BJTUYXD/MIT-States_star

MIT-States*

Dataset Introduction

MIT-States* dataset is a low-noise dataset in compositional zero-shot learning (CZSL), which is proposed in the paper "Warm-Cat: Warm-Started Test-Time Comprehensive Knowledge Accumulation for Compositional Zero-Shot Learning". Since original MIT-States datset suffer from substantial noise, with about 70% of its labels being incorrect, we performed a new round of annotation using INternVL-3-8b and filtration for refinement.

Dataset Statistics

• Total Attributes: 141
• Total Objects: 266
• Total Compositions: 1444
• Total Images: 14079

Uasge

Download data/MIT-States_star.zip and uncompress it. Load it using dataset.py from code project of any CZSL model.

Acknowledgement

This dataset is built upon MIT-States and relabeld by InternVL-3-8b. Thanks for their open sourse.

Citation

If you use this dataset in your research, please cite (this paper has not been released yet):

@inproceedings{WARMCAT,
  title     = {Warm-Cat: Warm-Started Test-Time Comprehensive Knowledge Accumulation for Compositional Zero-Shot Learning},
  author    = {Yan, Xudong and Feng, Songhe and Wang, Jiaxin and Su, Xin and Yi, Jin},
}