dronefreak/AnimalsDetection

Animals Detection Images Dataset (Object Detection)

Unofficial redistribution of an 80-class wildlife/animal object-detection subset of Google Open Images, reformatted into a standardized YOLO-compatible layout with a seeded validation split.

Disclaimer

This repository is not an official release of Open Images.

The images and their original box annotations were collected and released by Google (the Open Images project); the underlying photographs themselves were taken by their individual Flickr photographers, who retain copyright subject to the Creative Commons terms they published under. This repository does not claim ownership of any images, annotations, or metadata.

Two-hop provenance. This redistribution's raw files came from a Kaggle re-export, "Animals Detection Images Dataset" by Kaggle user antoreepjana, who downloaded an 80-class animal subset of Open Images V6+ via the community OIDv4_ToolKit. That Kaggle page tags the dataset CC0: Public Domain, but this is not accurate for the underlying data -- see License below for the real, verified terms (Open Images' own licensing page) and why they still permit this mirror.

---

Dataset Overview

This dataset covers 80 animal classes spanning mammals, birds, reptiles, fish, and invertebrates (e.g. Lion, Zebra, Eagle, Jellyfish, Butterfly, Centipede) -- the widest class vocabulary of any object-detection dataset DetectionBench has benchmarked. Boxes are large on average (median ~32% of image area) and images are typically single- or few-animal shots, though some contain many instances of one class (flocks, schools -- up to 123 boxes in one image).

---

Changes from the Official Release

• Cross-class duplicate images merged. Because OIDv4 downloads per class query, the same physical photo is sometimes fetched into more than one class folder when it contains multiple labeled animals (confirmed on this release: a horse/mule photo is byte-identical across the Horse and Mule folders, each folder's label file holding only that folder's class). This repository de-duplicates by filename within each split and merges every class's boxes for that image into one entry, rather than treating each class-folder copy as a separate image. This reduced the raw file-copy counts (22,566 train-side, 6,505 test-side) down to the true unique-image counts below.
• A validation split was carved out of train. The source ships only train/test. This repository holds out a seeded random 15% of train (seed 42) as valid, keeping the official test split as-is.
• Format converted. OIDv4's per-image <ClassName> <XMin> <YMin> <XMax> <YMax> absolute-pixel .txt labels were converted to YOLO-normalized .txt labels and to canonical COCO JSON ([x, y, w, h]) per split.
• Sharded for Hub limits. Images are split into shard_NNN/ subdirectories of ~3,000 files to stay under Hugging Face's practical per-directory file-count ceiling (train: 7 shards, valid: 2, test: 3). This has no effect on the data itself.
• No image pixel content was modified. No boxes were resized, moved, or reinterpreted.

---

Dataset Structure

<repo>/
├── README.md
├── animalsdet_banner.jpg
└── data/
    ├── data.yaml
    ├── images/
    │   ├── train/   shard_000 … shard_006/   (*.jpg + metadata.jsonl, ~3,000 per shard)
    │   ├── valid/   shard_000 … shard_001/
    │   └── test/    shard_000 … shard_002/
    └── labels/
        ├── train/   shard_000 … shard_006/   (*.txt, mirrors images)
        ├── valid/   shard_000 … shard_001/
        └── test/    shard_000 … shard_002/

where:

• data/images/<split>/shard_NNN/ contains the RGB images, plus a metadata.jsonl (file_name and objects.bbox as absolute-pixel COCO [x, y, w, h] with objects.categories) that drives the Hugging Face dataset viewer.
• data/labels/<split>/shard_NNN/ contains one YOLO-format .txt annotation file per image (class x_center y_center width height, normalized), mirroring the image shard layout.
• data/data.yaml is the Ultralytics dataset configuration file (class names, split paths, relative to data/).
• Splits: train 18,402 images / 24,091 boxes · valid 3,247 images / 4,123 boxes · test 6,003 images / 7,576 boxes (27,652 images / 35,790 boxes total).

Classes (80)

Full taxonomy spans mammals (Lion, Zebra, Elephant, Tiger, Giraffe, ...), birds (Eagle, Parrot, Owl, Sparrow, ...), reptiles/amphibians (Crocodile, Snake, Frog, Turtle, ...), fish/marine (Shark, Jellyfish, Seahorse, Sea lion, ...), and invertebrates (Butterfly, Spider, Crab, Snail, ...). See docs/datasets/animalsdet in the DetectionBench repository for the full per-class instance counts and box-geometry statistics.

Class distribution is heavily imbalanced: Butterfly is the most common class (2,930 boxes, 8.2%) while Squid is the rarest (29 boxes, 0.1%) -- roughly a 100x spread. Per-class metrics on the rarest classes are measured on very few examples.

---

Dataset Sources

Original Paper

The Open Images Dataset V4: Unified image classification, object detection, and visual relationship detection at scale

Alina Kuznetsova, Hassan Rom, Neil Alldrin, Jasper Uijlings, Ivan Krasin, Jordi Pont-Tuset, Shahab Kamali, Stefan Popov, Matteo Malloci, Alexander Kolesnikov, Tom Duerig, Vittorio Ferrari

International Journal of Computer Vision, 128:1956-1981, 2020. DOI: 10.1007/s11263-020-01316-z

Official Resources

• Official Website: https://storage.googleapis.com/openimages/web/index.html
• Official License Page: https://storage.googleapis.com/openimages/web/factsfigures_v7.html
• Download Toolkit Used: https://github.com/EscVM/OIDv4_ToolKit

Intermediate Export

• Kaggle re-export: https://www.kaggle.com/datasets/antoreepjana/animals-detection-images-dataset

---

Attribution

All credit for collecting and annotating this data belongs to Google (the Open Images project) and to the individual Flickr photographers who took the original photos. This repository does not modify, reinterpret, or take credit for the underlying imagery or labels -- it only merges per-class-folder duplicates, converts to a YOLO-compatible layout, and adds a validation split.

If you use this dataset in your research, please cite the original publication above.

---

License

Not CC0, despite the intermediate Kaggle re-export's tag. Open Images' own license page states the real terms:

The annotations are licensed by Google LLC under CC BY 4.0 license. The images are listed as having a CC BY 2.0 license, individually verifiable at the source Flickr pages linked from each image's Open Images metadata.

Source: https://storage.googleapis.com/openimages/web/factsfigures_v7.html (License section)

Accordingly:

• Redistribution is explicitly permitted, with attribution, for both the images (CC BY 2.0, per original photographer) and the box annotations (CC BY 4.0, Google) -- which is what this repository does.
• No share-alike obligation; no non-commercial restriction.
• This repository, and any further redistribution of it, should carry the same attribution.

---

Citation

If you use this dataset, please cite:

@article{OpenImages,
  title={The Open Images Dataset V4: Unified image classification, object detection, and visual relationship detection at scale},
  author={Kuznetsova, Alina and Rom, Hassan and Alldrin, Neil and Uijlings, Jasper and Krasin, Ivan and Pont-Tuset, Jordi and Kamali, Shahab and Popov, Stefan and Malloci, Matteo and Kolesnikov, Alexander and Duerig, Tom and Ferrari, Vittorio},
  journal={IJCV},
  volume={128},
  number={7},
  pages={1956--1981},
  year={2020},
  publisher={Springer}
}

---

Acknowledgements

We sincerely thank the Open Images team at Google, the original Flickr photographers whose images make up this dataset, and antoreepjana (Kaggle) and the OIDv4_ToolKit project for the per-class export this repository builds on.