sourcecycle-consistent adversarial networks, iccv 2017. ... (e.g., dog to cat) • can get confused...
TRANSCRIPT
![Page 1: SourceCycle-Consistent Adversarial Networks, ICCV 2017. ... (e.g., dog to cat) • Can get confused on images outside of the training domains (e.g., horse with rider) • Cannot close](https://reader034.vdocuments.site/reader034/viewer/2022052008/601cb25298f83944e7188228/html5/thumbnails/1.jpg)
Conditional generation
Source
Source
![Page 2: SourceCycle-Consistent Adversarial Networks, ICCV 2017. ... (e.g., dog to cat) • Can get confused on images outside of the training domains (e.g., horse with rider) • Cannot close](https://reader034.vdocuments.site/reader034/viewer/2022052008/601cb25298f83944e7188228/html5/thumbnails/2.jpg)
Outline• General conditional GANs• BigGAN• Paired image-to-image translation• Unpaired image-to-image translation:
CycleGAN
![Page 3: SourceCycle-Consistent Adversarial Networks, ICCV 2017. ... (e.g., dog to cat) • Can get confused on images outside of the training domains (e.g., horse with rider) • Cannot close](https://reader034.vdocuments.site/reader034/viewer/2022052008/601cb25298f83944e7188228/html5/thumbnails/3.jpg)
Conditional generation• Suppose we want to condition the generation
of samples on discrete side information (label) !• How do we add ! to the basic GAN framework?
" #"(%)#("(%))%
![Page 4: SourceCycle-Consistent Adversarial Networks, ICCV 2017. ... (e.g., dog to cat) • Can get confused on images outside of the training domains (e.g., horse with rider) • Cannot close](https://reader034.vdocuments.site/reader034/viewer/2022052008/601cb25298f83944e7188228/html5/thumbnails/4.jpg)
Conditional generation• Suppose we want to condition the generation
of samples on discrete side information (label) !• How do we add ! to the basic GAN framework?
"# #(", !)
'(# ", ! , !)
! !'
![Page 5: SourceCycle-Consistent Adversarial Networks, ICCV 2017. ... (e.g., dog to cat) • Can get confused on images outside of the training domains (e.g., horse with rider) • Cannot close](https://reader034.vdocuments.site/reader034/viewer/2022052008/601cb25298f83944e7188228/html5/thumbnails/5.jpg)
Conditional generation• Example: simple network for generating
28 x 28 MNIST digits
M. Mirza and S. Osindero, Conditional Generative Adversarial Nets, arXiv 2014
Class label: 10-dim one-hot
vector
Noise vector
Generator
Figure source: F. Fleuret
![Page 6: SourceCycle-Consistent Adversarial Networks, ICCV 2017. ... (e.g., dog to cat) • Can get confused on images outside of the training domains (e.g., horse with rider) • Cannot close](https://reader034.vdocuments.site/reader034/viewer/2022052008/601cb25298f83944e7188228/html5/thumbnails/6.jpg)
Conditional generation• Example: simple network for generating
28 x 28 MNIST digits
M. Mirza and S. Osindero, Conditional Generative Adversarial Nets, arXiv 2014
Class label: 10-dim one-hot
vector
Noise vector
Discriminator
Figure source: F. Fleuret
![Page 7: SourceCycle-Consistent Adversarial Networks, ICCV 2017. ... (e.g., dog to cat) • Can get confused on images outside of the training domains (e.g., horse with rider) • Cannot close](https://reader034.vdocuments.site/reader034/viewer/2022052008/601cb25298f83944e7188228/html5/thumbnails/7.jpg)
Conditional generation• Example: simple network for generating
28 x 28 MNIST digits
M. Mirza and S. Osindero, Conditional Generative Adversarial Nets, arXiv 2014
![Page 8: SourceCycle-Consistent Adversarial Networks, ICCV 2017. ... (e.g., dog to cat) • Can get confused on images outside of the training domains (e.g., horse with rider) • Cannot close](https://reader034.vdocuments.site/reader034/viewer/2022052008/601cb25298f83944e7188228/html5/thumbnails/8.jpg)
