2. ICCV
- ICCV : International Conference on Computer Vision
- 2019년 국내에서 첫 개최
: 10.27 ~ 11.2, COEX
- 2019 ICCV
: 약 7,500명 참가
총 1,077편의 논문
Yonsei University Severance Hospital CCIDS
3. ICCV
ICCV 2019 Summary
1. Visual Recognition for Medical Images
2. Image and Video Synthesis: How, Why and What if?
3. Interpretating and Explaining Visual AI Models
4. SinGAN: Learning a Generative Model from a Single Natural Images (Best Paper)
Yonsei University Severance Hospital CCIDS
4. Yonsei University Severance Hospital CCIDS
Visual Recognition
for Medical Images
https://sites.google.com/view/iccv19-vrmi
5. Visual Recognition for Medical Images
Yonsei University Severance Hospital CCIDS
Deep Learning in Neuroimaging and
Radiotherapy
Dinggang Shen
University of North Carolina at Chapel Hill
6. Visual Recognition for Medical Images
Yonsei University Severance Hospital CCIDS
Image Registration
Template T Subject S
Deep-learning based
Registration
J. Fan, X. Cao, Z. Xue, PT. Yap, D. Shen, “Adversarial Similarity Network for Evaluating Image Alignment in Deep Learning based Registration”
7. Visual Recognition for Medical Images
Yonsei University Severance Hospital CCIDS
Train the registration network by Similarity Metric
Template T
Subject S
Deep Registration
Network (U-Net)
T
S
68 * 68 * 68
68 * 68 * 68
Deformation
Field
3*28*28*28
Spatial
Transformation Warped
subject
28*28*28
Dissimilarity
Loss LD(T, S')
S'φ
J. Fan, X. Cao, Z. Xue, PT. Yap, D. Shen, “Adversarial Similarity Network for Evaluating Image Alignment in Deep Learning based Registration”
8. Visual Recognition for Medical Images
Yonsei University Severance Hospital CCIDS
Train the registration network by Similarity Metric (inspired by GAN)
Template T
Subject S
Deep Registration
Network (U-Net)
T
S
68 * 68 * 68
68 * 68 * 68
Deformation
Field
3*28*28*28
Spatial
Transformation Warped
subject
28*28*28
Discriminator
S'φ
J. Fan, X. Cao, Z. Xue, PT. Yap, D. Shen, “Adversarial Similarity Network for Evaluating Image Alignment in Deep Learning based Registration”
Generator
9. Visual Recognition for Medical Images
Yonsei University Severance Hospital CCIDS
On Unseen Testing Dataset
J. Fan, X. Cao, Z. Xue, PT. Yap, D. Shen, “Adversarial Similarity Network for Evaluating Image Alignment in Deep Learning based Registration”
10. Visual Recognition for Medical Images
Yonsei University Severance Hospital CCIDS
Registration Result
J. Fan, X. Cao, Z. Xue, PT. Yap, D. Shen, “Adversarial Similarity Network for Evaluating Image Alignment in Deep Learning based Registration”
11. Visual Recognition for Medical Images
Yonsei University Severance Hospital CCIDS
Image Synthesis
MRI CT
Can we synthesize CT image from
a single MRI Image?
CT images are highly desired for
- Specific diagnosis
- Dose planning
- PET attenuation correction
D. Nie et al,
“Medical Image Synthesis with Deep Convolutional Adversarial Networks”
12. Visual Recognition for Medical Images
Yonsei University Severance Hospital CCIDS
D. Nie et al,
“Medical Image Synthesis with Deep Convolutional Adversarial Networks”
Auto-context
Use context features
to iteratively refine
the training results
13. Visual Recognition for Medical Images
Yonsei University Severance Hospital CCIDS
D. Nie et al,
“Medical Image Synthesis with Deep Convolutional Adversarial Networks”
Auto-context Refinements
Difference maps
14. Visual Recognition for Medical Images
Yonsei University Severance Hospital CCIDS
Q&A
Q. Whole Image가 아니라 Patch를 사용하는 이유?
A. 데이터 수가 딥러닝을 학습하기에는 부족하기 때문이다. Patch를 사용하면 훨씬 많은 학습 데이터를
확보할 수 있기 때문에 patch를 사용한다.
