autoexpert-cvpr2026-workshop/ASA2026_dataset

Automated Species Annotation Dataset

Welcome to the public dataset repository for the CVPR 2026 Auto-Annotation Challenge. In this challenge, we use five fine-grained species-level datasets. The Species196 dataset is further split into three taxon-specific benchmarks: Insecta, Weeds, and Mollusca. Additionally, for a more comprehensive evaluation, we introduce two new datasets: Meta Album and RareSpecies. From Meta Album dataset, we select Insects2, Plankton, and PlantDoc for the competition.

Datasets list:

• Semi-Aves
• Species196 (Insecta, Weeds, Mollusca)
• FungiTastic
• Meta Album (Insects2, Plankton, PlantDoc) [New!]
• RareSpecies [New!]

To download this repository, run the following command:

git clone https://huggingface.co/datasets/cvpr2026-workshop-annoexpert/ASA2026_dataset

Directory Structure

The repository is organized as follows:

ASA2026_dataset/
|-- fungitastic/
|   |-- train/                          # few-shot data
|   |-- test/                           # test data
|   |-- fewshot16.txt                   # few-shot data annotation: img_path, label
|   |-- test.txt                        # test data ids: id, img_path
|   |-- fungitastic_description.csv     # short description for each species
|-- insecta/
|-- insects2/
|-- mollusca/
|-- plankton/
|-- plantdoc/
|-- rare_species/
|-- semi-aves/
|-- weeds/
|-- species196_download.py
|-- README.md                           # Dataset documentation

Details of Datasets

Dataset Composition

For each benchmark, we first filter out classes with ambiguous names or insufficient images. From the remaining classes, we randomly sample 16 images per class to form the few-shot training set, while keeping the original test sets for evaluation. Each class is also annotated with its scientific name, common name, and a short textual description. These metadata are provided in {DATASET}_description.csv.

dataset_id	dataset	# sampled cls	# sampled test
0	Semi-Aves	200	4,000
1	Insecta	78	2,910
2	Weeds	20	913
3	Mollusca	7	704
4	FungiTastic	196	3,805
5	Plankton [New!]	86	2,064
6	Insects2 [New!]	102	2,448
7	PlantDoc [New!]	27	648
8	RareSpecies [New!]	395	3,950

Detailed introductions and access instructions for each dataset are provided below.

Semi-Aves

The Semi-Aves dataset is designed for fine-grained visual recognition of bird species. It is sourced from iNaturalist, a large-scale, community-driven biodiversity dataset. The data exhibits high visual diversity and real-world complexity.

Download

You can download the training (trainval_images.tar.gz) and testing(test.tar.gz) sets from the official competition page on Kaggle. Once downloaded, please organize the files into the semi-aves/ folder as follows:

ASA2026_dataset/
|-- semi-aves/
|   |-- train/      # training images (extracted from trainval_images.tar.gz)
|   |-- test/       # test images (extracted from test.tar.gz)
|   |-- ...

Species196

The Species196 dataset (under the CC BY-NC-SA 4.0 license) is a large-scale semi-supervised dataset of 196-category invasive species. From this dataset, we derive three taxon-specific benchmarks: Insecta, Weeds, and Mollusca.

Download

Update: Due to the expiration of several original image URLs, the legacy download script is no longer functional. To request access to the full dataset, please complete our Google Form.

Deprecated Legacy Script: We provide a download script to fetch the training and test images for Insecta, Weeds, and Mollusca. Simply run:

cd ASA2026_dataset/

python species196_download.py

The script will automatically download all three datasets and organize the images into the following directory structure:

ASA2026_dataset/
|-- insecta/
|   |-- train/      # training images
|   |-- test/       # test images
|   |-- ...
|-- weeds/
|   |-- train/
|   |-- test/
|   |-- ...
|-- mollusca/
|   |-- train/
|   |-- test/
|   |-- ...

FungiTastic

FungiTastic (under the CC BY-NC-SA 4.0 license) is a large-scale, expert-verified, multi-modal dataset for wild fungi recognition. It comprises ~350k observations across 5,000+ species, collected over 20 years. The dataset is designed to support a range of research directions, including fine-grained classification, few-shot learning, multimodal learning, and domain adaptation under natural distribution shifts.

