conditional gan - idc.ac.il · conditional gan, mehdi mirza and simon osindero (nov 2014)...

61
Conditional GAN Deep Image Processing Seminar

Upload: others

Post on 12-Sep-2019

16 views

Category:

Documents


0 download

TRANSCRIPT

Page 1: Conditional GAN - idc.ac.il · Conditional Gan, Mehdi Mirza and Simon Osindero (Nov 2014) Conditional GAN min-max Train discriminator to maximize the probability of the training data

Conditional GANDeep Image Processing Seminar

Page 2: Conditional GAN - idc.ac.il · Conditional Gan, Mehdi Mirza and Simon Osindero (Nov 2014) Conditional GAN min-max Train discriminator to maximize the probability of the training data

Agenda

● Conditional Gan (Nov 2014)

● Image-to-Image Translation (Nov 2017)

● Unpaired Image-to-Image Translation using Cycle-Consistent Adversarial Networks (Feb 2018)

Page 3: Conditional GAN - idc.ac.il · Conditional Gan, Mehdi Mirza and Simon Osindero (Nov 2014) Conditional GAN min-max Train discriminator to maximize the probability of the training data

Conditional Gan

Mehdi Mirza,Simon OsinderoNov 2014

Page 4: Conditional GAN - idc.ac.il · Conditional Gan, Mehdi Mirza and Simon Osindero (Nov 2014) Conditional GAN min-max Train discriminator to maximize the probability of the training data

Conditional GAN

Dreal

fake

y_real(data)

y_fakeGz_noise

x

Conditional Gan, Mehdi Mirza and Simon Osindero (Nov 2014)

Page 5: Conditional GAN - idc.ac.il · Conditional Gan, Mehdi Mirza and Simon Osindero (Nov 2014) Conditional GAN min-max Train discriminator to maximize the probability of the training data

Why do we need a conditional GAN?

In an unconditioned generative model, there is no control on

modes of the data being generated.

In the Conditional GAN (CGAN), the generator learns to generate a fake sample with a specific condition or characteristics (such as a label associated with an image or more detailed tag) rather than a generic sample from unknown noise distribution.

Conditional Gan, Mehdi Mirza and Simon Osindero (Nov 2014)

Page 6: Conditional GAN - idc.ac.il · Conditional Gan, Mehdi Mirza and Simon Osindero (Nov 2014) Conditional GAN min-max Train discriminator to maximize the probability of the training data

Conditional Gan, Mehdi Mirza and Simon Osindero (Nov 2014)

Conditional generative adversarial nets for convolutional face generationJon Gauthier

Page 7: Conditional GAN - idc.ac.il · Conditional Gan, Mehdi Mirza and Simon Osindero (Nov 2014) Conditional GAN min-max Train discriminator to maximize the probability of the training data

Conditional GAN

Conditional Gan, Mehdi Mirza and Simon Osindero (Nov 2014)

In the discriminator x and y are presented as inputs to a discriminative function.

In the generator the prior input noise z, and y are combined in joint hidden representation.

Xy

X

Page 8: Conditional GAN - idc.ac.il · Conditional Gan, Mehdi Mirza and Simon Osindero (Nov 2014) Conditional GAN min-max Train discriminator to maximize the probability of the training data

Conditional GAN

Conditional Gan, Mehdi Mirza and Simon Osindero (Nov 2014)

Why do we still need z?

Page 9: Conditional GAN - idc.ac.il · Conditional Gan, Mehdi Mirza and Simon Osindero (Nov 2014) Conditional GAN min-max Train discriminator to maximize the probability of the training data

GAN min-max reminder

Train discriminator to maximize the probability of the training data

Train the discriminator to minimize the probability of the data sampled from the generator

Train the generator to maximize the probability that the discriminator assigns to its own sample

Conditional Gan, Mehdi Mirza and Simon Osindero (Nov 2014)

Page 10: Conditional GAN - idc.ac.il · Conditional Gan, Mehdi Mirza and Simon Osindero (Nov 2014) Conditional GAN min-max Train discriminator to maximize the probability of the training data

