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
コンピュータビジョン4.2節
Search
Takahiro Kawashima
June 13, 2018
Science
1
280
コンピュータビジョン4.2節
研究室のゼミで発表したRichard Szeliski 著,玉木徹ら訳の『コンピュータビジョン − アルゴリズムと応用』4.2節のスライド
Takahiro Kawashima
June 13, 2018
Tweet
Share
More Decks by Takahiro Kawashima
See All by Takahiro Kawashima
論文紹介:Precise Expressions for Random Projections
wasyro
0
210
ガウス過程入門
wasyro
0
250
論文紹介:Inter-domain Gaussian Processes
wasyro
0
130
論文紹介:Proximity Variational Inference (近接性変分推論)
wasyro
0
270
機械学習のための行列式点過程:概説
wasyro
0
1.3k
SOLVE-GP: ガウス過程の新しいスパース変分推論法
wasyro
1
1k
論文紹介:Stein Variational Gradient Descent
wasyro
0
890
次元削減(主成分分析・線形判別分析・カーネル主成分分析)
wasyro
0
640
論文紹介: Supervised Principal Component Analysis
wasyro
1
750
Other Decks in Science
See All in Science
白金鉱業Meetup Vol.15 DMLによる条件付処置効果の推定_sotaroIZUMI_20240919
brainpadpr
1
440
重複排除・高速バックアップ・ランサムウェア対策 三拍子そろったExaGrid × Veeam連携セミナー
climbteam
0
100
Machine Learning for Materials (Lecture 6)
aronwalsh
0
500
JSol'Ex : traitement d'images solaires en Java
melix
0
100
Transformers are Universal in Context Learners
gpeyre
0
490
多次元展開法を用いた 多値バイクラスタリング モデルの提案
kosugitti
0
190
Direct Preference Optimization
zchenry
0
270
Pericarditis Comic
camkdraws
0
670
The thin line between reconstruction, classification, and hallucination in brain decoding
ykamit
1
880
事業会社における 機械学習・推薦システム技術の活用事例と必要な能力 / ml-recsys-in-layerx-wantedly-2024
yuya4
3
210
As We May Interact: Challenges and Opportunities for Next-Generation Human-Information Interaction
