Handling Class-Imbalance for Improved Zero-Shot Domain Generalization
Ahmad Arfeen (IISc bangalore), Titir Dutta (Indian Institute of Science, Bangalore), Soma Biswas (Indian Institute of Science, Bangalore)*
The 33rd British Machine Vision Conference
Abstract
Zero-shot domain generalization (ZSDG) simultaneously addresses the challenges of dissimilar distribution and disjoint label-spaces of the training and test data in the context of classification. State-of-the-art ZSDG approaches leverage multiple source domain data and the semantic information of the classes to learn domain-agnostic features for handling both unseen domains and classes. Effective feature learning depends significantly on the training data characteristics, which has been largely overlooked for this task. In this work, we propose to handle one such important challenge, namely class-imbalance for the ZSDG problem. Towards this end, we propose a novel framework, Mixing-based Adaptive Margin Classifier Network (MAMC-Net) for handling this real-world challenge. Specifically, it consists of two components, (i) a novel adaptive-margin based semantic classifier for handling the data imbalance in the training data and (ii) a module for determining the mixing ratio when the input domains and classes are mixed, for better domain agnostic class-discrimination. Extensive experiments and analysis performed on multiple large-scale datasets, DomainNet and DomainNet-LS demonstrate the effectiveness of MAMC-Net to address the challenging ZSDG scenario.Video
Citation
@inproceedings{Arfeen_2022_BMVC,
author = {Ahmad Arfeen and Titir Dutta and Soma Biswas},
title = {Handling Class-Imbalance for Improved Zero-Shot Domain Generalization},
booktitle = {33rd British Machine Vision Conference 2022, {BMVC} 2022, London, UK, November 21-24, 2022},
publisher = {{BMVA} Press},
year = {2022},
url = {https://bmvc2022.mpi-inf.mpg.de/0728.pdf}
}
Copyright © 2022
The British Machine Vision Conference is organised by