Conditional GAN min-max

Train discriminator to maximize the probability of the training data

Train the discriminator to minimize the probability of the data sampled from the generator

Train the generator to maximize the probability that the discriminator assigns to its own sample

Conditional Gan, Mehdi Mirza and Simon Osindero (Nov 2014)

Page 11: Conditional GAN - idc.ac.il · Conditional Gan, Mehdi Mirza and Simon Osindero (Nov 2014) Conditional GAN min-max Train discriminator to maximize the probability of the training data

Experimental Results

Conditional Gan, Mehdi Mirza and Simon Osindero (Nov 2014)

Page 12: Conditional GAN - idc.ac.il · Conditional Gan, Mehdi Mirza and Simon Osindero (Nov 2014) Conditional GAN min-max Train discriminator to maximize the probability of the training data

Experimental Results - auto tagging of images

Conditional Gan, Mehdi Mirza and Simon Osindero (Nov 2014)

UCMUsergeneratedmetadata

Page 13: Conditional GAN - idc.ac.il · Conditional Gan, Mehdi Mirza and Simon Osindero (Nov 2014) Conditional GAN min-max Train discriminator to maximize the probability of the training data

Image-to-Image Translation

Philip Isola, Jun-Yan ZhuTinghui Zhou, Alexei A. EfrosNov 2017

Page 14: Conditional GAN - idc.ac.il · Conditional Gan, Mehdi Mirza and Simon Osindero (Nov 2014) Conditional GAN min-max Train discriminator to maximize the probability of the training data

Let’s start with a demo

Page 15: Conditional GAN - idc.ac.il · Conditional Gan, Mehdi Mirza and Simon Osindero (Nov 2014) Conditional GAN min-max Train discriminator to maximize the probability of the training data

Image to image mapping

Image-to-Image Translation, Philip Isola, Jun-Yan Zhu, Tinghui Zhou, Alexei A. Efros (Nov 2017)

Page 16: Conditional GAN - idc.ac.il · Conditional Gan, Mehdi Mirza and Simon Osindero (Nov 2014) Conditional GAN min-max Train discriminator to maximize the probability of the training data

Generalization

Image-to-Image Translation, Philip Isola, Jun-Yan Zhu, Tinghui Zhou, Alexei A. Efros (Nov 2017)

● Setting is always the same: predicting pixels from pixels● Conditional adversarial nets are a general-purpose solution that appears

to work well on a wide variety of these problems. Same architecture and objective, and simply train on different data.

● Similar direction: CNNs.Problem: designing effective losses.Loss: “Minimize Euclidean distance between predicted and ground truth”

● Better: Goal: “make output indistinguishable from reality”Loss: learn automatically from goal (GAN!)

● GANs loss functions adapt to data

Page 17: Conditional GAN - idc.ac.il · Conditional Gan, Mehdi Mirza and Simon Osindero (Nov 2014) Conditional GAN min-max Train discriminator to maximize the probability of the training data

Image-to-Image Translation, Philip Isola, Jun-Yan Zhu, Tinghui Zhou, Alexei A. Efros (Nov 2017)

Image → image transformations.Why do we need a conditional GAN for that?

Conditional generative model:We’ll condition on the input image

Page 18: Conditional GAN - idc.ac.il · Conditional Gan, Mehdi Mirza and Simon Osindero (Nov 2014) Conditional GAN min-max Train discriminator to maximize the probability of the training data

Method - training process

Image-to-Image Translation, Philip Isola, Jun-Yan Zhu, Tinghui Zhou, Alexei A. Efros (Nov 2017)

Page 19: Conditional GAN - idc.ac.il · Conditional Gan, Mehdi Mirza and Simon Osindero (Nov 2014) Conditional GAN min-max Train discriminator to maximize the probability of the training data

Objective

Image-to-Image Translation, Philip Isola, Jun-Yan Zhu, Tinghui Zhou, Alexei A. Efros (Nov 2017)

Tested importance of conditioning the discriminator:

Page 20: Conditional GAN - idc.ac.il · Conditional Gan, Mehdi Mirza and Simon Osindero (Nov 2014) Conditional GAN min-max Train discriminator to maximize the probability of the training data

Mixing the GAN objective

Image-to-Image Translation, Philip Isola, Jun-Yan Zhu, Tinghui Zhou, Alexei A. Efros (Nov 2017)

Previous approaches found it beneficial to mix the GAN objective with a traditional loss function.