Conditional generation• Another example: text-to-image synthesis
S. Reed, Z. Akata, X. Yan, L. Logeswaran, B. Schiele, H. Lee, Generative adversarial text to image synthesis, ICML 2016
LSTM text encoder Spatial replication, depth concatenation
![Page 9: SourceCycle-Consistent Adversarial Networks, ICCV 2017. ... (e.g., dog to cat) • Can get confused on images outside of the training domains (e.g., horse with rider) • Cannot close](https://reader034.vdocuments.site/reader034/viewer/2022052008/601cb25298f83944e7188228/html5/thumbnails/9.jpg)
Conditional generation• Another example: text-to-image synthesis
S. Reed, Z. Akata, X. Yan, L. Logeswaran, B. Schiele, H. Lee, Generative adversarial text to image synthesis, ICML 2016
Previously unseen captions (zero-shot
setting)
Captions seen in the training set
![Page 10: SourceCycle-Consistent Adversarial Networks, ICCV 2017. ... (e.g., dog to cat) • Can get confused on images outside of the training domains (e.g., horse with rider) • Cannot close](https://reader034.vdocuments.site/reader034/viewer/2022052008/601cb25298f83944e7188228/html5/thumbnails/10.jpg)
BigGAN• Synthesize ImageNet images, conditioned on
class label, up to 512 x 512 resolution
A. Brock, J. Donahue, K. Simonyan, Large scale GAN training for high fidelity natural image synthesis, arXiv 2018
![Page 12: SourceCycle-Consistent Adversarial Networks, ICCV 2017. ... (e.g., dog to cat) • Can get confused on images outside of the training domains (e.g., horse with rider) • Cannot close](https://reader034.vdocuments.site/reader034/viewer/2022052008/601cb25298f83944e7188228/html5/thumbnails/12.jpg)
BigGAN• Self-attention GAN to capture spatial structure• Spectral normalization for generator and discriminator• Conditioning the generator: conditional batch norm• Conditioning the discriminator: projection• 8x larger batch size, 50% more feature channels than
baseline• Hierarchical latent space: feed different “chunks” of noise
vector into multiple layers of the generator• Lots of other tricks (initialization, training, z sampling, etc.)• Training observed to be unstable, but good results are
achieved before collapse
![Page 13: SourceCycle-Consistent Adversarial Networks, ICCV 2017. ... (e.g., dog to cat) • Can get confused on images outside of the training domains (e.g., horse with rider) • Cannot close](https://reader034.vdocuments.site/reader034/viewer/2022052008/601cb25298f83944e7188228/html5/thumbnails/13.jpg)
BigGAN• Results at 512 x 512 resolution
![Page 14: SourceCycle-Consistent Adversarial Networks, ICCV 2017. ... (e.g., dog to cat) • Can get confused on images outside of the training domains (e.g., horse with rider) • Cannot close](https://reader034.vdocuments.site/reader034/viewer/2022052008/601cb25298f83944e7188228/html5/thumbnails/14.jpg)
BigGAN• Results at 512 x 512 resolution
![Page 15: SourceCycle-Consistent Adversarial Networks, ICCV 2017. ... (e.g., dog to cat) • Can get confused on images outside of the training domains (e.g., horse with rider) • Cannot close](https://reader034.vdocuments.site/reader034/viewer/2022052008/601cb25298f83944e7188228/html5/thumbnails/15.jpg)
BigGAN
Nearest neighbors in pixel space
Nearest neighbors in
FC7 space
![Page 16: SourceCycle-Consistent Adversarial Networks, ICCV 2017. ... (e.g., dog to cat) • Can get confused on images outside of the training domains (e.g., horse with rider) • Cannot close](https://reader034.vdocuments.site/reader034/viewer/2022052008/601cb25298f83944e7188228/html5/thumbnails/16.jpg)