물론 데이터 수가 충분히 많다면 patch가 아니라 whole image를 사용하는 것이 더욱 좋다.
15. Yonsei University Severance Hospital CCIDS
Image and Video Synthesis:
How, Why and What if
https://sites.google.com/berkeley.edu/iccv-2019-image-and-video-syn
16. Image and Video Synthesis: How, Why and What if?
Yonsei University Severance Hospital CCIDS
Seeing What
a GAN Cannot Generate
David Bau
MIT
http://ganseeing.csail.mit.edu//
17. Seeing What a GAN Cannot Generate
Yonsei University Severance Hospital CCIDS
https://software.intel.com/en-us/blogs/2017/08/21/mode-collapse-in-gans
Model Collapse in GAN is serious Problem
※ Model Collapse : A problem when all the generator outputs are identical
(all of them or most of the samples are equal)
18. Seeing What a GAN Cannot Generate
Yonsei University Severance Hospital CCIDS
Fake Images
Real Images
http://ganseeing.csail.mit.edu//
19. Seeing What a GAN Cannot Generate
Yonsei University Severance Hospital CCIDS
Fake Images
Real Images
Inception
Inception
Fake Inception
Feature Space
Real Inception
Feature Space
http://ganseeing.csail.mit.edu//
FID (Frechet Inception Distance)
Measuring
GAN Quality
20. Seeing What a GAN Cannot Generate
Yonsei University Severance Hospital CCIDS
Fake Images
Real Images
http://ganseeing.csail.mit.edu//
1. What is actually missing
in the distribution?
2. What is actually missing
in each image?
21. Seeing What a GAN Cannot Generate
Yonsei University Severance Hospital CCIDS
http://ganseeing.csail.mit.edu//
1. Understanding Omissions in the Distribution
Generated Image Semantic segmentation
Real Image Semantic segmentation
22. Seeing What a GAN Cannot Generate
Yonsei University Severance Hospital CCIDS
http://ganseeing.csail.mit.edu//
1. Understanding Omissions in the Distribution
Generated Image Semantic segmentation
Real Image Semantic segmentation
23. Seeing What a GAN Cannot Generate
Yonsei University Severance Hospital CCIDS
http://ganseeing.csail.mit.edu//
2. Understanding Omissions in Individual Images
Real Image x Synthesized Image G(z)
24. Seeing What a GAN Cannot Generate
Yonsei University Severance Hospital CCIDS
http://ganseeing.csail.mit.edu//
2. Understanding Omissions in Individual Images
Real Image x Synthesized Image G(z)
Pairs (x, G(z*)) reveals omissions
Objective : z* = argminz Loss(x, G(z))
25. Seeing What a GAN Cannot Generate
Yonsei University Severance Hospital CCIDS
http://ganseeing.csail.mit.edu//
Two steps to invert a large generator
26. Seeing What a GAN Cannot Generate
Yonsei University Severance Hospital CCIDS
http://ganseeing.csail.mit.edu//
x = G(z) x = G(z*) x = real x = G(z*)
Generated Reconstruction Real Photo Reconstruction
When G generates x,
reconstruction is precise
When reconstruction is imperfect
we know G cannot generate x
27. Seeing What a GAN Cannot Generate
Yonsei University Severance Hospital CCIDS
http://ganseeing.csail.mit.edu//
GANs don’t like people
28. Seeing What a GAN Cannot Generate
Yonsei University Severance Hospital CCIDS
http://ganseeing.csail.mit.edu//
The Cheese Hypothesis
Original Image
29. Seeing What a GAN Cannot Generate
Yonsei University Severance Hospital CCIDS
http://ganseeing.csail.mit.edu//
The Cheese Hypothesis
Original Image
Optimized z
30. Seeing What a GAN Cannot Generate
Yonsei University Severance Hospital CCIDS
http://ganseeing.csail.mit.edu//
The Cheese Hypothesis
Original Image
Adapted Cheese
31. Seeing What a GAN Cannot Generate
Yonsei University Severance Hospital CCIDS
