Prentz/Unified-Animals-Dataset

---
license: apache-2.0
language:
- en
tags:
- wildlife
- Animals
- image
- classification
pretty_name: Unified Animals Dataset
size_categories:
- 1K<n<10K
---

# Unified Animal Dataset

This dataset contains a large-scale collection of animal images designed for multi-class image classification tasks. It includes **57,329 images across 216 animal categories** and is suitable for training, evaluation, and benchmarking of computer vision models.

---

## Dataset Details

### Dataset Description

The Unified Animal Dataset is a curated multi-source dataset combining several publicly available animal image datasets into a unified classification benchmark. It spans a wide range of animal types including mammals, birds, insects, and marine species.

The dataset is pre-structured into training, validation, and test splits, making it directly usable for machine learning workflows.

* **Curated by:** Prentz
* **Shared by:** Prentz
* **License:** Apache 2.0

---

### Dataset Sources

The dataset is composed from multiple sources including:

* Stanford Dogs dataset
* Ultralytics datasets
* Additional animal datasets sourced from Kaggle

---

## Uses

### Direct Use

This dataset is intended for:

* Multi-class animal image classification
* Training and fine-tuning deep learning models
* Transfer learning experiments
* Benchmarking computer vision architectures
* Educational purposes in machine learning

---

### Out-of-Scope Use

This dataset is not suitable for:

* Human-related tasks (e.g., facial recognition, demographic prediction)
* Object detection or segmentation tasks (no bounding box annotations)
* Applications requiring balanced datasets without preprocessing

---

## Dataset Structure

The dataset follows a standard image classification folder structure:

```
data/
train/
val/
test/
class_1/
class_2/
...
```

### Statistics

* **Total Images:** 57,329
* **Total Classes:** 216
* **Dataset Size:** 4.14 GB

### Split Distribution

| Split | Images | Percentage |
| ---------- | ------ | ---------- |
| Train | 28,637 | 50.0% |
| Validation | 11,415 | 19.9% |
| Test | 17,277 | 30.1% |

### Class Distribution

* Average images per class: ~265
* Most represented class: ~4,800+ images
* Least represented classes: ~60 images
* Imbalance ratio: ~81×

---

## Dataset Creation

### Curation Rationale

The dataset was created to provide a large and diverse benchmark for animal classification tasks, combining multiple datasets into a single unified structure. The goal was to enable efficient experimentation and model development without requiring users to manually merge datasets.

---

### Source Data

#### Data Collection and Processing

Data was collected from multiple publicly available datasets and organized into a unified structure. The processing pipeline included:

* Merging datasets into a consistent directory format
* Cleaning corrupted or unreadable images
* Ensuring class consistency across splits
* Splitting into train, validation, and test sets

#### Source Data Producers

The original data was created by multiple dataset providers, including academic datasets and open-source contributors.

---

### Annotations

#### Annotation Process

Annotations are provided as **class labels via directory structure**. Each image belongs to exactly one class, determined by its folder.

#### Annotators

Annotations originate from the original datasets used (e.g., Stanford Dogs and others). No additional manual relabeling was performed beyond dataset organization.

#### Personal and Sensitive Information

The dataset does not contain personal or sensitive human data. It consists solely of animal images.

---

## Bias, Risks, and Limitations

* **Class imbalance:** Significant imbalance (~81×) between largest and smallest classes
* **Resolution variability:** Images vary widely in size and quality
* **Dataset bias:** Over-representation of certain animals (e.g., dogs, spiders)
* **Domain limitation:** Performance may degrade on out-of-distribution images

---

### Recommendations

Users should consider:

* Applying class weighting or focal loss
* Using data augmentation techniques
* Standardizing image resolution (e.g., 224×224)
* Evaluating per-class performance, not just overall accuracy

---

## Citation

**BibTeX:**

```
@dataset{Unified_animals_2026,
title={Unified Animal Dataset (216 Classes)},
author={Prentz},
year={2026}
}
```

**APA:**

Furly. (2026). *Unified Animal Dataset (216 Classes).*

---

## More Information

This dataset is intended for practical machine learning workflows and experimentation. Users should carefully handle preprocessing and class imbalance when training models.

---

## Dataset Card Authors

Prentz (author)

---

## Dataset Card Contact

none