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
【論文紹介】Attention-GAN
Search
nodaki
July 14, 2018
Research
0
760
【論文紹介】Attention-GAN
Attenotion-GAN for Object Transfiguration in Wild Images
https://arxiv.org/abs/1803.06798
nodaki
July 14, 2018
Tweet
Share
More Decks by nodaki
See All by nodaki
【論文紹介】Bayesian Conditional GAN (BC-GAN)
akihiro_noda
1
740
Other Decks in Research
See All in Research
LiDARセキュリティ最前線
kentaroy47
0
280
MegaParticles: GPUを利用したStein Particle Filterによる点群6自由度姿勢推定
koide3
1
540
メタ動画データセットによる動作認識の現状と可能性
yuyay
0
200
SSII2023 医療支援における画像処理研究の動向と展望
moda0
0
110
脳卒中患者・家族からみた循環器病対策推進基本計画の進捗に関する調査
japanstrokeassociation
0
540
ニフティのインナーソース導入事例 - InnerSource Commons #11
niftycorp
PRO
0
260
デフスポーツにおける支援技術 〜競技特性・ルールと技術との関係〜
slab
0
220
Weekly AI Agents News!
masatoto
12
3.6k
Bridging Continuous and Discrete Spaces: Interpretable Sentence Representation Learning via Compositional Operations
rudorudo11
0
160
クロスモーダル表現学習の研究動向: 音声関連を中心として
ryomasumura
3
610
DeepCrysTet: A Deep Learning Approach Using Tetrahedral Mesh for Predicting Properties of Crystalline Materials
tsurubee
0
380
Rの機械学習フレームワークの紹介〜tidymodelsを中心に〜 / machine_learning_with_r2024
s_uryu
0
230
Featured
See All Featured
ParisWeb 2013: Learning to Love: Crash Course in Emotional UX Design
dotmariusz
104
6.6k
GitHub's CSS Performance
jonrohan
1025
450k
Designing for Performance
lara
602
67k
In The Pink: A Labor of Love
frogandcode
138
21k
What's new in Ruby 2.0
geeforr
337
31k
Making the Leap to Tech Lead
cromwellryan
125
8.5k
Designing Dashboards & Data Visualisations in Web Apps
destraynor
226
51k
Robots, Beer and Maslow
schacon
PRO
155
7.9k
jQuery: Nuts, Bolts and Bling
dougneiner
59
7.2k
How to train your dragon (web standard)
notwaldorf
74
5.2k
Git: the NoSQL Database
bkeepers
PRO
423
63k
Fight the Zombie Pattern Library - RWD Summit 2016
marcelosomers
228
16k
Transcript
ATTENTION-GAN FOR OBJECT TRANSFIGURATION IN WILD IMAGES ʲจհʳ
֓ཁ Attention-Gan For Object Transfiguration In Wild Images ▸ ॻࢽใ
https://arxiv.org/abs/1803.06798 Submitted : 19 Mar 2018 ▸ ֓ཁ ը૾υϝΠϯؒͷࣸ૾ʢมʣΛֶश͢ΔωοτϫʔΫ AttentionػߏΛऔΓೖΕΔ͜ͱʹΑ͖ͬͯ͢ྖҬ͚ͩࣸ૾͢Δ͜ͱ͕Ͱ͖ɺ ΑΓ៉ྷͳը૾Λੜ͢Δ͜ͱʹޭ
Ϟσϧ Attention-GAN ྖҬΛΓग़͢Attention Network + มΛ୲͏Transformation Network
Ϟσϧ Attention GAN = Cycle GAN + Attention Mechanism ▸
Cycle consistency loss ▸ Attention loss X →Y →X Y→X →Y Attention cycle-consistent loss Attention sparse loss ྖҬ͕ՄೳͳݶΓখ͞ͳʢεύʔεͳʣྖҬʹͳΔΑ͏ʹL1ϊϧϜΛՃ
Ϟσϧ Supervised Learning ▸ Attention ≒ Segmentation Segmentation label͕͋ΔͷͳΒAttention network
segmentationΛղ͘Α͏ʹֶशͤ͞Εྑ͍ Attention supervised loss Total loss λcycle consistent loss ͱ attention loss ͷॏཁΛίϯτϩʔϧ͢Δ
ֶश Experiments ▸ Datasets ImageNet: tigert 1444 images, leopard 1396
images MSCOCO: horse, zebra 286 x 286 ʹϦαΠζͨ͠ޙɺϥϯμϜʹ256 x 256ͰΓग़͠ ▸ Training strategy Optimizer: Adam, LR: 0.0002 (~100epoch), 0ʹͳΔΑ͏ʹઢܗʹݮਰ(~200epoch) Batch size: 1
݁Ռ Results
݁Ռ Comparison with CycleGAN ▸ ఆੑൺֱʢࠨஈ: input, தஈ: AttentionGAN, ӈஈ:
CycleGAN) ▸ ఆྔൺֱ AttentionGANͰ ಛʹinputͷഎܠใΛอ͍ͯͯΔ AttentionGAN, CycleGANͷ ͲͪΒ͕ΑΓྑ͍ը૾Λ ੜ͍ͯ͠Δ͔ͷΞϯέʔτ݁Ռ (੨: AttentionGAN, : CycleGAN)
ߟ Ablation Analysis ▸ AttentionͷॏΈΛมߋ Sparse lossͷΛมߋ͍ͤͯ͘͞ͱ… ࣮ݧతʹ λ=1 ͕ϕετͩͬͨ
ߟ Comparison of Supervised Results segmentation label ࠐΈͰֶश͢Δํ͕ྑ͍݁Ռʹ
·ͱΊ Conclusion ▸ GANʹAttentionػߏΛΈࠐΉࣄʹΑΓɺΑΓϦΞϧͳυϝΠ ϯؒࣸ૾ΛՄೳʹͨ͠ ಛʹഎܠ෦ͷྼԽΛ͑Δ͜ͱʹޭ ▸ ྖҬʹΑΓҙΛ͏Α͏ʹɺattention cycle consistent-
loss, attention sparse lossΛఏҊ