Page 21: Conditional GAN - idc.ac.il · Conditional Gan, Mehdi Mirza and Simon Osindero (Nov 2014) Conditional GAN min-max Train discriminator to maximize the probability of the training data

Image-to-Image Translation, Philip Isola, Jun-Yan Zhu, Tinghui Zhou, Alexei A. Efros (Nov 2017)

DropoutAddressing overfitting concerns:

● Using z, past cGANs used Gaussian noise as z

● Here: provide noise only in form of dropout

Still:Output not stochastic enough

Page 22: Conditional GAN - idc.ac.il · Conditional Gan, Mehdi Mirza and Simon Osindero (Nov 2014) Conditional GAN min-max Train discriminator to maximize the probability of the training data

Network Architectures - Generator

Image-to-Image Translation, Philip Isola, Jun-Yan Zhu, Tinghui Zhou, Alexei A. Efros (Nov 2017)

Page 23: Conditional GAN - idc.ac.il · Conditional Gan, Mehdi Mirza and Simon Osindero (Nov 2014) Conditional GAN min-max Train discriminator to maximize the probability of the training data

Network Architectures - Discriminator

Image-to-Image Translation, Philip Isola, Jun-Yan Zhu, Tinghui Zhou, Alexei A. Efros (Nov 2017)

For low frequencies, L1/L2 loss will suffice for enforcing correctness.

We’ll restrict the GAN discriminator to only model high frequencies.

Sufficient to restrict attention only to local image patches: PatchGAN

Page 24: Conditional GAN - idc.ac.il · Conditional Gan, Mehdi Mirza and Simon Osindero (Nov 2014) Conditional GAN min-max Train discriminator to maximize the probability of the training data

Network Architectures - Discriminator

Image-to-Image Translation, Philip Isola, Jun-Yan Zhu, Tinghui Zhou, Alexei A. Efros (Nov 2017)

PatchGAN● Classifies if each

N*N patch in the image is real or fake

● This discriminator is run convolutionally across the image, averaging all responses to provide D’s output.

● Patch is smaller than image

Page 25: Conditional GAN - idc.ac.il · Conditional Gan, Mehdi Mirza and Simon Osindero (Nov 2014) Conditional GAN min-max Train discriminator to maximize the probability of the training data

Experiments and Evaluation

Image-to-Image Translation, Philip Isola, Jun-Yan Zhu, Tinghui Zhou, Alexei A. Efros (Nov 2017)

Page 26: Conditional GAN - idc.ac.il · Conditional Gan, Mehdi Mirza and Simon Osindero (Nov 2014) Conditional GAN min-max Train discriminator to maximize the probability of the training data

Experiments and Evaluation

Image-to-Image Translation, Philip Isola, Jun-Yan Zhu, Tinghui Zhou, Alexei A. Efros (Nov 2017)

400 images in dataset, less than 2 hours of training

Page 27: Conditional GAN - idc.ac.il · Conditional Gan, Mehdi Mirza and Simon Osindero (Nov 2014) Conditional GAN min-max Train discriminator to maximize the probability of the training data

Experiments and Evaluation

Image-to-Image Translation, Philip Isola, Jun-Yan Zhu, Tinghui Zhou, Alexei A. Efros (Nov 2017)

Page 28: Conditional GAN - idc.ac.il · Conditional Gan, Mehdi Mirza and Simon Osindero (Nov 2014) Conditional GAN min-max Train discriminator to maximize the probability of the training data