signer
PRO
0
130
科学で迫る勝敗の法則(名城大学公開講座.2024年10月) / The principle of victory discovered by science (Open lecture in Meijo Univ. 2024)
konakalab
0
190
Featured
See All Featured
Bootstrapping a Software Product
garrettdimon
PRO
305
110k
4 Signs Your Business is Dying
shpigford
180
21k
RailsConf 2023
tenderlove
29
880
Rails Girls Zürich Keynote
gr2m
93
13k
Distributed Sagas: A Protocol for Coordinating Microservices
caitiem20
328
21k
How STYLIGHT went responsive
nonsquared
95
5.2k
Java REST API Framework Comparison - PWX 2021
mraible
PRO
28
7.9k
How to Think Like a Performance Engineer
csswizardry
19
1.1k
A Tale of Four Properties
chriscoyier
156
23k
Build The Right Thing And Hit Your Dates
maggiecrowley
32
2.4k
Making the Leap to Tech Lead
cromwellryan
132
8.9k
Designing on Purpose - Digital PM Summit 2013
jponch
115
6.9k
Transcript
4.2 અ Τοδ ౡوେ June 11, 2018 ిؾ௨৴େֶ ঙݚڀࣨ B4
࣍ 1. Τοδͷݕग़ 2. Τοδͷ࿈݁ 2
Τοδͷݕग़
Τοδͷݕग़ ྠֲઢͳͲͷΤοδ͖ΘΊͯଟ͘ͷใΛؚΉ ਓखʹΑΔΤοδݕग़ (ਤ 4.31) ˠ͜ΕΛύιίϯ༷ʹΒ͍ͤͨ 3
Τοδͷݕग़ ୯७ͳΤοδͷݕग़ํ๏ɿΤοδΛٸܹͳًมԽͱͯ͠ѻ͏ ˠًͷޯΛߟ͑Δ I(x) ΛϐΫηϧ x = (x, y)⊤ ্ͷًͱ͢Δͱɼًޯ
J(x) J(x) = ∇I(x) = ( ∂I ∂x , ∂I ∂y ) (x) (4.19) 4
Τοδͷݕग़ ϕΫτϧ J(x) ͷ • ͖ɿًؔͷ࠷ٸޯํ • େ͖͞ɿًؔͷมԽ߹͍ 5
Τοδͷݕग़ ߴपʹϊΠζ͕ଟ͍ ˠϩʔύεϑΟϧλͰฏԽ͔ͯ͠ΒޯΛܭࢉ ローパス フィルタ 6
Τοδͷݕग़ ϑΟϧλద༻ޙޯͷ͖͕ਖ਼͘͠อଘ͞Ε͍ͯͯ΄͍͠ ˠԁܗͷϑΟϧλ ՄೳͳԁܗϑΟϧλΨεϑΟϧλͷΈ (3.2 અɼਤ 3.14) ˠΤοδݕग़ͷͨΊͷϩʔύεϑΟϧλΨγΞϯ͕ఆ൪ 7
Τοδͷݕग़ ඍઢܗԋࢉͰ͋ΔͷͰଞͷϑΟϧλԋࢉͱՄ ΨεϑΟϧλؔΛ Gσ(x) = 1 2πσ2 exp ( −
x2 + y2 2σ2 ) ͱ͢Δ ฏԽޙͷը૾ͷޯΛ Jσ(x) ͱॻ͘ͱɼ Jσ(x) = ∇[Gσ(x) ∗ I(x)] = [∇Gσ(x)] ∗ I(x) (4.20) ͱͳΓɼΨεϑΟϧλؔͷඍͱͷͨͨΈࠐΈͰදݱͰ͖Δ 8
