Used articles and papers
- Unpaired Image-to-Image Translation using Cycle-Consistent Adversarial Networks
Implementation: https://github.com/junyanz/pytorch-CycleGAN-and-pix2pix
Paper: https://arxiv.org/pdf/1703.10593.pdf
- Article having (1.) as basis
https://hackernoon.com/gender-and-race-change-on-your-selfie-with-neural-nets-9a9a1c9c5c16
Goal
Having Cycle GANs for ethnicity transformation, e.g.
- Black and White
- White and Asian
Used Datasets
- CelebA ~ 200.000 images
http://mmlab.ie.cuhk.edu.hk/projects/CelebA.html
- UTKFace ~ 20.000 images
http://aicip.eecs.utk.edu/wiki/UTKFace
Faced Problems
- CelebA is a huge dataset but does not have ethnicity labels unfortunately.
Approach: train Classifier on UTKFace and LFW to label CelebA.
Improvements 2. Replace L1 pixel to pixel identity loss with p2 Norm of feature map distance. see https://cs.stanford.edu/people/jcjohns/papers/eccv16/JohnsonECCV16.pdf