Team Ai
Datasetpublic

PBatch23888/birds-object-detection

Birds Object Detection 640x640 A single-class bird object-detection dataset prepared for Ultralytics at a 640 x 640 input resolution. Positive images use a deterministic approximately 90% training and 10% validation split. Reviewed garden negatives are added to the training split. Contents data/yolo_birds_640x640.tar.gz: complete Ultralytics dataset archive. manifest.json: archive size, SHA-256 checksum, and dataset metadata. After extraction, the archive… See the full description on the dataset page: https://huggingface.co/datasets/PBatch23888/birds-object-detection.

sourceHugging Faceupdated 11d agoView on Hugging Face
0likes44downloads
Dataset Card

Birds Object Detection 640x640

A single-class bird object-detection dataset prepared for Ultralytics at a 640 x 640 input resolution. Positive images use a deterministic approximately 90% training and 10% validation split. Reviewed garden negatives are added to the training split.

Contents

  • —data/yolo_birds_640x640.tar.gz: complete Ultralytics dataset archive.
  • —manifest.json: archive size, SHA-256 checksum, and dataset metadata.

After extraction, the archive contains data.yaml and matching image and YOLO label trees:

text
data.yaml
images/train/
images/val/
labels/train/
labels/val/

Download

python
from huggingface_hub import hf_hub_download

archive = hf_hub_download(
    repo_id="PBatch23888/birds-object-detection",
    repo_type="dataset",
    filename="data/yolo_birds_640x640.tar.gz",
)

Private repositories require an authenticated Hugging Face account with access to the dataset.

Dataset composition

The data builder combines bird annotations from Open Images, COCO 2017, Pascal VOC 2012, Birdsnap, and NABirds. Open Images group-of and depiction annotations are excluded. All retained annotations are mapped to the single class bird.

Training also includes human-reviewed bird-free garden images from Open Images and Places365, represented by empty YOLO label files. Places365 source images are 256 x 256 pixels and are resized with padding.

Licensing

This is a compilation of multiple upstream datasets. Their respective terms and licenses continue to apply. Review each upstream dataset's license before redistributing or using this compilation.