Tanxunze/WhosAya

BanG Dream! Aya Maruyama Detection Model

This repository contains YOLOv8-based object detection models trained to recognize Aya Maruyama, a character from BanG Dream! Girls Band Party!, along with the dataset used for training.

Model Overview

The model is trained to detect Aya Maruyama in various scenarios, including official card illustrations, anime screenshots, and fan artworks.

Training Details

• Base Model: YOLOv8n
• Dataset Size: 747 images
• Training Epochs: 300
• Input Size: 640×640
• Optimizer: Default YOLOv8 optimizer
• Batch Size: 32

Performance

• mAP50: 0.9848
• mAP50-95: 0.6612
• Precision: 0.9869
• Recall: 0.955

Dataset

Dataset Composition

• Total Images: 747
• Split: 70% training, 20% validation, 10% testing
• Image Sources:
  • Official card illustrations from the game
  • Anime screenshots
  • Official promotional materials
  • High-quality fan artworks

Annotations

• Format: YOLO format
• Class: Single class (Aya Maruyama)
• Annotation Type: Bounding boxes around face and upper body

Data Augmentation

• Horizontal Flip
• Rotation: ±15°
• Brightness Adjustment: ±15%
• Saturation: ±25%
• Noise: Up to 0.54% of pixels

Usage

from ultralytics import YOLO

# Load the model
model = YOLO('best.pt')

# Perform prediction
results = model.predict(
    source='image.jpg',
    conf=0.5,
    save=True
)

Limitations

• Model performance may vary on heavily stylized artwork
• Best performance on frontal and side views
• May have reduced accuracy in complex group scenes

Future Improvements

• Increase dataset diversity
• Add more challenging scenarios
• Fine-tune hyperparameters
• Test different model architectures

License

This project is available under MIT License, with respect to BanG Dream!'s copyright and terms of use.

Acknowledgments

Thanks to Bushiroad and Craft Egg for creating BanG Dream! and the character Aya Maruyama(with Amita).