Experiments and Evaluation

Image-to-Image Translation, Philip Isola, Jun-Yan Zhu, Tinghui Zhou, Alexei A. Efros (Nov 2017)

Page 29: Conditional GAN - idc.ac.il · Conditional Gan, Mehdi Mirza and Simon Osindero (Nov 2014) Conditional GAN min-max Train discriminator to maximize the probability of the training data

Experiments and Evaluation

Image-to-Image Translation, Philip Isola, Jun-Yan Zhu, Tinghui Zhou, Alexei A. Efros (Nov 2017)

Page 30: Conditional GAN - idc.ac.il · Conditional Gan, Mehdi Mirza and Simon Osindero (Nov 2014) Conditional GAN min-max Train discriminator to maximize the probability of the training data

Experiments and Analysis

Image-to-Image Translation, Philip Isola, Jun-Yan Zhu, Tinghui Zhou, Alexei A. Efros (Nov 2017)

How do you evaluate synthesized images?

● Plausibility to a human observed is the ultimate goal →With actual people: using AMT

● Measuring to see if off-the-shelf recognition systems can recognize them →FCN-score, using pre-training semantic classifiers

Page 31: Conditional GAN - idc.ac.il · Conditional Gan, Mehdi Mirza and Simon Osindero (Nov 2014) Conditional GAN min-max Train discriminator to maximize the probability of the training data

Objective function analysis

Image-to-Image Translation, Philip Isola, Jun-Yan Zhu, Tinghui Zhou, Alexei A. Efros (Nov 2017)

● L1 alone - blurry, greyish

● cGAN alone - imaginativeBetter with colors

● Removing conditioning data from the discriminator resulted in: only cares that the output is realistic

Page 32: Conditional GAN - idc.ac.il · Conditional Gan, Mehdi Mirza and Simon Osindero (Nov 2014) Conditional GAN min-max Train discriminator to maximize the probability of the training data

Generator analysis

Image-to-Image Translation, Philip Isola, Jun-Yan Zhu, Tinghui Zhou, Alexei A. Efros (Nov 2017)

U-Net architecture allowed low-level information to shortcut across the network.Did this improve results?

Page 33: Conditional GAN - idc.ac.il · Conditional Gan, Mehdi Mirza and Simon Osindero (Nov 2014) Conditional GAN min-max Train discriminator to maximize the probability of the training data

Discriminator analysis

Image-to-Image Translation, Philip Isola, Jun-Yan Zhu, Tinghui Zhou, Alexei A. Efros (Nov 2017)

PatchGANTested the effect of varying the patch size.

Page 34: Conditional GAN - idc.ac.il · Conditional Gan, Mehdi Mirza and Simon Osindero (Nov 2014) Conditional GAN min-max Train discriminator to maximize the probability of the training data

Organic Applications

Image-to-Image Translation, Philip Isola, Jun-Yan Zhu, Tinghui Zhou, Alexei A. Efros (Nov 2017)

Twitter #edges2cats

Page 35: Conditional GAN - idc.ac.il · Conditional Gan, Mehdi Mirza and Simon Osindero (Nov 2014) Conditional GAN min-max Train discriminator to maximize the probability of the training data

Organic Applications

Image-to-Image Translation, Philip Isola, Jun-Yan Zhu, Tinghui Zhou, Alexei A. Efros (Nov 2017)

Page 36: Conditional GAN - idc.ac.il · Conditional Gan, Mehdi Mirza and Simon Osindero (Nov 2014) Conditional GAN min-max Train discriminator to maximize the probability of the training data

Organic Applications

Image-to-Image Translation, Philip Isola, Jun-Yan Zhu, Tinghui Zhou, Alexei A. Efros (Nov 2017)

Page 37: Conditional GAN - idc.ac.il · Conditional Gan, Mehdi Mirza and Simon Osindero (Nov 2014) Conditional GAN min-max Train discriminator to maximize the probability of the training data

More results

Image-to-Image Translation, Philip Isola, Jun-Yan Zhu, Tinghui Zhou, Alexei A. Efros (Nov 2017)

