࿦จLT: Objects as Points h"ps:/ /arxiv.org/abs/1904.07850 2019/04/19 ౻ຊ༟հ 1

໨࣍ • ஶऀ৘ใ • ֓ཁ • ͜Ε·ͰͷϞσϧͱͷҧ͍ • ਫ਼౓ • ͦͷଞײ૝ 2

ஶऀ৘ใ • Xingyi Zhou(UT Aus1n) • Dequan Wang(UC Berkeley) • Philipp Krähenbühl(UT Aus1n) 3

ಛ௃ • ෺ମݕग़Ϟσϧ • ༗໊ͳྫ: SSD, YOLOv3, Re.naNet, M2Det... • ݕग़ͷΈͳΒͣ࢟੎ɾdepthɾ޲͖ɾ3d size ౳ʹ΋ద༻͍ͯ͠Δ • backbone ͱͯ͠ DLA(deep layers aggrega.on) ΍ Hourglass(CornerNet ౳Ͱ࢖ ༻) Λ࢖༻ 4

ಛ௃ • bounding box Λ࢖Θͣʹݕग़Λߦ͏Ϟσϧ(keypointਪఆ) • bounding box ༻ͷ grid ͷ୅ΘΓʹ໨͕ࡉ͔͍ heatmap(H, W Λ4Ͱׂͬͨఔ ౓ͷ΋ͷ) Λग़ྗ • heatmap ͕ߴ͍৔ॴ(௖఺) Λ෺ମͷத৺ͱਪఆ • த৺ͱͳΔ৔ॴͷ feature ͔Β෺ମͷେ͖͞ɾ཭ࢄԽޡࠩΛਪఆ • ཭ࢄԽޡࠩ = heatmap ʹͨ͠ࡍͷޡࠩ • େ͖͞ʹ͍ͭͯ͸ scale ͍ͯ͠ͳ͍(ͦͷ··ͷ஋) 5

ಛ௃ • ༧ଌϘοΫε = heatmap ͷ࠲ඪ + ༧ଌϘοΫεαΠζ + ༧ଌ཭ࢄԽޡࠩ • ֶशʹ࢖͏ heatmap ͷ఺͸ 1෺ମʹ͖ͭ 1ͭͷΈ • SSD ౳ͷΑ͏ʹ IoU ͷॏͳΓ۩߹Ͱ background ͔൱͔Λ෼͚ͳ͍ • ෳ਺ box ग़͞ͳ͍͜ͱΛલఏͱ͍ͯ͠Δ • ಉ͡ΫϥεͰॏͳͬͯ͠·͏৔߹͕͋Δ͕શମͷ 0.1 % ະຬͰ RCNN(2% ະ ຬ) ΑΓখ͍͞ 6

Πϝʔδਤ 7

͜Ε·ͰͷϞσϧͱͷҧ͍ • Object detec*on with implicit anchors(SSD, YOLO, Re*naNet ౳)ͱͷҧ͍ • CenterNet͸ശͷॏͳΓͰ͸ͳ͘ҐஔͷΈʹج͍ͮͯʮΞϯΧʔʯΛׂ౰ • લܠͱഎܠͷ෼ྨʹؔ͢Δखಈͷ͖͍͠஋͸ͳ͍(IoU 0.5 > ͱ͔) • ෺ମຖʹϙδςΟϒͳΞϯΧʔ͸1͚ͭͩͳͷͰ NMS Λඞཁͱ͠ͳ͍ • We simply extract local peaks in the keypoint heatmap • keypoint heatmap ͔ΒϩʔΧϧϐʔΫΛநग़͢Δ͚ͩͰྑ͍ 8

͜Ε·ͰͷϞσϧͱͷҧ͍ • Object detec*on with implicit anchors(SSD, YOLO, Re*naNet ౳)ͱͷҧ͍ • CenterNet͸ΑΓେ͖ͳग़ྗղ૾౓Λ࢖͏ • mask r-cnn ͱ͔ͱൺֱͯ͠ • output stride of 16 • ͜ΕʹΑΓෳ਺ͷΞϯΧʔ͕ෆཁͱͳΔʁʁʁʁ • [1711.08189] An Analysis of Scale Invariance in Object Detec*on - SNIP 9

͜Ε·ͰͷϞσϧͱͷҧ͍ • Object detec*on by keypoint es*ma*on(CornerNet, ExtremeNet ౳)ͱͷҧ͍ • ্ه 2ͭ͸ keypoint ݕग़ޙʹ ૊Έ߹ΘͤΛ grouping ͢Δඞཁ͕͋Δ • ஗͘ͳͬͯ͠·͏ • CenterNet ͸ඞཁͱ͠ͳ͍ • ଎͍ʂ 10

ਫ਼౓ 11

ਫ਼౓(M2Det ͷ݁ՌΛ໨ࢹͰ௥Ճͯ͠Έͨ) 12

ͦͷଞײ૝ • Backbone ͱͯ͠ DLA Λ࢖͑ΔͷΛॳΊͯ஌ͬͨ • Ή͠Ζ DLA ॳΊͯ஌Γ·ͨ͠ ! • NMS ͕ෆཁʹͳΔͷ͸஍ຯʹخ͍͠ • anchor ͕ফ͑Δͷ΋خ͍͠ • খ͍͞෺ମʹରͯ͠ͲΕ͚ͩରԠͰ͖Δ͔֬ೝ͠ͳ͍ͱ 13