Prentz/Unified-Animals-Dataset

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