RealBirdID

🎉 RealBirdID is finally available!

RealBirdID is now available for public release. It is a bird identification and abstention benchmark with answerable species examples and unanswerable examples labeled at genus level.

Paper · Code and starter notebooks

Load the data

from datasets import load_dataset

answerable = load_dataset("cvl-umass/RealBirdID", "answerable", split="validation", streaming=True)
unanswerable = load_dataset("cvl-umass/RealBirdID", "unanswerable", split="validation", streaming=True)
example = next(iter(answerable))
print(example["taxon_name"], example["license"], example["photo_attribution"])

Both configs contain validation data. There is no training split. Answerable examples use the original answerable v1 set. Unanswerable examples use the later unanswerable v2 set.

Release scope

Config	Original rows	Released rows	Rows without a CC photo license
answerable	31,885	31,885	0
unanswerable	3,253	3,253	457
Total	35,138	35,138	457

The release includes 457 unanswerable rows whose source photo license is null. These rows are labeled all-rights-reserved when the source attribution explicitly says so, or unknown otherwise. They are not CC-licensed. Inclusion does not grant permission to reuse or redistribute these images. The original null value is preserved in source_license_code. These rows are isolated in data/unanswerable/validation-restricted-*.parquet for separate handling and removal. To use only CC-licensed images, filter on license or load only validation-licensed-*.parquet files.

The public schema contains images, observation and photo identifiers, taxon labels, categorical abstention reasons, photo licenses, attribution, source links, and source-version labels. It excludes raw observation snapshots, user profiles, coordinates, discussion text, and local filesystem paths. taxonomy.json lists benchmark species. license_counts.json records source, included, and excluded counts.

Licenses

Images retain their individual licenses. This dataset does not grant a blanket MIT license over third-party images. Each row includes license, source_license_code, license_url, photo_attribution, photo_url, and source_url. The license values are taken from the selected photo record, not from its observation record. They reflect the locally retained source snapshot rather than a new check of each source page.

Photo license	Answerable rows	Unanswerable rows	Total
CC0	713	53	766
CC BY	3,239	267	3,506
CC BY-SA	543	32	575
CC BY-ND	125	9	134
CC BY-NC	24,980	2,258	27,238
CC BY-NC-SA	501	45	546
CC BY-NC-ND	1,784	132	1,916
No CC license (source null)	0	457	457

Preserve attribution and source links. Noncommercial, share-alike, and no-derivatives conditions apply where the row's license specifies them. Consult the linked license before reuse. Non-CC rows have an empty license_url; their source links and attribution do not constitute a redistribution license. Image bytes are packaged from the benchmark's existing local JPEG files without further resizing or encoding. The unanswerable source preparation converted downloaded photos to RGB JPEG; the source links identify the original media.

The reasons are categorical benchmark annotations. Photo licenses do not grant rights to copy observation discussions or other third-party text. Raw observation licenses are separate, and 799 source observations have null or empty observation-license values; their discussions and raw records are not distributed here.

Evaluation

Use answerable rows to measure species identification and unanswerable rows to measure abstention. The starter notebooks show data loading, attributed image display, and simple baseline scoring. The baseline notebook is a workflow demonstration, not a reproduction of the paper's evaluation suite. Pin a dataset revision and state whether evaluation includes the 457 non-CC rows.

Citation

@inproceedings{lawrence2026realbirdid,
  title={RealBirdID: Benchmarking Bird Species Identification in the Era of MLLMs},
  author={Lawrence, Logan and Chasmai, Mustafa and Daroya, Rangel and Liu, Wuao and Jeong, Seoyun and Sun, Aaron and Hamilton, Max and Delattre, Fabien and Saha, Oindrila and Maji, Subhransu and Van Horn, Grant},
  booktitle={Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition},
  year={2026}
}