exp ( −d2 i 2s2k2
i ) δ (vi > 0) ∑ i δ (vi > 0) ∈ [0,1] di : ؔઅ఺Jͷਪఆ࠲ඪͱਖ਼ղ࠲ඪͷڑ཭ s : ਓ෺ͷαΠζ ki : ؔઅ఺ͷछྨຖʹઃఆ͞ΕΔఆ਺ ਪఆ͕೉͍ؔ͠અ΄Ͳେ͖͍஋Λઃఆ vi : ؔઅ఺͕Ξϊςʔγϣϯ͞Ε͍ͯΔ͔Ͳ͏͔ 0,4͕ᮢ஋Ҏ্Ͱ͋Ε͹༧ଌ͕ਖ਼ղͱΈͳ͢ 0,4ͷ஋Λʙ·ͰࠁΈʹมಈ͠ͳ͕Β༧ଌਫ਼౓ΛධՁ͢Δ N"3 N"1 0,3 0CKFDU,FZQPJOU4JNJMBSJUZ

࣮ݧ݁Ռ $SPXE1PTF %BUB $SPXE*OEFY &BTZ d .FEJVN d )BSE d
ࠞࡶঢ়گʹ͓͍ͯɺಛʹطଘख๏ΑΓߴ͍ਫ਼౓ΛތΔ

࣮ݧ݁Ռ .4$0$0 ࠞࡶ͕΄ͱΜͲͳ͍.4$0$0ʹؔͯ͠΋405"

"CMBUJPO4UVEJFT )FBENBQͷֶशʹ͓͍ͯ ޡͬͨδϣΠϯτͷϩεΛೖΕΔ͔Ͳ͏͔ͷӨڹ ϚονϯάΛղ͘ͷͰ͸ͳ͘ /PO.BYJNVN4VQSFTTJPO ʴͦͷ೿ੜ ͱͷൺֱ Lossi = 1
K K ∑ k=1 MSE [Pk i , Tk i + μCk i ] ͜ͷ߲Λআ֎ ਓͷݕग़ δϣΠϯτީิͷݕग़ /.4 (SFFEZ

2VBMJUBUJWF"OBMZTJT

࣍ʹಡΈ͍ͨ࿦จ ษڧ͍ͨ͜͠ͱ w )VOHBSJBO"MHPSJUIN w )JHIFS)3/FU4DBMF"XBSF3FQSFTFOUBUJPO-FBSOJOH GPS#PUUPN6Q)VNBO1PTF&TUJNBUJPO $713`   #PUUPNVQ"QQSPBDI 
$SPXE1PTF%BUBTFUͰ405"

ࢀߟจݙɾαΠτ w $SPXE1PTF&⒏DJFOU$SPXEFE4DFOFT1PTF&TUJNBUJPOBOE"/FX#FODINBSL $713` w %FFQ1PTF)VNBO1PTF&TUJNBUJPOWJB%FFQ/FVSBM/FUXPSLT $713` w
&⒏DJFOU0CKFDU-PDBMJ[BUJPO6TJOH$POWPMVUJPOBM/FUXPSLT $713` w "HVJEFUP)VNBO1PTF&TUJNBUJPOXJUI%FFQ-FBSOJOH w ίϯϐϡʔλϏδϣϯͷ࠷৽࿦จௐࠪ%)VNBO1PTF&TUJNBUJPOฤ w (JUIVC "MQIBQPTFͷΦϓγϣϯͱͯ͠૊Έࠐ·Ε͍ͯΔ   ʻ஫ҙ఺ɹҎԼͷ̎ͭͷϑΝΠϧΛผ్μ΢ϯϩʔυ͢Δඞཁ͋Γʼ  ZPMPWTQQXFJHIUT  EVD@TFQUI

CrowdPose: Efficient Crowded Scenes Pose Estima...