BigGAN• Results at 512 x 512 resolution
Easy classes Difficult classes
![Page 17: SourceCycle-Consistent Adversarial Networks, ICCV 2017. ... (e.g., dog to cat) • Can get confused on images outside of the training domains (e.g., horse with rider) • Cannot close](https://reader034.vdocuments.site/reader034/viewer/2022052008/601cb25298f83944e7188228/html5/thumbnails/17.jpg)
Image-to-image translation
P. Isola, J.-Y. Zhu, T. Zhou, A. Efros, Image-to-Image Translation with Conditional Adversarial Networks, CVPR 2017
![Page 18: SourceCycle-Consistent Adversarial Networks, ICCV 2017. ... (e.g., dog to cat) • Can get confused on images outside of the training domains (e.g., horse with rider) • Cannot close](https://reader034.vdocuments.site/reader034/viewer/2022052008/601cb25298f83944e7188228/html5/thumbnails/18.jpg)
Image-to-image translation• Produce modified image ! conditioned on
input image " (note change of notation)• Generator receives ! as input• Discriminator receives an !, # pair and has to
decide whether it is real or fake
![Page 19: SourceCycle-Consistent Adversarial Networks, ICCV 2017. ... (e.g., dog to cat) • Can get confused on images outside of the training domains (e.g., horse with rider) • Cannot close](https://reader034.vdocuments.site/reader034/viewer/2022052008/601cb25298f83944e7188228/html5/thumbnails/19.jpg)
Image-to-image translation• Generator architecture: U-Net
• Note: no ! used as input, transformation is basically deterministic
![Page 20: SourceCycle-Consistent Adversarial Networks, ICCV 2017. ... (e.g., dog to cat) • Can get confused on images outside of the training domains (e.g., horse with rider) • Cannot close](https://reader034.vdocuments.site/reader034/viewer/2022052008/601cb25298f83944e7188228/html5/thumbnails/20.jpg)
Image-to-image translation• Generator architecture: U-Net
Figure source
Encode: convolution → BatchNorm → ReLU
Decode: transposed convolution → BatchNorm → ReLU
![Page 21: SourceCycle-Consistent Adversarial Networks, ICCV 2017. ... (e.g., dog to cat) • Can get confused on images outside of the training domains (e.g., horse with rider) • Cannot close](https://reader034.vdocuments.site/reader034/viewer/2022052008/601cb25298f83944e7188228/html5/thumbnails/21.jpg)
Image-to-image translation• Generator architecture: U-Net
Effect of adding skip connections to the generator
![Page 22: SourceCycle-Consistent Adversarial Networks, ICCV 2017. ... (e.g., dog to cat) • Can get confused on images outside of the training domains (e.g., horse with rider) • Cannot close](https://reader034.vdocuments.site/reader034/viewer/2022052008/601cb25298f83944e7188228/html5/thumbnails/22.jpg)
Image-to-image translation• Generator loss: GAN loss plus L1 reconstruction
penalty
!∗ = argmin"max#ℒ"$% !,, + . /&
0& − !(3&) '
Generated output ((*&) should be close to
ground truth target ,&
![Page 23: SourceCycle-Consistent Adversarial Networks, ICCV 2017. ... (e.g., dog to cat) • Can get confused on images outside of the training domains (e.g., horse with rider) • Cannot close](https://reader034.vdocuments.site/reader034/viewer/2022052008/601cb25298f83944e7188228/html5/thumbnails/23.jpg)
Image-to-image translation• Generator loss: GAN loss plus L1 reconstruction
penalty
!∗ = argmin"max#ℒ"$% !,, + . /&
0& − !(3&) '
![Page 24: SourceCycle-Consistent Adversarial Networks, ICCV 2017. ... (e.g., dog to cat) • Can get confused on images outside of the training domains (e.g., horse with rider) • Cannot close](https://reader034.vdocuments.site/reader034/viewer/2022052008/601cb25298f83944e7188228/html5/thumbnails/24.jpg)