Download

The training and test sets can be downloaded directly from this repository.

Meta Album

Meta Album (under the CC BY-NC 4.0 license) is a meta-dataset of 40 datasets across 10 domains, designed for few-shot, meta-, and continual learning. For this challenge, we select three specific subsets: Insects2, Plankton, and PlantDoc.

Download

The training and test sets can be downloaded directly from this repository.

RareSpecies

RareSpecies (under the CC BY-NC 4.0 license) is a fine-grained classification benchmark, designed for zero-shot tasks on threatened biodiversity, focusing on species categorized as Near Threatened to Extinct in the Wild on the IUCN Red List.

Download

The training and test sets can be downloaded directly from this repository.

Terms of Use

By downloading these datasets you agree to the following terms:

• You will abide by the iNaturalist Terms of Service, Species196 Terms of Service, license of FungiTastic, license of Meta Album, and license of RareSpecies.
• You will use the data only for non-commercial research and educational purposes.
• The AutoExpert workshop makes no representations or warranties regarding the data, including but not limited to warranties of non-infringement or fitness for a particular purpose.
• You accept full responsibility for your use of the data and shall defend and indemnify the AutoExpert workshop, including its employees, officers, and agents, against any and all claims arising from your use of the data, including but not limited to your use of any copies of copyrighted images that you may create from the data.

Citation

🙏 We thank the providers of the datasets used in this benchmark.

If you use this benchmark in your research or challenge submission, please cite our CVPR 2026 Workshop, the baseline paper, and the original dataset paper:

@article{liu2025poc,
title={Surely Large Multimodal Models (Don't) Excel in Visual Species Recognition?}, 
author={Liu, Tian and Basu, Anwesha and Caverlee, James and Kong, Shu},
journal={arXiv preprint arXiv:2512.15748},
year={2025}
}


@article{su2021semi,
  title={The semi-supervised inaturalist-aves challenge at fgvc7 workshop},
  author={Su, Jong-Chyi and Maji, Subhransu},
  journal={arXiv preprint arXiv:2103.06937},
  year={2021}
}


@article{he2023species196,
  title={Species196: A one-million semi-supervised dataset for fine-grained species recognition},
  author={He, Wei and Han, Kai and Nie, Ying and Wang, Chengcheng and Wang, Yunhe},
  journal={Advances in Neural Information Processing Systems},
  volume={36},
  pages={44957--44975},
  year={2023}
}


@inproceedings{picek2025fungitastic,
  title={Fungitastic: A multi-modal dataset and benchmark for image categorization},
  author={Picek, Lukas and Janouskova, Klara and Cermak, Vojtech and Matas, Jiri},
  booktitle={2025 IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops (CVPRW)},
  pages={2037--2047},
  year={2025},
  organization={IEEE}
}


@inproceedings{meta-album-2022,
  title={Meta-Album: Multi-domain Meta-Dataset for Few-Shot Image Classification},
  author={Ullah, Ihsan and Carrion, Dustin and Escalera, Sergio and Guyon, Isabelle M and Huisman, Mike and Mohr, Felix and van Rijn, Jan N and Sun, Haozhe and Vanschoren, Joaquin and Vu, Phan Anh},
  booktitle={Thirty-sixth Conference on Neural Information Processing Systems Datasets and Benchmarks Track},
  url = {https://meta-album.github.io/},
  year = {2022}
}


@inproceedings{stevens2024bioclip,
title = {{B}io{CLIP}: A Vision Foundation Model for the Tree of Life}, 
author = {Samuel Stevens and Jiaman Wu and Matthew J Thompson and Elizabeth G Campolongo and Chan Hee Song and David Edward Carlyn and Li Dong and Wasila M Dahdul and Charles Stewart and Tanya Berger-Wolf and Wei-Lun Chao and Yu Su},
booktitle={Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)},
year = {2024},
pages = {19412-19424}
}