Domain-Sum Feature Transformation For Multi-Target Domain Adaptation
Takumi Kobayashi (National Institute of Advanced Industrial Science and Technology),* Lincon Souza (National Institute of Advanced Industrial Science and Technology (AIST)), Kazuhiro Fukui (University of Tsukuba)
The 34th British Machine Vision Conference
Abstract
Domain adaptation effectively transfers a learner from a source domain to a target domain. Recent deep methods are based on detailed comparison between a pair of source and target domains, which makes it less applicable to multiple domains. In this paper, we address the domain adaptation on the basis of subspace which provides more robust metric. We analyze the subspace methods in domain adaptation to theoretically derive a subspace-based feature transformation in an efficient form of simple summation. It intrinsically contributes to closing a gap between source and target subspaces in an end-to-end deep framework. Besides, due to the robust representation of subspace and the simple transformation, the proposed method naturally deals with multiple domains both for source and target in contrast to previous approaches. Multi-target domain adaptation especially provides efficient inference to process multiple target domains by only a single model. In the experiments on visual domain adaptation tasks, the proposed method exhibits favorable performance in a scenario of the multi-target domains.Video
Citation
@inproceedings{Kobayashi_2023_BMVC,
author = {Takumi Kobayashi and Lincon Souza and Kazuhiro Fukui},
title = {Domain-Sum Feature Transformation For Multi-Target Domain Adaptation},
booktitle = {34th British Machine Vision Conference 2023, {BMVC} 2023, Aberdeen, UK, November 20-24, 2023},
publisher = {BMVA},
year = {2023},
url = {https://papers.bmvc2023.org/0197.pdf}
}
Copyright © 2023
The British Machine Vision Conference is organised by