Dataset Card for Ultra-FGVC

Dataset Summary

Ultra-FGVC (Ultra-Fine-Grained Visual Categorization) is a benchmark dataset designed for ultra-fine-grained image classification tasks, particularly in biological domains such as plant phenotype analysis. It consists of several subsets featuring subtle inter-class visual differences. This dataset was originally proposed in the ICCV 2021 paper.

This Hugging Face version mirrors the original dataset and is provided for convenience and research reproducibility.

Datasets	# of Categories	Training	Test
SoyAgeing	198	4,950	4,950
SoyGene	1,110	12,763	11,143
Cotton80	80	240	240
SoyLocal	200	600	600
SoyGlobal	1,938	5,814	5,814

Supported Tasks

• image-classification: Standard supervised classification on labeled subsets.

Dataset Structure

Each subset (e.g., SoyAgeing-R1, SoyAgeing-R3, etc.) contains:

ultra-fgvc/
├── SoyAgeing/
│   ├── SoyAgeing-R1.zip/
│   │   ├── anno/
│   │   │   └── {test, train, val}.txt
│   │   └── images/
│   │   │   └── yd012_R1_{1-10}_{train,test}.png
│   └── ...
├── SoyGene/
│   ├── SoyGene-Part-1.zip
│   └── ...
└── README.md

Citation

If you use this dataset, please cite the following papers:

@inproceedings{yu2021benchmark,
  title     = {Benchmark Platform for Ultra-Fine-Grained Visual Categorization Beyond Human Performance},
  author    = {Yu, Xiaohan and Zhao, Yang and Gao, Yongsheng and Yuan, Xiaohui and Xiong, Shengwu},
  booktitle = {Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV)},
  pages     = {10285--10295},
  year      = {2021}
}

License

This dataset is distributed under the MIT License.

Source

• Original dataset: https://github.com/XiaohanYu-GU/Ultra-FGVC