Horama/animal-200-detection

Subset of Horama/animal-200 (scraped wildlife images) annotated automatically with Grounding DINO to produce bounding boxes for animal detection.

---

Origin

Property	Value
Source dataset	Horama/animal_scrapped (aka Horama/inat-raw)
Image sources	DuckDuckGo, Wikimedia Commons, Wikipedia
Species coverage	~200 animal species
Pre-classification	CLIP zero-shot (openai/clip-vit-large-patch14)
Filtering	alive=True AND watermark=False only

CLIP pre-classification

Before annotation, every image in Horama/animal_scrapped was classified with CLIP along three axes:

Axis	Positive prompts	Negative prompts	Purpose
alive	"a photo of a living animal in nature/zoo"	dead, skeleton, drawing, taxidermy, statue, no animal…	Discard non-living depictions
watermark	large/intrusive watermark, stock watermark…	clean photo, subtle corner text…	Discard images with obstructive watermarks
distance	close-up / medium / far (3 prompt sets)	—	Control the distance distribution per split

Only images classified as alive and without watermark are kept. The distance label is then used to subsample with distance-aware quotas:

Split	Base images/species	Distance targets
train	230 (+120 for 90 priority species)	40 % close-up, 50 % medium, 10 % far
val	25	33 % / 34 % / 33 % balanced
test	25	33 % / 34 % / 33 % balanced

Close-up images include synthetic crops generated from medium-distance detections (15 % padding around the bounding box).

---

Annotation pipeline

Horama/animal_scrapped
        │
        ▼
  Filter (alive, no watermark)
        │
        ▼
  Subsample per species / distance
        │
        ▼
  Grounding DINO (IDEA-Research/grounding-dino-base)
    prompt: "animal ."
    confidence threshold: 0.25
        │
        ▼
  COCO-format annotations (category 0 = animal)
        │
        ▼
  Push to Horama/animal_annotated

The annotation was parallelized across 4 workers (pods), each processing a disjoint partition of species. Results are stored under worker_0/ through worker_3/.

---

Dataset structure

Horama/animal_annotated/
├── worker_0/
│   ├── annotations/
│   │   ├── train.json          # COCO JSON for this worker's train split
│   │   ├── val.json
│   │   └── test.json
│   ├── images/
│   │   ├── train/
│   │   │   ├── part_0000/      # sharded to stay < 10 000 files per dir
│   │   │   │   ├── Species_name_001.jpg
│   │   │   │   └── ...
│   │   │   └── part_0001/
│   │   ├── val/
│   │   │   └── part_0000/
│   │   └── test/
│   │       └── part_0000/
│   ├── dataset_stats.json
│   └── checkpoint.json
├── worker_1/
│   └── ...
├── worker_2/
│   └── ...
└── worker_3/
    └── ...

Each worker_*/annotations/{split}.json follows the standard COCO Object Detection format:

{
  "images": [
    {"id": 1, "file_name": "part_0000/Species_name_001.jpg", "width": 1024, "height": 768}
  ],
  "annotations": [
    {"id": 1, "image_id": 1, "category_id": 0, "bbox": [x, y, w, h]}
  ],
  "categories": [
    {"id": 0, "name": "animal"}
  ]
}

• bbox format: [x, y, width, height] in absolute pixels (COCO convention)
• category_id: always 0 (animal) — person annotations are added later during merge
• Images are sharded into part_XXXX/ sub-directories (max 9 000 files each)

---

Workers

4 workers processed species in parallel on separate compute pods:

Worker	Prefix	Species partition
0	worker_0/	Species 0, 4, 8, 12, …
1	worker_1/	Species 1, 5, 9, 13, …
2	worker_2/	Species 2, 6, 10, 14, …
3	worker_3/	Species 3, 7, 11, 15, …

Species are assigned round-robin by alphabetical index, so each worker handles a roughly equal share of the ~200 species.

---

Model details

Property	Value
Model	IDEA-Research/grounding-dino-base
Prompt	"animal ."
Confidence threshold	0.25
Framework	HuggingFace Transformers (AutoModelForZeroShotObjectDetection)
Inference	torch.no_grad(), best device (CUDA > MPS > CPU)

---

License

This dataset is released under AGPL-3.0, which means it can be freely used, modified and redistributed as long as all derivative work remains open-source and publicly available under the same license.

If you need to use this dataset in a closed-source or commercial context, please contact us to discuss a commercial license arrangement.

---

Limitations

• Annotations are automatically generated (no manual review) — expect noise, missed detections on unusual poses/species, and occasional false positives.

• The confidence threshold of 0.25 favors recall over precision.

• Some species may have fewer images than the target quota if the source dataset had insufficient qualifying images (alive, no watermark).

  ---

Citation

@misc{horama_animal200_detection_2026,
  title   = {Animal-200-Detection: Grounding DINO Annotated Wildlife Images (200 Species)},
  author  = {Horama},
  year    = {2026},
  url     = {https://huggingface.co/datasets/Horama/animal-200-detection},
  note    = {Bounding box annotations generated with Grounding DINO for animal detection across 200 species}
}