training unpaired but with ID matching?

[omriio]

Hi,
I'm doing a project to convert images taken from camera A to look like camera B, and to tackle it like the horse-zebra problem, but with the exception that I have images of facial parts taken from both cameras but they are not aligned with each other (shooting angle and posing), and I would like to verify it learns features from the same person and not from a random image, to apply to old image without matching B image.
would there be a training sequence with cyclegan-turbo that will bring the best solution to transform from A to B

[lizhaohu]

Hi, I'm doing a project to convert images taken from camera A to look like camera B, and to tackle it like the horse-zebra problem, but with the exception that i have images of facial parts taken from both cameras but they are not aligned with each other (shooting angle and posing), and i would like to verify it learn features each from the same person and not from random image, to apply to image without B image. would there be a training sequence with cyclegan-turbo that will bring best solution to transform from A to B

Hello, I have a similar request. Have you managed to achieve it?

[GaParmar]

Hi @omriio

I think I did not understand your task description. Could you provide some visual examples of the setting that you are describing?

-Gaurav

[omriio]

hi @GaParmar, thank you for responding!
as I understand there are 2 scenarios:
pairs: two sets with images paired perfectly - > use img2img
unpairs: two sets with the same "concepts" (horse-zebras) - > use cyclegan
my case: corresponding pairs of images for each "object", they're not aligned (angle, lighting, posing, etc'), but I need to keep as many as possible characters from the same specific object (textures, shape) to be able to identify it. in this example with full faces its obvious who the characteristics are, but if I want lips-only or forehead-only images, what technique would make the right transformation but keep the object the same?
I did try modifying the "class UnpairedDataset" always to choose couples from the same id and I am quite happy with the results, but would like to hear your input, and thank you for sharing your research