Just Say the Name: Online Continual Learning with Category Names Only via Data Generation

We provide the dataset used for Name-only continual learning, generated using Stable Diffusion XL, DALL.E-2, CogView2, and DeepFloyd IF models.

Disclaimer

This dataset is created solely for academic purposes. We minimized human intervention to ensure a fair comparison with the baseline methods discussed in our paper. Despite our efforts, the extensive size of the dataset prevented us from examining each image individually.

Therefore, we emphasize that this dataset should only be used for academic purposes, and we disclaim any issues arising from improper use. Specifically, this dataset is recommended solely for assessing the performance of classification models derived from datasets generated by generative models.

Dataset Composition

We have constructed generative model versions of three publicly available datasets: PACS[1], DomainNet[2], and CCT[2]. All datasets will be released sequentially. Additionally, we plan to upload all datasets that have been processed using our proposed RMD-based ensemble method.

PACS dataset from generative model
DomainNet dataset from generative model
CCT dataset from generative model

Download the dataset

clone this repository by following the commands below.

git clone https://huggingface.co/datasets/seongwon980/Nameonly_generated

After downloading the dataset, you may check the directories and the expected outcome is as below.

cd Nameonly_generated/raw_datasets/PACS_sdxl
ls
dog   elephant   giraffe   guitar    horse    house    person

Related work that uses this dataset

Just Say the Name: Online Continual Learning with Category Names Only via Data Generation
Minhyuk Seo , Diganta Misra , Seongwon Cho , Minjae Lee , Jonghyun Choi
arXiv 2024

References

[1] Li, Da, et al. "Deeper, broader and artier domain generalization." ICCV, 2017.

[2] Peng, Xingchao, et al. "Moment matching for multi-source domain adaptation." ICCV, 2019.

[3] Beery, Sara, Grant Van Horn, and Pietro Perona. "Recognition in terra incognita." ECCV, 2018.

Citation

If you find this repository useful, please cite this repository.

@article{seo2024just,
  title={Just Say the Name: Online Continual Learning with Category Names Only via Data Generation},
  author={Seo, Minhyuk and Misra, Diganta and Cho, Seongwon and Lee, Minjae and Choi, Jonghyun},
  journal={arXiv preprint arXiv:2403.10853},
  year={2024}
}