Image-to-image translation• Discriminator: PatchGAN
• Given input image ! and second image ", decide whether " is a ground truth target or produced by the generator
![Page 25: SourceCycle-Consistent Adversarial Networks, ICCV 2017. ... (e.g., dog to cat) • Can get confused on images outside of the training domains (e.g., horse with rider) • Cannot close](https://reader034.vdocuments.site/reader034/viewer/2022052008/601cb25298f83944e7188228/html5/thumbnails/25.jpg)
Image-to-image translation• Discriminator: PatchGAN
• Given input image ! and second image ", decide whether " is a ground truth target or produced by the generator
• Output is a 30 x 30 map where each value (0 to 1) represents the quality of the corresponding section of the output image
• Fully convolutional network, effective patch size canbe increased by increasing the depth
Figure source
![Page 26: SourceCycle-Consistent Adversarial Networks, ICCV 2017. ... (e.g., dog to cat) • Can get confused on images outside of the training domains (e.g., horse with rider) • Cannot close](https://reader034.vdocuments.site/reader034/viewer/2022052008/601cb25298f83944e7188228/html5/thumbnails/26.jpg)
Image-to-image translation• Discriminator: PatchGAN
• Given input image ! and second image ", decide whether " is a ground truth target or produced by the generator
• Output is a 30 x 30 map where each value (0 to 1) represents the quality of the corresponding section of the output image
• Fully convolutional network, effective patch size can be increased by increasing the depth
Effect of discriminator patch size on generator output
![Page 27: SourceCycle-Consistent Adversarial Networks, ICCV 2017. ... (e.g., dog to cat) • Can get confused on images outside of the training domains (e.g., horse with rider) • Cannot close](https://reader034.vdocuments.site/reader034/viewer/2022052008/601cb25298f83944e7188228/html5/thumbnails/27.jpg)
Image-to-image translation: Results• Translating between maps and aerial photos
![Page 28: SourceCycle-Consistent Adversarial Networks, ICCV 2017. ... (e.g., dog to cat) • Can get confused on images outside of the training domains (e.g., horse with rider) • Cannot close](https://reader034.vdocuments.site/reader034/viewer/2022052008/601cb25298f83944e7188228/html5/thumbnails/28.jpg)
Image-to-image translation: Results• Translating between maps and aerial photos• Human study:
![Page 29: SourceCycle-Consistent Adversarial Networks, ICCV 2017. ... (e.g., dog to cat) • Can get confused on images outside of the training domains (e.g., horse with rider) • Cannot close](https://reader034.vdocuments.site/reader034/viewer/2022052008/601cb25298f83944e7188228/html5/thumbnails/29.jpg)
Image-to-image translation: Results• Semantic labels to scenes
![Page 30: SourceCycle-Consistent Adversarial Networks, ICCV 2017. ... (e.g., dog to cat) • Can get confused on images outside of the training domains (e.g., horse with rider) • Cannot close](https://reader034.vdocuments.site/reader034/viewer/2022052008/601cb25298f83944e7188228/html5/thumbnails/30.jpg)
Image-to-image translation: Results• Semantic labels to scenes• Evaluation: FCN score
• The higher the quality of the output, the better theFCN should do at recovering the original semantic labels
![Page 31: SourceCycle-Consistent Adversarial Networks, ICCV 2017. ... (e.g., dog to cat) • Can get confused on images outside of the training domains (e.g., horse with rider) • Cannot close](https://reader034.vdocuments.site/reader034/viewer/2022052008/601cb25298f83944e7188228/html5/thumbnails/31.jpg)
