Unpaired image-to-image translation aims to learn proper mappings that can map images from one domain to another domain while preserving the content of the input image. However, with large enough capacities, the network can learn to map the inputs to any random permutation of images in another domain. Existing methods treat two domains as discrete and propose different assumptions to address this problem. In this paper, we start from a different perspective and consider the paths connecting the two domains. We assume that the optimal path length between the input and output image should be the shortest among all possible paths. Based on this assumption, we propose a new method to allow generating images along the path and present a simple way to encourage the network to find the shortest path without pair information. Extensive experiments on various tasks demonstrate the superiority of our approach.
python train.py --dataroot=datasets/cityscapes --direction=BtoA --lambda_path=0.1 --tag=santa
checkpoints/cityscapes
python test.py --dataroot=datasets/cityscapes --name=cityscapes --direction=BtoA
--lambda_path
, --path_layers
, --path_interval_min
and --path_interval_max
. The dataset is constructed from the UTKFace dataset. Then I apply super-resolution model and divide the output images into old/young according to the age. The dataset contains 1500 training and 500 testing images for each domain.
Following shows the first six training images in each domain.
If you use this code for your research, please consider citing our paper:
@inproceedings{xie2023unpaired,
title={Unpaired Image-to-Image Translation With Shortest Path Regularization},
author={Xie, Shaoan and Xu, Yanwu and Gong, Mingming and Zhang, Kun},
booktitle={Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition},
pages={10177--10187},
year={2023}
}