CrowdPose: Efficient Crowded Scenes Pose Estimation and A New Benchmark (CVPR'19)

Akihiro

More Decks by Akihiro

Other Decks in Research

Featured

Transcript

CrowdPose Efﬁcient Crowded Scenes Pose Estimation and A New Benchmark

%.VMUJ1PTF&TUJNBUJPO ೖྗ ग़ྗ ᶃτοϓμ΢ϯܕ ᶄϘτϜΞοϓܕ ਓͷݕ஌ ؔઅ఺ݕ஌ ؔઅ఺ݕ஌ ؔઅ఺ͷϚονϯά ෛՙਪఆ

छྨ τοϓμ΢ϯܕ ϘτϜΞοϓܕ ྲྀΕ ϝϦοτ ੾Γ෼͚ͨλεΫʹ͓͍ͯผ෼໺ͷطଘݚڀΛྲྀ༻Ͱ͖Δ ෺ମݕ஌ɾ̍ਓͷࠎ֨ਪఆ ਓ਺͕૿͑ͯ΋ਪ࿦଎౓͕མͪʹ͍͘ ̍ճͷॱ఻ൖͰશͯͷީิ఺Λग़ྗͰ͖Δ σϝϦοτ

طଘख๏ͷ໰୊఺ ਓ͕ࠞࡶ͢Ε͹͢Δ΄Ͳ ਫ਼౓͕ஶ͘͠௿Լ͢Δ $SPXE*OEFY = 1 n n ∑ i=1

ఏҊख๏ͷྲྀΕ ᶃ෺ମݕ஌ :0-0W ᶄδϣΠϯτީิͷग़ྗ ᶅδϣΠϯτϚονϯά

ఏҊख๏ͷྲྀΕ ᶃ෺ମݕ஌ :0-0W ᶄδϣΠϯτީิͷग़ྗ ᶅδϣΠϯτϚονϯά ϙΠϯτɿ͋Δਓͷ#PVOEJOH#PYʹೖ͍ͬͯΔશͯ ຊਓҎ֎ؚ ͷδϣΠϯτީิΛ ॏΈΛม͑ͯ ग़ྗ͢Δ

δϣΠϯτɿ෦Ґ w -FFET4QPSUT1PTF -41 ɿ  ຕͷ୯Ұਓ෺ը૾͔Β੒Δσʔληοτ w .1**)VNBO1PTFɿ  ໿ສਓ෺ͷؔઅ఺࠲ඪ͕Ξϊςʔγϣϯ͞Εͨ

δϣΠϯτީิͷਪఆ 1. DeepPose: Human Pose Estimation via Deep Neural Networks

ީิͷਪఆ ώʔτϚοϓʹΑΔ༧ଌ ࠞࡶ͍ͯ͠Δঢ়گͰ͸ɺ͋ΔਓΛݕ஌ͨ͠##ʹଞͷਓͷδϣΠϯτ͕ ࣸΓࠐΉՄೳੑ͕͋Δ δϣΠϯτΛݕ஌͍ͨ͠ ଞਓͷδϣΠϯτ͸ଞਓͷϚονϯάʹ ༗༻ͳ৘ใͱͳΔ͜ͱ͕͋Δ ର৅ऀͷδϣΠϯτ 5BSHFU+PJOUT ର৅ऀͷҎ֎ͷδϣΠϯτ

֤δϣΠϯτຖʹώʔτϚοϓΛग़ྗ͢ΔωοτϫʔΫΛ࡞੒ɾֶश͢Δ L൪໨ͷδϣΠϯτͷJ൪໨ͷத৺࠲ඪ த৺ ඪ४ภࠩ ͷਖ਼ن෼෍ L൪໨ͷδϣΠϯτͷJ൪໨ͷਓͷਖ਼ղώʔτϚοϓ