Image-to-image translation: Results• Scenes to semantic labels
![Page 32: SourceCycle-Consistent Adversarial Networks, ICCV 2017. ... (e.g., dog to cat) • Can get confused on images outside of the training domains (e.g., horse with rider) • Cannot close](https://reader034.vdocuments.site/reader034/viewer/2022052008/601cb25298f83944e7188228/html5/thumbnails/32.jpg)
Image-to-image translation: Results• Scenes to semantic labels• Accuracy is worse than that of regular FCNs
or generator with L1 loss
![Page 33: SourceCycle-Consistent Adversarial Networks, ICCV 2017. ... (e.g., dog to cat) • Can get confused on images outside of the training domains (e.g., horse with rider) • Cannot close](https://reader034.vdocuments.site/reader034/viewer/2022052008/601cb25298f83944e7188228/html5/thumbnails/33.jpg)
Image-to-image translation: Results• Semantic labels to facades
![Page 34: SourceCycle-Consistent Adversarial Networks, ICCV 2017. ... (e.g., dog to cat) • Can get confused on images outside of the training domains (e.g., horse with rider) • Cannot close](https://reader034.vdocuments.site/reader034/viewer/2022052008/601cb25298f83944e7188228/html5/thumbnails/34.jpg)
Image-to-image translation: Results• Day to night
![Page 35: SourceCycle-Consistent Adversarial Networks, ICCV 2017. ... (e.g., dog to cat) • Can get confused on images outside of the training domains (e.g., horse with rider) • Cannot close](https://reader034.vdocuments.site/reader034/viewer/2022052008/601cb25298f83944e7188228/html5/thumbnails/35.jpg)
Image-to-image translation: Results• Edges to photos
![Page 37: SourceCycle-Consistent Adversarial Networks, ICCV 2017. ... (e.g., dog to cat) • Can get confused on images outside of the training domains (e.g., horse with rider) • Cannot close](https://reader034.vdocuments.site/reader034/viewer/2022052008/601cb25298f83944e7188228/html5/thumbnails/37.jpg)
Image-to-image translation: Limitations• Visual quality could be improved• Requires !, # pairs for training• Does not model conditional distribution $(#|!), returns a single mode instead
![Page 38: SourceCycle-Consistent Adversarial Networks, ICCV 2017. ... (e.g., dog to cat) • Can get confused on images outside of the training domains (e.g., horse with rider) • Cannot close](https://reader034.vdocuments.site/reader034/viewer/2022052008/601cb25298f83944e7188228/html5/thumbnails/38.jpg)
Unpaired image-to-image translation• Given two unordered image collections ! and ",
learn to “translate” an image from one into the other and vice versa
J.-Y. Zhu, T. Park, P. Isola, A. Efros, Unpaired Image-to-Image Translation Using Cycle-Consistent Adversarial Networks, ICCV 2017
![Page 39: SourceCycle-Consistent Adversarial Networks, ICCV 2017. ... (e.g., dog to cat) • Can get confused on images outside of the training domains (e.g., horse with rider) • Cannot close](https://reader034.vdocuments.site/reader034/viewer/2022052008/601cb25298f83944e7188228/html5/thumbnails/39.jpg)
Unpaired image-to-image translation• Given two unordered image collections ! and ",
learn to “translate” an image from one into the other and vice versa
J.-Y. Zhu, T. Park, P. Isola, A. Efros, Unpaired Image-to-Image Translation Using Cycle-Consistent Adversarial Networks, ICCV 2017
![Page 40: SourceCycle-Consistent Adversarial Networks, ICCV 2017. ... (e.g., dog to cat) • Can get confused on images outside of the training domains (e.g., horse with rider) • Cannot close](https://reader034.vdocuments.site/reader034/viewer/2022052008/601cb25298f83944e7188228/html5/thumbnails/40.jpg)
