Upgrade to Pro
— share decks privately, control downloads, hide ads and more …
Speaker Deck
Features
Speaker Deck
PRO
Sign in
Sign up for free
Search
Search
Paper-Survey: Objects as Points
Search
fam_taro
April 19, 2019
Science
0
2.3k
Paper-Survey: Objects as Points
fam_taro
April 19, 2019
Tweet
Share
More Decks by fam_taro
See All by fam_taro
NeRFの概要と 派生系についてのふんわり紹介
fam_taro
3
4.2k
実践 PyTorch Lightning (2019/11/30 分析コンペLT会 #1)
fam_taro
3
4.6k
Paper:ShapeMask
fam_taro
0
71
Summary: Objects as Points
fam_taro
0
3.2k
Tensorコアを使った PyTorch の高速化について
fam_taro
4
4k
Sequence to Sequence Learning with Neural Networks
fam_taro
1
1.1k
Other Decks in Science
See All in Science
MCMCのR-hatは分散分析である
moricup
0
480
実力評価性能を考慮した弓道高校生全国大会の大会制度設計の提案 / (konakalab presentation at MSS 2025.03)
konakalab
2
210
Lean4による汎化誤差評価の形式化
milano0017
1
340
データマイニング - コミュニティ発見
trycycle
PRO
0
160
データマイニング - ノードの中心性
trycycle
PRO
0
280
SciPyDataJapan 2025
schwalbe10
0
270
生成検索エンジン最適化に関する研究の紹介
ynakano
2
1.4k
Collective Predictive Coding as a Unified Theory for the Socio-Cognitive Human Minds
tanichu
0
110
凸最適化からDC最適化まで
santana_hammer
1
310
Accelerated Computing for Climate forecast
inureyes
0
120
学術講演会中央大学学員会府中支部
tagtag
0
310
Transport information Geometry: Current and Future II
lwc2017
0
210
Featured
See All Featured
Speed Design
sergeychernyshev
32
1.2k
Rebuilding a faster, lazier Slack
samanthasiow
84
9.2k
Evolution of real-time – Irina Nazarova, EuRuKo, 2024
irinanazarova
9
1k
How to Create Impact in a Changing Tech Landscape [PerfNow 2023]
tammyeverts
55
3k
Exploring the Power of Turbo Streams & Action Cable | RailsConf2023
kevinliebholz
36
6.1k
RailsConf & Balkan Ruby 2019: The Past, Present, and Future of Rails at GitHub
eileencodes
140
34k
Principles of Awesome APIs and How to Build Them.
keavy
127
17k
Testing 201, or: Great Expectations
jmmastey
45
7.7k
JavaScript: Past, Present, and Future - NDC Porto 2020
reverentgeek
52
5.7k
XXLCSS - How to scale CSS and keep your sanity
sugarenia
249
1.3M
Rails Girls Zürich Keynote
gr2m
95
14k
Helping Users Find Their Own Way: Creating Modern Search Experiences
danielanewman
30
2.9k
Transcript
จLT: Objects as Points h"ps:/ /arxiv.org/abs/1904.07850 2019/04/19 ౻ຊ༟հ 1
࣍ • ஶऀใ • ֓ཁ • ͜Ε·ͰͷϞσϧͱͷҧ͍ • ਫ਼ •
ͦͷଞײ 2
ஶऀใ • Xingyi Zhou(UT Aus1n) • Dequan Wang(UC Berkeley) •
Philipp Krähenbühl(UT Aus1n) 3
ಛ • ମݕग़Ϟσϧ • ༗໊ͳྫ: SSD, YOLOv3, Re.naNet, M2Det... •
ݕग़ͷΈͳΒͣ࢟ɾdepthɾ͖ɾ3d size ʹద༻͍ͯ͠Δ • backbone ͱͯ͠ DLA(deep layers aggrega.on) Hourglass(CornerNet Ͱ ༻) Λ༻ 4
ಛ • bounding box ΛΘͣʹݕग़Λߦ͏Ϟσϧ(keypointਪఆ) • bounding box ༻ͷ grid
ͷΘΓʹ͕ࡉ͔͍ heatmap(H, W Λ4Ͱׂͬͨఔ ͷͷ) Λग़ྗ • heatmap ͕ߴ͍ॴ() Λମͷத৺ͱਪఆ • த৺ͱͳΔॴͷ feature ͔Βମͷେ͖͞ɾࢄԽޡࠩΛਪఆ • ࢄԽޡࠩ = heatmap ʹͨ͠ࡍͷޡࠩ • େ͖͞ʹ͍ͭͯ scale ͍ͯ͠ͳ͍(ͦͷ··ͷ) 5
ಛ • ༧ଌϘοΫε = heatmap ͷ࠲ඪ + ༧ଌϘοΫεαΠζ + ༧ଌࢄԽޡࠩ
• ֶशʹ͏ heatmap ͷ 1ମʹ͖ͭ 1ͭͷΈ • SSD ͷΑ͏ʹ IoU ͷॏͳΓ۩߹Ͱ background ͔൱͔Λ͚ͳ͍ • ෳ 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
• CenterNetശͷॏͳΓͰͳ͘ҐஔͷΈʹج͍ͮͯʮΞϯΧʔʯΛׂ • લܠͱഎܠͷྨʹؔ͢Δखಈͷ͖͍͠ͳ͍(IoU 0.5 > ͱ͔) • ମຖʹϙδςΟϒͳΞϯΧʔ1͚ͭͩͳͷͰ 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
• CenterNetΑΓେ͖ͳग़ྗղ૾Λ͏ • mask r-cnn ͱ͔ͱൺֱͯ͠ • output stride of 16 • ͜ΕʹΑΓෳͷΞϯΧʔ͕ෆཁͱͳΔʁʁʁʁ • [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 ͷ݁ՌΛࢹͰՃͯ͠Έͨ) 12
ͦͷଞײ • Backbone ͱͯ͠ DLA Λ͑ΔͷΛॳΊͯͬͨ • Ή͠Ζ DLA ॳΊͯΓ·ͨ͠
! • NMS ͕ෆཁʹͳΔͷຯʹخ͍͠ • anchor ͕ফ͑Δͷخ͍͠ • খ͍͞ମʹରͯ͠ͲΕ͚ͩରԠͰ͖Δ͔֬ೝ͠ͳ͍ͱ 13