Τοδͷݕग़ ΨεϑΟϧλؔͷඍͷධՁ ∇Gσ(x) = ( ∂ ∂x , ∂ ∂y
)⊤ Gσ(x) = ( ∂ ∂x , ∂ ∂y )⊤ 1 2πσ2 exp ( − x2 + y2 2σ2 ) = 1 σ2 (−x, − y)⊤ 1 2πσ2 exp ( − x2 + y2 2σ2 ) ((4.21) ࣜͱ߹Θͳ͍͕ͨͿΜ͜ΕͰ͍͋ͬͯΔ) 9
Τοδͷݕग़ thinning ΤοδΛ 1 ըૉͷଠ͞Ͱදݱ͍ͨ͠߹͕ଟ͍ (ࡉઢԽ; thinning) (ը૾ [1] ΑΓ)
10
Τοδͷݕग़ thinning ʮΤοδʹରͯ͠ਨͳํͷޯڧ͕࠷େʹͳΔ࠲ඪʯΛٻ ΊΕΑ͍ ˠًͷ 2 ֊ඍ (ϥϓϥγΞϯ) Λߟ͑ΕΑͦ͞͏ͩ ͜ͷ
2 ֊ඍͷ Sσ(x) ɼ∇2 = ∇ · ∇(= div grad) ΑΓ Sσ(x) = ∇ · Jσ(x) = [∇2Gσ(x)] ∗ I(x) (4.22) 11
Τοδͷݕग़ thinning ΨεϑΟϧλͷϥϓϥγΞϯͷධՁ ∇2Gσ(x) = ∇ · [ 1 σ2
(−x, − y)⊤ 1 2πσ2 exp ( − x2 + y2 2σ2 )] = ∂ ∂x [ − x 2πσ4 exp ( − x2 + y2 2σ2 )] + ∂ ∂y [ − y 2πσ4 exp ( − x2 + y2 2σ2 )] = 1 2πσ2 ( x2 + y2 − 2σ2 σ4 ) exp ( − x2 + y2 2σ2 ) 12
Τοδͷݕग़ thinning ∇2Gσ(x) ͷΛແࢹˠ LoG(Laplacian of Gaussian) ϑΟϧλ LoG(x) =
( x2 + y2 − 2σ2 σ4 ) exp ( − x2 + y2 2σ2 ) 13
Τοδͷݕग़ thinning Sσ(x) ͷූ߸͕มԽ ˠ૬ରతͳ໌Δ͕͞มԽ Sσ(x) ͷθϩަࠩΛ୳ͤ Α͍ 14
Τοδͷݕग़ thinning sign(Sσ(xi)) ̸= sign(Sσ(xj)) ͱͳΔྡϐΫηϧ xi, xj ͓Αͼθ ϩަࠩ
xz Λ୳͢ Sσ(xi) ͱ Sσ(xj) ͱΛ݁Ϳઢ͕θϩͱަࠩ͢Δ xz ΛٻΊΔ 15
Τοδͷݕग़ thinning Sσ(xj) − Sσ(xi) xj − xi (xz −
xi) + Sσ(xi) = 0 ∴ xz = xiSσ(xj) + xjSσ(xi) Sσ(xj) + Sσ(xi) ͕ಘΒΕΔɽ3 ࣍ݩҎ্ͷ߹ಉ༷ʹ xz = xiSσ(xj) + xjSσ(xi) Sσ(xj) + Sσ(xi) (4.25) Ͱ͋Δ 16
Τοδͷݕग़ εέʔϧબͱϘέྔਪఆ LoG ʹదͳ σ ΛઃఆˠӶ͍/ಷ͍ΤοδΛநग़ (ਤ 4.32, (b), (c))
17
Τοδͷݕग़ εέʔϧબͱϘέྔਪఆ ͍ײͰΤοδΛͱΓ͍ͨͳΒʁ ˠεέʔϧεϖʔεͷΞϓϩʔν 1. ͍͔ͭ͘ͷ σ Λ༻ҙ 2. ͦΕͧΕͷ
σ ʹ͍ͭͯޯ ͱ 2 ֊ඍΛܭࢉ 3. ҆ఆʹΤοδΛݕग़Ͱ͖Δ ࠷খͷ σ ΛબɼͦΕΑΓ େ͖͍ σ Ͱݕग़͞ΕͨΤο δΛՃ 18
Τοδͷݕग़ εέʔϧબͱϘέྔਪఆ ͍ σ ͰΤοδΛநग़ (ਤ 4.32, (f)) 19
Τοδͷݕग़ Χϥʔը૾ͰͷΤοδݕग़ Χϥʔը૾ͰΤοδݕग़Λ͍ͨ͠ ୯७ʹًޯΛݟΔͱɼً৭ؒͷΤοδΛݕग़Ͱ͖ͳ͍ ղܾҊ 1ɿRGB ֤͝ͱʹًޯΛܭࢉ͢Δ • ֤৭Ͱූ߸ͷҟͳΔޯ͕ग़Δͱɼ୯७ͳ͠߹ΘͤͰ૬ ࡴ͕ى͜Δ
