2. WHO AM I 2
▸ Chung Minki
▸ BS, KAIST, IE, 2016
▸ MS, SNU, IE, 2018..?!
▸ Vision Projects
▸ Working on Semantic Image Inpainting
3. WHAT IS VISUAL ATTENTION 3
▸ Attention is HOT nowadays
▸ http://openaccess.thecvf.com/CVPR2017_search.py
▸ http://search.iclr2018.smerity.com/search/?query=attention
4. WHAT IS VISUAL ATTENTION 4
▸ Maybe heard of
▸ "Neural Machine Translation by Jointly Learning to Align and Translate"
▸ "Show, Attend, and Tell: Neural Image Caption"
Dzmitry Bahdanau, Kyunghyun Cho, Yoshua Bengio, 2015, ICLR. "Neural Machine Translation by Jointly Learning to Align and Translate"
Kelvin Xu, Jimmy Ba, Ryan Kiros, Kyunghyun Cho, Aaron Courville, Ruslan Salakhutdinov, Richard Zemel, Yoshua Bengio, 2015, ICML.
"Show, Attend, and Tell: Neural Image Caption Generation with Visual Attention"
5. WHAT IS VISUAL ATTENTION 5
▸ More,
Jimmy Lei Ba, Volodymyr Mnih, Koray Kavukcuoglu, 2015, ILCR. "Multiple Object Recognition With Visual Attention"
Max Jaderberg, Karen Simonyan, Andrew Zisserman, Koray Kavukcuoglu, NIPS, 2014. "Spatial Transformer Network"
Jianlong Fu, Heliang Zheng, Tao Mei, 2017, CVPR. "Look Closer to See Better: Recurrent Attention Convolutional Neural Network for Fine-
grained Image Recognition"
Siavash Gorji, James J. Clark, 2017, CVPR. "Attentional Push: A Deep Convolutional Network for Augmenting Image Salience
with Shared Attention Modeling in Social Scenes"
6. WHAT IS VISUAL ATTENTION 6
▸ Visual Attention:
▸ Attend on certain part of image to solve a task more efficiently
▸ Deep learning, the black box model → Interpretability
7. TABLE OF CONTENTS 7
▸ Early Works
▸ Recurrent Attention Model (RAM)
▸ Spatial Transformer Network (STN)
▸ Recent Works of visual attention
▸ in ICLR
▸ in CVPR
10. RECURRENT ATTENTION MODEL 10
▸ Volodymyr Mnih, Nicolas Heess, Alex Graves, Koray Kavukcuoglu, 2014, NIPS.
"Recurrent Models of Visual Attention"
▸ Google DeepMind, 563 citations
▸ Motivation: Confronted by large image, human process image sequentially,
selecting where and what to look
▸ Tackle ConvNet limitation: poor scalability with increasing input image size
11. RECURRENT ATTENTION MODEL 11
▸ Multiple Object Recognition with Visual Attention (DRAM), 2015, ICLR
▸ Refined architecture version of RAM
▸ RNN Structure with multi-resolution crop, called glimpse
▸ Architecture:
Jimmy Lei Ba, Volodymyr Mnih, Koray Kavukcuoglu, 2015, ILCR. "Multiple Object Recognition With Visual Attention"
12. RECURRENT ATTENTION MODEL 12
▸ Architecture:
Jimmy Lei Ba, Volodymyr Mnih, Koray Kavukcuoglu, 2015, ILCR. "Multiple Object Recognition With Visual Attention"
WHERE TO SEE
WHAT TO SEE
provide initial state
locate glimpse
outputs the inputs for rnn(1)
for multiple objects
13. RECURRENT ATTENTION MODEL 13
▸ Demo
▸ Single object classification
https://github.com/kevinzakka/recurrent-visual-attention
14. RECURRENT ATTENTION MODEL 14
▸ Training:
▸ maximize
Jimmy Lei Ba, Volodymyr Mnih, Koray Kavukcuoglu, 2015, ILCR. "Multiple Object Recognition With Visual Attention"
