Retrieval Guided Unsupervised Multi-domain Image-to-Image Translation

TL;DR

This paper introduces a retrieval-guided approach to improve unsupervised multi-domain image-to-image translation by leveraging image retrieval systems to enhance image quality and diversity, especially with limited data.

Contribution

It proposes integrating an image retrieval system with translation models to refine and enhance multi-domain image translation, utilizing unlabeled data for better results.

Findings

Improved image quality in multi-domain translation.

Effective use of unlabeled data via retrieval system.

Enhanced diversity and reliability of generated images.

Abstract

Image to image translation aims to learn a mapping that transforms an image from one visual domain to another. Recent works assume that images descriptors can be disentangled into a domain-invariant content representation and a domain-specific style representation. Thus, translation models seek to preserve the content of source images while changing the style to a target visual domain. However, synthesizing new images is extremely challenging especially in multi-domain translations, as the network has to compose content and style to generate reliable and diverse images in multiple domains. In this paper we propose the use of an image retrieval system to assist the image-to-image translation task. First, we train an image-to-image translation model to map images to multiple domains. Then, we train an image retrieval model using real and generated images to find images similar to a query one in content but in a different domain. Finally, we exploit the image retrieval system to fine-tune the image-to-image translation model and generate higher quality images. Our experiments show the effectiveness of the proposed solution and highlight the contribution of the retrieval network, which can benefit from additional unlabeled data and help image-to-image translation models in the presence of scarce data.