Active Gradual Domain Adaptation: Dataset and Approach
Published in IEEE Transaction on Multimedia, 2022
Recommended citation: @ARTICLE{Zhou2021, author={Zhou, Shiji and Wang, Lianzhe and Zhang, Shanghang and Wang, Zhi and Zhu, Wenwu}, journal={IEEE Transactions on Multimedia}, title={Active Gradual Domain Adaptation: Dataset and Approach}, year={2022}, pages={1-1}, doi={10.1109/TMM.2022.3142524} } https://ieeexplore.ieee.org/document/9681347
In this work, we proposed the active gradual self-training (AGST) algorithm with active pseudolabeling and gradual semi-supervised domain adaptation to adapt deep neural networks to the changing environments, especially for online web applications, where the data distribution changes gradually due to the evolving environments.
We also created a new dataset – Evolving-Image-Search (EVIS), which shows the evolution of web images in the time range of 12 years. Designed a crawler based on search engine to automatically collect this dataset.
The code and EVIS dataset have been released:
EVIS dataset is released at here.
Code is available at here.
Paper PDF (Not available yet)