LOWERBOUND F
multiple object case
15. RECURRENT ATTENTION MODEL 15
▸ Cont'd:
Jimmy Lei Ba, Volodymyr Mnih, Koray Kavukcuoglu, 2015, ILCR. "Multiple Object Recognition With Visual Attention"
REINFORCE
16. RECURRENT ATTENTION MODEL 16
▸ Experiments & Results
▸ MNIST, SVHN
Jimmy Lei Ba, Volodymyr Mnih, Koray Kavukcuoglu, 2015, ILCR. "Multiple Object Recognition With Visual Attention"
17. SPATIAL TRANSFORMER NETWORK 17
▸ Max Jaderberg, Karen Simonyan, Andrew Zisserman, Koray Kavukcuoglu, 2014
NIPS. "Spatial Transformer Network"
▸ Google DeepMind, 624 citations
▸ Motivation: Human process distorted objects by un-distorting it
▸ ConvNet is not actually invariant to large transformation(only realised over a
deep hierarchy of max-pooling)
Max Jaderberg, Karen Simonyan, Andrew Zisserman, Koray Kavukcuoglu, 2014, NIPS. "Spatial Transformer Network"
https://kevinzakka.github.io/2017/01/18/stn-part2/
18. SPATIAL TRANSFORMER NETWORK 18
▸ Architecture:
▸ three parts: localisation net, sampling grid, sampler
▸ Assume 𝛵𝜃 is 2D affine transformation A𝜃,
Max Jaderberg, Karen Simonyan, Andrew Zisserman, Koray Kavukcuoglu, 2014, NIPS. "Spatial Transformer Network"
regression
H,W,C H',W',C
19. SPATIAL TRANSFORMER NETWORK 19
▸ 𝛵𝜃, for attention becomes:
▸ Allowing cropping, translation, isotropic scaling
▸ In case if a bilinear sampling kernel,
▸ Differentiable, Modular,
Max Jaderberg, Karen Simonyan, Andrew Zisserman, Koray Kavukcuoglu, 2014, NIPS. "Spatial Transformer Network"
20. SPATIAL TRANSFORMER NETWORK 20
▸ Experiments and Results
▸ MNIST
▸ SVHN
Max Jaderberg, Karen Simonyan, Andrew Zisserman, Koray Kavukcuoglu, 2014, NIPS. "Spatial Transformer Network"
21. SPATIAL TRANSFORMER NETWORK 21
▸ Experiments and Results
▸ Fine-grained classification (CUB-200-211 bird classification dataset)
Max Jaderberg, Karen Simonyan, Andrew Zisserman, Koray Kavukcuoglu, 2014, NIPS. "Spatial Transformer Network"
22. SPATIAL TRANSFORMER NETWORK 22
▸ Already implemented in Tensorlayer
Max Jaderberg, Karen Simonyan, Andrew Zisserman, Koray Kavukcuoglu, 2014, NIPS. "Spatial Transformer Network"
23. RECURRENT ATTENTIONAL NETWORKS FOR SALIENCY DETECTION 23
▸ Jason Kuen, Zhenhua Wang, Gang Wang, 2016, CVPR. "Recurrent Attentional
Networks for Saliency Detection"
▸ RAM(Glimpse system) + STN(Differentiability) for Saliency Detection
Jason Kuen, Zhenhua Wang, Gang Wang, 2016, CVPR. "Recurrent Attentional Networks for Saliency Detection"
24. RECURRENT ATTENTIONAL NETWORKS FOR SALIENCY DETECTION 24
▸ Recurrent Attentional Convolutional-Deconvolutional Network (RACDNN)
▸ Architecture
Jason Kuen, Zhenhua Wang, Gang Wang, 2016, CVPR. "Recurrent Attentional Networks for Saliency Detection"
25. RECURRENT ATTENTIONAL NETWORKS FOR SALIENCY DETECTION 25
▸ Experiments & Results
Jason Kuen, Zhenhua Wang, Gang Wang, 2016, CVPR. "Recurrent Attentional Networks for Saliency Detection"
27. GENERATIVE IMAGE INPAINTING WITH CONTEXTUAL ATTENTION 27
▸ Jiahui Yu, Zhe Lin, Jimei Yang, Xiaohui Shen, Xin Lu, Thomas S. Huang, 2018, CVPR.
