Unsupervised Conditional Generation*Domain Transfer

Conditional Generation*Style Transfer)

CycleGAN

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

1.Main Porpose

Unpaired Source&target Data for domain transfer

2.Structure

3.Loss Func

Photo generation from paintings :+Identity Loss

4.Metrics

• Qualitative：AMT studies
• Quantitative:FNC score:for labels->photo,FNC to detect labels in photo
• Quantitative:Standard metrics from Dataset, photo->labels:per pixel,per-class,class IOU

5.Datasets

• Cellection style transfer:Flicker Wikiart
• Object transfiguration ImageNet
• Season transfer Flickr
• Photo Enhancement:Flickr

StarGAN

1. Main Porpose

ConditionalGAN(Just Gene,No transfer)+CycleGAN(One side)

2.Structure

3. Loss Func

• Adversail loss
  
• Domain Classification loss
  
  
• Cycle loss
  
  

4. Metric

• Qualitative :give result pics
• Quantitative:AMT
• Quantitative:classification error of facial expression

5.DataSets

• CelebA
• RaFD

6.Other Contribution

Mask Vector for Multi Datasets trained all together

PairedCycleGAN

PairedCycleGAN: Asymmetric Style Transfer for Applying and Removing Makeup

1. Main Porpose

Different style Make-up and tide-off for face,asymmetric Net

2 .Net Structure

3.Loss Func

• Adversarial loss
  
  
• Identity loss

• Style loss

• whether same makeup style
  

4 .Metric

• Qualitative:picture shown
• Quantitative:AMT on single/multi task
• Quantitative:classification of facial expression

5.Datasets

from Youtube makeup videos