Real Image x
32. Seeing What a GAN Cannot Generate
Yonsei University Severance Hospital CCIDS
Optimized
vector
z*
G
Real Image x Reconstructed Image G(z*)
z* = argminz Loss(x, G(z))
33. Seeing What a GAN Cannot Generate
Yonsei University Severance Hospital CCIDS
Optimized
vector
z*
G
Real Image x Reconstructed Image G(z*, θ*)
θ*
z*, θ* = argminz,θ Loss(x, G(z)) + R(θ)
Regularizer
Inspired by Deep Image Prior [Ulyanove et al, 2018]
34. Yonsei University Severance Hospital CCIDS
Interpretating and Explaining
Visual AI Models
http://xai.unist.ac.kr/workshop/2019/
35. Interpretating and Explaining Visual AI Models
Yonsei University Severance Hospital CCIDS
Recent progress towards XAI
at UC Berkeley
Trevor Darrel
UC Berkeley
37. Recent progress towards XAI at UC Berkeley
Yonsei University Severance Hospital CCIDS
Explainable NN
38. Recent progress towards XAI at UC Berkeley
Yonsei University Severance Hospital CCIDS
Salience for Introspection
Image LIME GradCAM RISE
V Petsiuk et al. RISE: Randomized Input Sampling for Explanation of Black-box Models
39. Recent progress towards XAI at UC Berkeley
Yonsei University Severance Hospital CCIDS
Overview of RISE
V Petsiuk et al. RISE: Randomized Input Sampling for Explanation of Black-box Models
40. Recent progress towards XAI at UC Berkeley
Yonsei University Severance Hospital CCIDS
Pros of RISE
V Petsiuk et al. RISE: Randomized Input Sampling for Explanation of Black-box Models
Cons of RISE
A more general framework
as the importance map is obtained with
access to only the input and
output of the base model
Time and resource complexity
41. Recent progress towards XAI at UC Berkeley
Yonsei University Severance Hospital CCIDS
LRP : Layer-wise Relevance Propagation
Explaining Decisions of Neural Networks by LRP. Alexander Binder
@ Deep Learning: Theory, Algorithms, and Applications. Berlin, June 2017
42. Recent progress towards XAI at UC Berkeley
Yonsei University Severance Hospital CCIDS
LRP : Layer-wise Relevance Propagation
Explaining Decisions of Neural Networks by LRP. Alexander Binder
@ Deep Learning: Theory, Algorithms, and Applications. Berlin, June 2017
43. Recent progress towards XAI at UC Berkeley
Yonsei University Severance Hospital CCIDS
LRP : Layer-wise Relevance Propagation
Explaining Decisions of Neural Networks by LRP. Alexander Binder
@ Deep Learning: Theory, Algorithms, and Applications. Berlin, June 2017
44. Recent progress towards XAI at UC Berkeley
Yonsei University Severance Hospital CCIDS
Unmasking Clever Hans Predictors
S Lapuschkin et al. Unmasking Clever Hans Predictors and Assessing What Machines Really Learn
Images with a copyright watermark (a, c) are classified to “horse” class,
but images without an added copyright watermark (b, d) aren’t.
45. Recent progress towards XAI at UC Berkeley
Yonsei University Severance Hospital CCIDS
Understanding Learning Behaviour
S Lapuschkin et al. Unmasking Clever Hans Predictors and Assessing What Machines Really Learn
Model learns
1. Track the ball
2. Focus on paddle
3. Focus in the tunnel
46. Yonsei University Severance Hospital CCIDS
Learning a Generative Model
from a Single Natural Image
https://sites.google.com/berkeley.edu/iccv-2019-image-and-video-syn
SinGAN
Tamar Rott Shaham
Google AI
ICCV 2019 Best Paper
50. SinGAN
Yonsei University Severance Hospital CCIDS
▶ Same model
▶ No extra training
SinGAN for single image animation
https://www.youtube.com/watch?v=xk8bWLZk4DU&feature=youtu.be