CycleGAN• Given: domains ! and "• Train two generators # and $ and two
discriminators %! and %"• # translates from $ to %, & translates from % to $• '! recognizes images from $, '" from %• We want &(#())) ≈ ) and #(&(,)) ≈ ,
![Page 41: SourceCycle-Consistent Adversarial Networks, ICCV 2017. ... (e.g., dog to cat) • Can get confused on images outside of the training domains (e.g., horse with rider) • Cannot close](https://reader034.vdocuments.site/reader034/viewer/2022052008/601cb25298f83944e7188228/html5/thumbnails/41.jpg)
CycleGAN: Architecture• Generators:
• Discriminators: PatchGAN on 70 x 70 patches
Figure source
![Page 42: SourceCycle-Consistent Adversarial Networks, ICCV 2017. ... (e.g., dog to cat) • Can get confused on images outside of the training domains (e.g., horse with rider) • Cannot close](https://reader034.vdocuments.site/reader034/viewer/2022052008/601cb25298f83944e7188228/html5/thumbnails/42.jpg)
CycleGAN: Loss• Requirements:
• ! translates from " to #, $ translates from # to "• %& recognizes images from ", %' from #• We want $(!())) ≈ ) and !($(,)) ≈ ,
• CycleGAN discriminator loss: LSGANℒ./0 %' = 23~56787(3) (%' , − 1); + 2=~56787(=) %' ! ) ;
ℒ./0 %& = 2=~56787(=) (%& ) − 1); + 23~56787(3) %& $ , ;
• CycleGAN generator loss:ℒ>?> !, $ = 2=~56787(=) %' ! ) − 1 ; + 23~56787(3) %& $ , − 1 ;
+ 2=~56787(=) $ ! ) − ) A + 23~56787(3) ! $ , − , A
![Page 43: SourceCycle-Consistent Adversarial Networks, ICCV 2017. ... (e.g., dog to cat) • Can get confused on images outside of the training domains (e.g., horse with rider) • Cannot close](https://reader034.vdocuments.site/reader034/viewer/2022052008/601cb25298f83944e7188228/html5/thumbnails/43.jpg)
CycleGAN• Illustration of cycle consistency:
![Page 44: SourceCycle-Consistent Adversarial Networks, ICCV 2017. ... (e.g., dog to cat) • Can get confused on images outside of the training domains (e.g., horse with rider) • Cannot close](https://reader034.vdocuments.site/reader034/viewer/2022052008/601cb25298f83944e7188228/html5/thumbnails/44.jpg)
CycleGAN: Results• Translation between maps and aerial photos
![Page 45: SourceCycle-Consistent Adversarial Networks, ICCV 2017. ... (e.g., dog to cat) • Can get confused on images outside of the training domains (e.g., horse with rider) • Cannot close](https://reader034.vdocuments.site/reader034/viewer/2022052008/601cb25298f83944e7188228/html5/thumbnails/45.jpg)
CycleGAN: Results• Other pix2pix tasks
![Page 46: SourceCycle-Consistent Adversarial Networks, ICCV 2017. ... (e.g., dog to cat) • Can get confused on images outside of the training domains (e.g., horse with rider) • Cannot close](https://reader034.vdocuments.site/reader034/viewer/2022052008/601cb25298f83944e7188228/html5/thumbnails/46.jpg)
CycleGAN: Results• Scene to labels and labels to scene
• Worse performance than pix2pix due to lack of paired training data
![Page 47: SourceCycle-Consistent Adversarial Networks, ICCV 2017. ... (e.g., dog to cat) • Can get confused on images outside of the training domains (e.g., horse with rider) • Cannot close](https://reader034.vdocuments.site/reader034/viewer/2022052008/601cb25298f83944e7188228/html5/thumbnails/47.jpg)
CycleGAN: Results• Tasks for which paired data is unavailable
![Page 48: SourceCycle-Consistent Adversarial Networks, ICCV 2017. ... (e.g., dog to cat) • Can get confused on images outside of the training domains (e.g., horse with rider) • Cannot close](https://reader034.vdocuments.site/reader034/viewer/2022052008/601cb25298f83944e7188228/html5/thumbnails/48.jpg)