ղܾҊ 2ɿ֤ըૉͷपลͰہॴతͳ౷ܭྔΛ͍Ζ͍ΖௐΔ • ୯७ͳًɾ໌ɾ৭͚ͩͰͳ͘ɼςΫενϟͷมԽͳͲ ଊ͑ΒΕΔ 20
Τοδͷݕग़ ਤ 4.33ɽBGɿ໌ɼCGɿ৭ɼTGɿςΫενϟ 21
Τοδͷ࿈݁
Τοδͷ࿈݁ நग़͞ΕͨΤοδΛ࿈݁ͯ͠Ұܨ͗ʹ͍ͨ͠ thinning ͞ΕͨΤοδͷըૉใΛ͍࣋ͬͯΔͱָ ˠ͍ۙΛ୳ࡧͯ͠ܨ͛Α͍ ΤοδΛ࿈݁͢ΔͱΑΓѹॖͨ͠දݱ͕ՄೳʹͳΔ 22
Τοδͷ࿈݁ νΣΠϯίʔυ 8 ͭͷํ֯ (N, NE, E, SE, S, SW,
W, NW) Λ 3bit ͰίʔυԽ (ਤ 4.34) 23
Τοδͷ࿈݁ νΣΠϯίʔυ νΣΠϯίʔυͰͷΤϯίʔυޙɼϥϯϨϯάεූ߸Ͱ͞Βʹѹ ॖͰ͖Δ ϥϯϨϯάεූ߸ ܁Γฦ͠ͷจࣈΛͦͷճͰදݱ AAAABBBCCCCC ˠ A4B3C5 24
Τοδͷ࿈݁ arc-length parameterization ʮހʯͷ͞ͱΤοδ࠲ඪΛ༻͍ͯදݱ (ਤ 4.35) 1. x0 = (1,
0.5)⊤ ͔Βελʔτ 2. s = 0 ʹ x0 ͷ࠲ඪΛͦΕͧΕϓϩοτ 3. x1 = (2, 0.5)⊤ 4. s = ∥x1 − x0∥ = 1 ʹ x1 ͷ࠲ඪΛͦΕͧΕϓϩοτ 5. ࢝ʹΔ·Ͱ܁Γฦ͢ 25
Τοδͷ࿈݁ arc-length parameterization Q. Կ͕͏Ε͍͠ͷ͔ʁ A. ϚονϯάฏԽͳͲͷॲཧ͕༰қʹͳΔ ܗঢ়ͷࣅͨΤοδΛߟ͑Δ (ਤ 4.36)
26
Τοδͷ࿈݁ arc-length parameterization 1. Τοδͷ࠲ඪͷฏۉ ¯ x0 = ∫ S
x(s)ds Λݮࢉ 2. s Λ 0 ∼ S ͔Β 0 ∼ 1 ʹਖ਼نԽ 3. ͦΕͧΕʹ͍ͭͯϑʔϦΤม 27
Τοδͷ࿈݁ arc-length parameterization ͱͷΤοδಉ͕࢜εέʔϦϯάͱճసͷҧ͍͔͠ͳ͍ ˠϑʔϦΤมͷ݁ՌڧͱҐ૬ͷζϨ͔͠ҟͳΒͳ͍ͣ (։͕࢝ҟͳΔͱઢܗͷҐ૬ͷζϨग़Δ) 28
Τοδͷ࿈݁ arc-length parameterization ࢄԽ࣌ʹੜ͡ΔϊΠζͷฏԽʹ༗ޮ ͔͠͠ී௨ʹฏԽϑΟϧλΛ͔͚Δͱॖখͯ͠ฏԽ͞ΕΔ ਤ 4.37(a), ԁͷܘ͕ॖখ͍ͯ͠Δ 29
Τοδͷ࿈݁ arc-length parameterization 2 ֊ඍʹجͮ͘Φϑηοτ߲Λ͔͢ɼΑΓେ͖ͳ (ͦ͢ͷ ͍ʁ) ฏԽϑΟϧλΛ༻͍Δ ਤ 4.37(b)
30
·ͱΊ • άϨʔεέʔϧը૾ͰًޯͰΤοδΛݕग़ ϊΠζআڈಉ࣌ʹߦ͏ͨΊʹΨγΞϯϑΟϧλͷ 1 ֊ඍ ͱͨͨΈࠐΉ • thinning ͍ͨ͠߹
LoG ϑΟϧλΛ͔͚ͯθϩަࠩΛٻ ΊΔ • Χϥʔը૾ͷΤοδݕग़໌ɾ৭ɾςΫενϟͳͲͷ౷ܭ ྔ͕༗ޮ • thinning ͞ΕͨΤοδͷ࿈݁νΣΠϯίʔυ arc-length parameterization ͕༗ޮ • arc-length parameterization ޙϚονϯάϊΠζআڈΛ͠ ͍͢ 31
References I [1] R. Rao. Image sampling, pyramids, and edge
detection. https://courses.cs.washington.edu/courses/cse455/ 09wi/Lects/lect3.pdf, 2009.