"Generative Image Inpainting with Contextual Attention"
▸ Adobe Research
Jiahui Yu, Zhe Lin, Jimei Yang, Xiaohui Shen, Xin Lu, Thomas S. Huang, 2018, CVPR. "Generative Image Inpainting with Contextual Attention
28. GENERATIVE IMAGE INPAINTING WITH CONTEXTUAL ATTENTION 28
▸ Architecture
▸ Two-stage(coarse to fine)
▸ Global and Local W-GANS
▸ Spatially discounted reconstruction loss(𝑙1): 𝛾
Jiahui Yu, Zhe Lin, Jimei Yang, Xiaohui Shen, Xin Lu, Thomas S. Huang, 2018, CVPR. "Generative Image Inpainting with Contextual Attention
USE W-GAN
attention
𝑙
29. GENERATIVE IMAGE INPAINTING WITH CONTEXTUAL ATTENTION 29
▸ Attention
Jiahui Yu, Zhe Lin, Jimei Yang, Xiaohui Shen, Xin Lu, Thomas S. Huang, 2018, CVPR. "Generative Image Inpainting with Contextual Attention
fx,y
bx,y
Calculate cosine similarity:
30. GENERATIVE IMAGE INPAINTING WITH CONTEXTUAL ATTENTION 30
▸ Experiments & Results
Jiahui Yu, Zhe Lin, Jimei Yang, Xiaohui Shen, Xin Lu, Thomas S. Huang, 2018, CVPR. "Generative Image Inpainting with Contextual Attention
31. LEARN TO PAY ATTENTION 31
▸ Saumya Jetley, Nicholas A. Lord, Namhoon Lee, Philip H. S. Torr, 2018, ICLR. "Learn
to Pay Attention"
▸ Very simple
Saumya Jetley, Nicholas A. Lord, Namhoon Lee, Philip H. S. Torr, 2018, ICLR. "Learn to Pay Attention"
32. LEARN TO PAY ATTENTION 32
▸ Architecture
Saumya Jetley, Nicholas A. Lord, Namhoon Lee, Philip H. S. Torr, 2018, ICLR. "Learn to Pay Attention"
Attention
Compatibility
function(dot
product)
33. LEARN TO PAY ATTENTION 33
▸ Experiments & Results
▸ Image classification and fine-grained recognition
Saumya Jetley, Nicholas A. Lord, Namhoon Lee, Philip H. S. Torr, 2018, ICLR. "Learn to Pay Attention"
34. LEARN TO PAY ATTENTION 34
▸ Experiments & Results
▸ Weakly supervised semantic segmentation
Saumya Jetley, Nicholas A. Lord, Namhoon Lee, Philip H. S. Torr, 2018, ICLR. "Learn to Pay Attention"
35. LOOK CLOSER TO SEE BETTER 35
▸ Jianlong Fu, Heliang Zheng, Tao Mei, 2017, CVPR. "Look Closer to See Better:
Recurrent Attention Convolutional Neural Network for Fine-grained Image
Recognition"
▸ Fine-grained image recognition:
▸ Discriminative region localization + fine-grained feature learning
Jianlong Fu, Heliang Zheng, Tao Mei, 2017, CVPR. "Look Closer to See Better: Recurrent Attention Convolutional Neural Network for Fine-
grained Image Recognition"
36. LOOK CLOSER TO SEE BETTER 36
▸ Recurrent Attention Convolutional Neural Network (RA-CNN)
▸ Multi-scale networks: classification sub-network, attention proposal sub-
network(APN)
▸ Finer-scale network (coarse to fine)
▸ Intra-scale softmax loss for classification, inter-scale pairwise ranking loss for
APN
Jianlong Fu, Heliang Zheng, Tao Mei, 2017, CVPR. "Look Closer to See Better: Recurrent Attention Convolutional Neural Network for Fine-
grained Image Recognition"
37. LOOK CLOSER TO SEE BETTER 37
▸ RA-CNN architecture:
Jianlong Fu, Heliang Zheng, Tao Mei, 2017, CVPR. "Look Closer to See Better: Recurrent Attention Convolutional Neural Network for Fine-
grained Image Recognition"
bilinear
interpolation
to amplify
38. LOOK CLOSER TO SEE BETTER 38
▸ Training:
▸ Multi-task loss:
Jianlong Fu, Heliang Zheng, Tao Mei, 2017, CVPR. "Look Closer to See Better: Recurrent Attention Convolutional Neural Network for Fine-
grained Image Recognition"
forces
39. LOOK CLOSER TO SEE BETTER 39
▸ Experiments & Results
▸ CUB-200-211 Bird Dataset
Jianlong Fu, Heliang Zheng, Tao Mei, 2017, CVPR. "Look Closer to See Better: Recurrent Attention Convolutional Neural Network for Fine-
grained Image Recognition"
40. LOOK CLOSER TO SEE BETTER 40
▸ Experiments & Results
▸ Stanford Dogs, Stanford Cars
Jianlong Fu, Heliang Zheng, Tao Mei, 2017, CVPR. "Look Closer to See Better: Recurrent Attention Convolutional Neural Network for Fine-
grained Image Recognition"
41. SUMMARY 41
▸ Attention for efficiency, better performance, interpretability
▸ Many types of Attention:
▸ RAM
▸ STN
▸ RAM+STN
▸ Others