Frequency Decomposition to Tap the Potential of Single Domain for Generalization


Hongjing Niu (University of Science and Technology of China), Qingyue Yang (University of Science and Technology of China), Pengfei Xia (University of Science and Technology of China), Wei Zhang (University of Science and Technology of China), Bin Li (University of Science and Technology of China), Feng Zhao (University of Science and Technology of China)
The 35th British Machine Vision Conference
PDFPosterVideo (Right click to download)Supplementary

Abstract

Domain generalization (DG) is essential for general artificial intelligence, enabling models to operate across unseen domains. This paper addresses the challenge of DG when limited to single-source domain training, where identifying domain-invariant features is difficult due to lack of comparative data. We propose that these invariant features are embedded in single-source domain samples and focus on extracting them. Our hypothesis suggests a close relationship between these features and frequency. We introduce a novel method that leverages multiple frequency domains. The approach involves dividing each image's frequency domain into subdomains and extracting features via a dual-branch network. This technique forces the model to learn from a narrowly defined spectrum, enhancing the detection of domain-invariant features that might be overshadowed by easily learned features. Extensive experiments show that frequency decomposition aids in learning complex features and our method surpasses existing single-source domain generalization methods.

Citation

@inproceedings{Niu_2024_BMVC,
author    = {Hongjing Niu and Qingyue Yang and Pengfei Xia and Wei Zhang and Bin Li and Feng Zhao},
title     = {Frequency Decomposition to Tap the Potential of Single Domain for Generalization},
booktitle = {35th British Machine Vision Conference 2024, {BMVC} 2024, Glasgow, UK, November 25-28, 2024},
publisher = {BMVA},
year      = {2024},
url       = {https://papers.bmvc2024.org/0740.pdf}
}


Copyright © 2024 The British Machine Vision Association and Society for Pattern Recognition
The British Machine Vision Conference is organised by The British Machine Vision Association and Society for Pattern Recognition. The Association is a Company limited by guarantee, No.2543446, and a non-profit-making body, registered in England and Wales as Charity No.1002307 (Registered Office: Dept. of Computer Science, Durham University, South Road, Durham, DH1 3LE, UK).

Imprint | Data Protection