Page 38: Conditional GAN - idc.ac.il · Conditional Gan, Mehdi Mirza and Simon Osindero (Nov 2014) Conditional GAN min-max Train discriminator to maximize the probability of the training data

Image-to-Image Translation, Philip Isola, Jun-Yan Zhu, Tinghui Zhou, Alexei A. Efros (Nov 2017)

Page 39: Conditional GAN - idc.ac.il · Conditional Gan, Mehdi Mirza and Simon Osindero (Nov 2014) Conditional GAN min-max Train discriminator to maximize the probability of the training data

Unpaired Image-to-Image Translation using Cycle-Consistent Adversarial Networks

Jun-Yan Zhu, Taesung Park, Phillip Isola, Alexei A. EfrosFeb 2018

Page 40: Conditional GAN - idc.ac.il · Conditional Gan, Mehdi Mirza and Simon Osindero (Nov 2014) Conditional GAN min-max Train discriminator to maximize the probability of the training data

Unpaired Image to Image

Unpaired image-to-Image Translation, Jun-Yan Zhu, Taesung Park, Phillip Isola, Alexei A. Efros (Feb 2018)

Given any two unordered image collections X and Y, our algorithm learns to automatically “translate” an image from one into the other and vice versa

Page 41: Conditional GAN - idc.ac.il · Conditional Gan, Mehdi Mirza and Simon Osindero (Nov 2014) Conditional GAN min-max Train discriminator to maximize the probability of the training data

Unpaired Image to Image

Unpaired image-to-Image Translation, Jun-Yan Zhu, Taesung Park, Phillip Isola, Alexei A. Efros (Feb 2018)

We are not using a training set of aligned image pairsWe’ll translate an image from a source domain X to a target

domain Y in the absence of paired examples

Page 42: Conditional GAN - idc.ac.il · Conditional Gan, Mehdi Mirza and Simon Osindero (Nov 2014) Conditional GAN min-max Train discriminator to maximize the probability of the training data

Unpaired Image to Image

Unpaired image-to-Image Translation, Jun-Yan Zhu, Taesung Park, Phillip Isola, Alexei A. Efros (Feb 2018)

“Our approach builds on the “pix2pix” framework of Isola et al., which uses a conditional generative adversarial network to learn a mapping from input to output

images”

Page 43: Conditional GAN - idc.ac.il · Conditional Gan, Mehdi Mirza and Simon Osindero (Nov 2014) Conditional GAN min-max Train discriminator to maximize the probability of the training data

Unpaired Image to Image

Unpaired image-to-Image Translation, Jun-Yan Zhu, Taesung Park, Phillip Isola, Alexei A. Efros (Feb 2018)

We want:Mapping such that the output is indistinguishable from images by an adversary trained to classify apart from .

Problems:Meaningful pairingIn practice - mode collapse

We need to add more structure to our objective

Page 44: Conditional GAN - idc.ac.il · Conditional Gan, Mehdi Mirza and Simon Osindero (Nov 2014) Conditional GAN min-max Train discriminator to maximize the probability of the training data

Unpaired Image to Image

Unpaired image-to-Image Translation, Jun-Yan Zhu, Taesung Park, Phillip Isola, Alexei A. Efros (Feb 2018)

We need to add more structure to our objective →Translation should be “cycle consistent”

Combining this loss with adversarial losses on domains X and Y yieldsour full objective for unpaired image-to-image translation

Page 45: Conditional GAN - idc.ac.il · Conditional Gan, Mehdi Mirza and Simon Osindero (Nov 2014) Conditional GAN min-max Train discriminator to maximize the probability of the training data

Consistency Cycle loss

Unpaired image-to-Image Translation, Jun-Yan Zhu, Taesung Park, Phillip Isola, Alexei A. Efros (Feb 2018)

Page 46: Conditional GAN - idc.ac.il · Conditional Gan, Mehdi Mirza and Simon Osindero (Nov 2014) Conditional GAN min-max Train discriminator to maximize the probability of the training data