CycleGAN: Results• Style transfer
![Page 49: SourceCycle-Consistent Adversarial Networks, ICCV 2017. ... (e.g., dog to cat) • Can get confused on images outside of the training domains (e.g., horse with rider) • Cannot close](https://reader034.vdocuments.site/reader034/viewer/2022052008/601cb25298f83944e7188228/html5/thumbnails/49.jpg)
CycleGAN: Failure cases• Failure cases
![Page 50: SourceCycle-Consistent Adversarial Networks, ICCV 2017. ... (e.g., dog to cat) • Can get confused on images outside of the training domains (e.g., horse with rider) • Cannot close](https://reader034.vdocuments.site/reader034/viewer/2022052008/601cb25298f83944e7188228/html5/thumbnails/50.jpg)
CycleGAN: Failure cases• Failure cases
![Page 51: SourceCycle-Consistent Adversarial Networks, ICCV 2017. ... (e.g., dog to cat) • Can get confused on images outside of the training domains (e.g., horse with rider) • Cannot close](https://reader034.vdocuments.site/reader034/viewer/2022052008/601cb25298f83944e7188228/html5/thumbnails/51.jpg)
CycleGAN: Limitations• Cannot handle shape changes (e.g., dog to
cat)• Can get confused on images outside of the
training domains (e.g., horse with rider)• Cannot close the gap with paired translation
methods• Does not account for the fact that one
transformation direction may be morechallenging than the other
![Page 52: SourceCycle-Consistent Adversarial Networks, ICCV 2017. ... (e.g., dog to cat) • Can get confused on images outside of the training domains (e.g., horse with rider) • Cannot close](https://reader034.vdocuments.site/reader034/viewer/2022052008/601cb25298f83944e7188228/html5/thumbnails/52.jpg)
Multimodal image-to-image translation
J.Y. Zhu, R. Zhang, D. Pathak, T. Darrell, A. A. Efros, O. Wang, E. Shechtman, Toward Multimodal Image-to-Image Translation, NIPS 2017
![Page 53: SourceCycle-Consistent Adversarial Networks, ICCV 2017. ... (e.g., dog to cat) • Can get confused on images outside of the training domains (e.g., horse with rider) • Cannot close](https://reader034.vdocuments.site/reader034/viewer/2022052008/601cb25298f83944e7188228/html5/thumbnails/53.jpg)
Unsupervised image-to-image translation
M.-Y. Liu, T. Breuel, and J. Kautz, Unsupervised Image-to-Image Translation Networks, NIPS 2017
![Page 54: SourceCycle-Consistent Adversarial Networks, ICCV 2017. ... (e.g., dog to cat) • Can get confused on images outside of the training domains (e.g., horse with rider) • Cannot close](https://reader034.vdocuments.site/reader034/viewer/2022052008/601cb25298f83944e7188228/html5/thumbnails/54.jpg)
M.-Y. Liu, T. Breuel, and J. Kautz, Unsupervised Image-to-Image Translation Networks, NIPS 2017
Unsupervised image-to-image translation
![Page 55: SourceCycle-Consistent Adversarial Networks, ICCV 2017. ... (e.g., dog to cat) • Can get confused on images outside of the training domains (e.g., horse with rider) • Cannot close](https://reader034.vdocuments.site/reader034/viewer/2022052008/601cb25298f83944e7188228/html5/thumbnails/55.jpg)
M.-Y. Liu, T. Breuel, and J. Kautz, Unsupervised Image-to-Image Translation Networks, NIPS 2017
Unsupervised image-to-image translation
![Page 56: SourceCycle-Consistent Adversarial Networks, ICCV 2017. ... (e.g., dog to cat) • Can get confused on images outside of the training domains (e.g., horse with rider) • Cannot close](https://reader034.vdocuments.site/reader034/viewer/2022052008/601cb25298f83944e7188228/html5/thumbnails/56.jpg)
Interesting New Yorker article
https://www.newyorker.com/magazine/2018/11/12/in-the-age-of-ai-is-seeing-still-believing