ఏҊख๏ͷྲྀΕ ᶃ෺ମݕ஌ :0-0W ᶄδϣΠϯτީิͷग़ྗ ᶅδϣΠϯτϚονϯά ϙΠϯτɿݸʑਓ୯ҐͰ͸ແ͘શମͰϚονϯά໰୊Λղ͘͜ͱʹΑΓɺ ಛʹࠞࡶঢ়گʹ͓͍ͯ શମͷ੔߹ੑΛߟྀͨ͠ࠎ֨ਪఆΛ࣮ݱ

άϥϑͷߏ੒ ϊʔυ h1 h2 vhead 1 ℋ

άϥϑͷߏ੒ Τοδ ℰ = {ek i,j : ∀i, j, k}

ྫ άϥϑͷߏ੒ vhead 1 vhead 3 vR−knee

άϥϑͷߏ੒ ϊʔυͷϚʔδ p(k) 1 − p(k) 2 2 ≤ min

ྫ άϥϑͷߏ੒ Ϛʔδલ vhead 1 vhead 3

ྫ άϥϑͷߏ੒ Ϛʔδޙ vhead 1 vhead 3

ϚονϯάఆࣜԽ ม਺ dk ij = { 1 0 ϢʔβJ͕δϣΠϯτLʹOPEFKΛ࣋ͭ ͦͷଞ

ྫ άϥϑͷߏ੒ Ϛονϯάલ vhead 1 vhead 3

ྫ άϥϑͷߏ੒ Ϛονϯάޙ vR−knee 2 vhead 1

= ((ℋ, ), ℰ) 1FSTPO+PJOU(SBQI ͸ҎԼͷΑ͏ʹLݸͷ ෦෼άϥϑʹ෼ղͰ͖Δɻ k = ((ℋ,

Ϛονϯάʹ͍ͭͯ max d = max d ∑ i,j,k w(k) i,j

Ϛονϯάͷܭࢉྔʹ͍ͭͯ ఆٛ ೚ҙͷۭͰͳ͍෦෼άϥϑ ͕ OPEF਺Λ ͱද͢ ࠷େ ຊͷࢬ͔࣋ͨ͠ͳ͍࣌ʹ άϥϑ ͸

Ϛονϯάͷܭࢉྔʹ͍ͭͯ 4QBSTFͳάϥϑͷׂ౰໰୊ -JOFBS"TTJHONFOU1SPCMFN ͷܭࢉྔ͸ ͳͷͰ $BSQBOFUPFUBM ෦෼άϥϑ ͷϚονϯά໰୊͸ Ͱղ͚Δɻ

ධՁࢦඪʹ͍ͭͯ OKS = ∑ i exp ( −d2 i 2s2k2

࣮ݧ݁Ռ $SPXE1PTF %BUB $SPXE*OEFY &BTZ d .FEJVN d )BSE d

࣮ݧ݁Ռ .4$0$0 ࠞࡶ͕΄ͱΜͲͳ͍.4$0$0ʹؔͯ͠΋405"

"CMBUJPO4UVEJFT )FBENBQͷֶशʹ͓͍ͯ ޡͬͨδϣΠϯτͷϩεΛೖΕΔ͔Ͳ͏͔ͷӨڹ ϚονϯάΛղ͘ͷͰ͸ͳ͘ /PO.BYJNVN4VQSFTTJPO ʴͦͷ೿ੜ ͱͷൺֱ Lossi = 1

2VBMJUBUJWF"OBMZTJT

࣍ʹಡΈ͍ͨ࿦จ ษڧ͍ͨ͜͠ͱ w )VOHBSJBO"MHPSJUIN w )JHIFS)3/FU4DBMF"XBSF3FQSFTFOUBUJPO-FBSOJOH GPS#PUUPN6Q)VNBO1PTF&TUJNBUJPO $713`   #PUUPNVQ"QQSPBDI

ࢀߟจݙɾαΠτ w $SPXE1PTF&⒏DJFOU$SPXEFE4DFOFT1PTF&TUJNBUJPOBOE"/FX#FODINBSL $713` w %FFQ1PTF)VNBO1PTF&TUJNBUJPOWJB%FFQ/FVSBM/FUXPSLT $713` w