Consistency Cycle loss

Unpaired image-to-Image Translation, Jun-Yan Zhu, Taesung Park, Phillip Isola, Alexei A. Efros (Feb 2018)

Also the opposite way

Cycle consistency loss:

Adversarial losses:

Page 47: Conditional GAN - idc.ac.il · Conditional Gan, Mehdi Mirza and Simon Osindero (Nov 2014) Conditional GAN min-max Train discriminator to maximize the probability of the training data

Full objective

Unpaired image-to-Image Translation, Jun-Yan Zhu, Taesung Park, Phillip Isola, Alexei A. Efros (Feb 2018)

Page 48: Conditional GAN - idc.ac.il · Conditional Gan, Mehdi Mirza and Simon Osindero (Nov 2014) Conditional GAN min-max Train discriminator to maximize the probability of the training data

Experiments and Evaluation

Image-to-Image Translation, Philip Isola, Jun-Yan Zhu, Tinghui Zhou, Alexei A. Efros (Nov 2017)Unpaired image-to-Image Translation, Jun-Yan Zhu, Taesung Park, Phillip Isola, Alexei A. Efros (Feb 2018)

Page 49: Conditional GAN - idc.ac.il · Conditional Gan, Mehdi Mirza and Simon Osindero (Nov 2014) Conditional GAN min-max Train discriminator to maximize the probability of the training data

Experiments and Analysis

How do you evaluate synthesized images?

● Plausibility to a human observed is the ultimate goal →With actual people: using AMT

● Measuring to see if off-the-shelf recognition systems can recognize them →FCN-score, using pre-training semantic classifiers

Image-to-Image Translation, Philip Isola, Jun-Yan Zhu, Tinghui Zhou, Alexei A. Efros (Nov 2017)Unpaired image-to-Image Translation, Jun-Yan Zhu, Taesung Park, Phillip Isola, Alexei A. Efros (Feb 2018)

Page 50: Conditional GAN - idc.ac.il · Conditional Gan, Mehdi Mirza and Simon Osindero (Nov 2014) Conditional GAN min-max Train discriminator to maximize the probability of the training data

Experiments and Analysis

Image-to-Image Translation, Philip Isola, Jun-Yan Zhu, Tinghui Zhou, Alexei A. Efros (Nov 2017)Unpaired image-to-Image Translation, Jun-Yan Zhu, Taesung Park, Phillip Isola, Alexei A. Efros (Feb 2018)

Page 51: Conditional GAN - idc.ac.il · Conditional Gan, Mehdi Mirza and Simon Osindero (Nov 2014) Conditional GAN min-max Train discriminator to maximize the probability of the training data

Experiments and Analysis

Image-to-Image Translation, Philip Isola, Jun-Yan Zhu, Tinghui Zhou, Alexei A. Efros (Nov 2017)Unpaired image-to-Image Translation, Jun-Yan Zhu, Taesung Park, Phillip Isola, Alexei A. Efros (Feb 2018)

Page 52: Conditional GAN - idc.ac.il · Conditional Gan, Mehdi Mirza and Simon Osindero (Nov 2014) Conditional GAN min-max Train discriminator to maximize the probability of the training data

Experiments and Analysis

Image-to-Image Translation, Philip Isola, Jun-Yan Zhu, Tinghui Zhou, Alexei A. Efros (Nov 2017)Unpaired image-to-Image Translation, Jun-Yan Zhu, Taesung Park, Phillip Isola, Alexei A. Efros (Feb 2018)

Page 53: Conditional GAN - idc.ac.il · Conditional Gan, Mehdi Mirza and Simon Osindero (Nov 2014) Conditional GAN min-max Train discriminator to maximize the probability of the training data

Image-to-Image Translation, Philip Isola, Jun-Yan Zhu, Tinghui Zhou, Alexei A. Efros (Nov 2017)Unpaired image-to-Image Translation, Jun-Yan Zhu, Taesung Park, Phillip Isola, Alexei A. Efros (Feb 2018)

