search

yasserben / DATUM

[CVPR-W 2023] Official Implementation of One-shot Unsupervised Domain Adaptation with Personalized Diffusion Models
https://arxiv.org/abs/2303.18080
Other
72 stars 4 forks source link

How to obtain training images of the target domain? #3

Closed ddghjikle closed 1 year ago

ddghjikle commented 1 year ago

Hi, thanks for releasing codes. Could you please specify how to randomly obtain training images of the target domain? It seems that released codes do not contain the function of randomly choosing.

yasserben commented 1 year ago

Hi @ddghjikle , sorry for the late reply ! Indeed, the code doesn't contain a script which randomly chooses the single target domain. As it was specified in the README, you can choose any image you want from the training set of Cityscapes and use it to personalize your text-to-image diffusion model. In our paper we chose the following three random images and we averaged the results over them : stuttgart_000016_000019_leftImg8bit.png cologne_000115_000019_leftImg8bit.png cologne_000006_000019_leftImg8bit.png