Horama/WOW_iNat_annoted

---
license: apache-2.0
task_categories:
- object-detection
tags:
- animals
- wildlife
- detection
- bounding-box
- iNaturalist
- grounding-dino
size_categories:
- 10K<n<100K
---

# WOW iNat Annotated - Animal Detection Dataset

Dataset de detection d'animaux base sur **Horama/WOW_iNat_augmented**, annote automatiquement
avec **Grounding DINO** pour la detection de bounding boxes.

## Source

- Dataset source: `Horama/WOW_iNat_augmented`
- Detection: `IDEA-Research/grounding-dino-base` (Grounding DINO)
- Double-check: `openai/clip-vit-large-patch14` (CLIP)

## Statistiques

| Metrique | Valeur |
|----------|--------|
| **Images totales** | 4,695 |
| **Images avec animal** | 4,645 (98.9%) |
| **Images sans animal** | 50 (1.1%) |

### Detail par split

| Split | Total | Avec animal | Sans animal |
|-------|-------|-------------|-------------|
| `test_prod` | 3,334 | 3,316 | 18 |
| `test_gen` | 1,361 | 1,329 | 32 |

## Splits

- **test_prod**
- **test_gen**

## Colonnes

| Colonne | Type | Description |
|---------|------|-------------|
| `image` | Image | L'image PIL |
| `is_animal` | bool | True si au moins un animal detecte |
| `has_animal` | bool | Alias de `is_animal` |
| `decision` | str | kept_dino, kept_clip, removed_clip, removed_dino |
| `num_detections` | int | Nombre de bounding boxes |
| `objects` | dict | Bounding boxes, confidences, labels |
| `clip_score` | float | Score CLIP animal (null si non verifie) |

### Format des objets

```python
example["objects"] = {
"bbox": [[x_min, y_min, x_max, y_max], ...],
"confidence": [0.85, ...],
"label": ["animal", ...],
}
```

## Utilisation

```python
from datasets import load_dataset

ds_test_prod = load_dataset("Horama/WOW_iNat_annoted", split="test_prod")
ds_test_gen = load_dataset("Horama/WOW_iNat_annoted", split="test_gen")

# Filtrer
animals_only = ds_train.filter(lambda x: x['is_animal'])
```

---

**Cree par [Horama](https://huggingface.co/Horama)**