Page 54: Conditional GAN - idc.ac.il · Conditional Gan, Mehdi Mirza and Simon Osindero (Nov 2014) Conditional GAN min-max Train discriminator to maximize the probability of the training data

Image-to-Image Translation, Philip Isola, Jun-Yan Zhu, Tinghui Zhou, Alexei A. Efros (Nov 2017)Unpaired image-to-Image Translation, Jun-Yan Zhu, Taesung Park, Phillip Isola, Alexei A. Efros (Feb 2018)

Page 55: Conditional GAN - idc.ac.il · Conditional Gan, Mehdi Mirza and Simon Osindero (Nov 2014) Conditional GAN min-max Train discriminator to maximize the probability of the training data

Image-to-Image Translation, Philip Isola, Jun-Yan Zhu, Tinghui Zhou, Alexei A. Efros (Nov 2017)Unpaired image-to-Image Translation, Jun-Yan Zhu, Taesung Park, Phillip Isola, Alexei A. Efros (Feb 2018)

Page 56: Conditional GAN - idc.ac.il · Conditional Gan, Mehdi Mirza and Simon Osindero (Nov 2014) Conditional GAN min-max Train discriminator to maximize the probability of the training data

Image-to-Image Translation, Philip Isola, Jun-Yan Zhu, Tinghui Zhou, Alexei A. Efros (Nov 2017)Unpaired image-to-Image Translation, Jun-Yan Zhu, Taesung Park, Phillip Isola, Alexei A. Efros (Feb 2018)

Page 57: Conditional GAN - idc.ac.il · Conditional Gan, Mehdi Mirza and Simon Osindero (Nov 2014) Conditional GAN min-max Train discriminator to maximize the probability of the training data

Face to Ramen

Image-to-Image Translation, Philip Isola, Jun-Yan Zhu, Tinghui Zhou, Alexei A. Efros (Nov 2017)Unpaired image-to-Image Translation, Jun-Yan Zhu, Taesung Park, Phillip Isola, Alexei A. Efros (Feb 2018)

Page 58: Conditional GAN - idc.ac.il · Conditional Gan, Mehdi Mirza and Simon Osindero (Nov 2014) Conditional GAN min-max Train discriminator to maximize the probability of the training data

Resurrecting Ancient Cities

Image-to-Image Translation, Philip Isola, Jun-Yan Zhu, Tinghui Zhou, Alexei A. Efros (Nov 2017)Unpaired image-to-Image Translation, Jun-Yan Zhu, Taesung Park, Phillip Isola, Alexei A. Efros (Feb 2018)

Page 59: Conditional GAN - idc.ac.il · Conditional Gan, Mehdi Mirza and Simon Osindero (Nov 2014) Conditional GAN min-max Train discriminator to maximize the probability of the training data

Limitations

Image-to-Image Translation, Philip Isola, Jun-Yan Zhu, Tinghui Zhou, Alexei A. Efros (Nov 2017)Unpaired image-to-Image Translation, Jun-Yan Zhu, Taesung Park, Phillip Isola, Alexei A. Efros (Feb 2018)

Page 60: Conditional GAN - idc.ac.il · Conditional Gan, Mehdi Mirza and Simon Osindero (Nov 2014) Conditional GAN min-max Train discriminator to maximize the probability of the training data

References

Page 61: Conditional GAN - idc.ac.il · Conditional Gan, Mehdi Mirza and Simon Osindero (Nov 2014) Conditional GAN min-max Train discriminator to maximize the probability of the training data

● Conditional Gan, Mehdi Mirza and Simon Osindero (Nov 2014)● Image-to-Image Translation, Philip Isola, Jun-Yan Zhu, Tinghui Zhou, Alexei A. Efros

(Nov 2017)● Unpaired Image-to-Image Translation using Cycle-Consistent Adversarial Networks

Jun-Yan Zhu, Taesung Park, Phillip Isola, Alexei A. Efros(Feb 2018)