N-o-1/pokemon-images

Pokemon Competitive Team Dataset

A curated image dataset featuring 6 specific Pokemon commonly used in competitive play: Arceus, Marshadow, Sandy Shocks, Slaking, Reshiram, and Magearna. This dataset is designed for training computer vision models to recognize these Pokemon in various contexts.

Dataset Overview

Pokemon	Image Count	Percentage
Arceus	644	49.7%
Magearna	200	15.4%
Slaking	152	11.7%
Reshiram	118	9.1%
Marshadow	101	7.8%
Sandy Shocks	75	5.8%
Total	1,290	100%

Pokemon Details

Arceus

• Type: Normal (can change with Plates/Z-Crystals)
• Generation: IV (Diamond/Pearl)
• Role: Versatile support/offensive Pokemon
• Images: 644 (largest class - includes various formes)

Marshadow

• Type: Fighting/Ghost
• Generation: VII (Sun/Moon)
• Role: Physical sweeper with unique typing
• Images: 101

Sandy Shocks

• Type: Electric/Ground
• Generation: IX (Scarlet/Violet)
• Role: Paradox Pokemon, special attacker
• Images: 75 (smallest class - newer Pokemon)

Slaking

• Type: Normal
• Generation: III (Ruby/Sapphire)
• Role: High-power attacker with Truant ability
• Images: 152

Reshiram

• Type: Dragon/Fire
• Generation: V (Black/White)
• Role: Legendary special attacker
• Images: 118

Magearna

• Type: Steel/Fairy
• Generation: VII (Sun/Moon)
• Role: Support/tank with Soul-Heart ability
• Images: 200

Image Characteristics

File Formats

• Supported: JPG, JPEG, PNG
• Primary: JPG (70%), PNG (30%)

Image Sources

• Official Pokemon artwork
• Game screenshots (various Pokemon games)
• Trading card game artwork
• Anime screenshots
• Fan art (high-quality, recognizable)

Image Quality

• Resolution: Varies (typically 200x200 to 1920x1080)
• Aspect Ratios: Mixed (square, 16:9, 4:3, portrait)
• Content: Pokemon-focused with various backgrounds
• Lighting: Natural variety from different sources

Content Variety

• Poses: Multiple angles and positions
• Contexts: Battle scenes, portraits, environment shots
• Styles: Official art, game renders, anime style, realistic interpretations
• Backgrounds: Transparent, solid colors, natural environments, battle arenas

Dataset Characteristics

Class Imbalance

The dataset exhibits significant class imbalance:

• Most represented: Arceus (49.7% of total)
• Least represented: Sandy Shocks (5.8% of total)
• Imbalance ratio: 8.6:1 (Arceus vs Sandy Shocks)

Balanced Training Strategy

When training models on this dataset, we recommend:

• Balanced sampling: Use equal samples per class per epoch
• Weighted loss functions: Account for class imbalance
• Data augmentation: Especially for underrepresented classes
• Stratified splits: Maintain class ratios in train/val/test splits

Recommended Usage

Training

# Balanced sampling approach
samples_per_class = 75  # Based on smallest class (Sandy Shocks)

Data Augmentation

Recommended augmentations for this dataset:

• Horizontal flips (Pokemon can face either direction)
• Rotation (±15 degrees)
• Color jitter (brightness, contrast, saturation)
• Random crops and resizing
• Avoid: Vertical flips (Pokemon don't appear upside down)

Validation Strategy

• Stratified sampling: Maintain class proportions
• Temporal split: If timestamps available, use chronological splits
• Balanced metrics: Use balanced accuracy, not raw accuracy

Technical Specifications

File Organization

• Each Pokemon has its own subdirectory
• Consistent naming convention within directories
• No duplicate images across classes
• Clean filenames (no special characters)

Data Quality

• Manually curated: All images verified for correct Pokemon
• Deduplication: Removed obvious duplicates
• Quality filtering: Excluded very low resolution or corrupted images
• Labeling accuracy: 100% (single Pokemon per image)

Potential Challenges

Class Imbalance

• Standard accuracy metrics may be misleading
• Model may be biased toward Arceus
• Requires careful sampling strategy

Visual Similarity

• Some Pokemon share similar color schemes
• Legendary Pokemon may have similar poses
• Steel-type Pokemon (Magearna) may share metallic features

Context Variation

• Wide variety of backgrounds and contexts
• Different art styles may confuse models
• Lighting and angle variations

Evaluation Metrics

For this dataset, use:

• Balanced Accuracy: Accounts for class imbalance
• Per-class Precision/Recall: Individual Pokemon performance
• Confusion Matrix: Identify misclassification patterns
• F1-scores: Harmonic mean of precision/recall

Use Cases

Primary Applications

• Pokemon recognition models: Computer vision training
• Competitive analysis: Team composition recognition
• Content filtering: Pokemon-specific content moderation
• Educational tools: Pokemon identification applications

Research Applications

• Class imbalance handling: Testing balancing techniques
• Transfer learning: Fine-tuning pre-trained models
• Multi-class classification: Benchmark dataset for 6-class problems

Data Collection Methodology

1. Pokemon Selection: Chosen based on competitive viability and team synergy
2. Source Diversification: Multiple art styles and contexts
3. Quality Control: Manual verification of each image
4. Deduplication: Automated and manual duplicate removal
5. Organization: Systematic file naming and directory structure

Licensing and Attribution

Dataset License

This dataset is provided under the Creative Commons Attribution 4.0 (CC BY 4.0) license for research and educational purposes.

Pokemon Copyright

• Pokemon characters are © The Pokémon Company/Nintendo
• This dataset is for non-commercial research/educational use
• Images sourced from publicly available content
• Fair use applies for research and educational purposes

Attribution

If you use this dataset, please cite:

@dataset{pokemon_team_dataset,
  title={Pokemon Competitive Team Dataset},
  author={Steven Van Ingelgem},
  year={2025},
  url={https://huggingface.co/datasets/your-username/pokemon-team-dataset},
  note={6-class Pokemon image dataset for computer vision research}
}

Download and Usage

Requirements

• Python 3.8+
• PIL/Pillow for image loading
• OpenCV (optional, for advanced preprocessing)

Loading Example

from pathlib import Path
from PIL import Image

# Load dataset
dataset_path = Path("images")
pokemon_names = ["arceus", "marshadow", "sandy-shocks", "slaking", "reshiram", "magearna"]

# Example: Load all Arceus images
arceus_images = []
arceus_dir = dataset_path / "arceus"
for img_path in arceus_dir.glob("*"):
    if img_path.suffix.lower() in {'.jpg', '.jpeg', '.png'}:
        image = Image.open(img_path)
        arceus_images.append(image)

print(f"Loaded {len(arceus_images)} Arceus images")

Version History

• v1.0: Initial release with 1,290 images across 6 Pokemon classes
• Focus on competitive Pokemon team composition
• Curated for computer vision model training

Contact

For questions, issues, or contributions to this dataset, please contact:

• Author: Steven Van Ingelgem
• Email: [email protected]
• GitHub: https://github.com/svaningelgem

---

Disclaimer: This dataset is intended for research and educational purposes. All Pokemon-related content is the property of The Pokémon Company and Nintendo. Use in accordance with fair use guidelines and applicable copyright laws.