YOLOv12 Comic Panel Detection

📌 Notebook moved to Kaggle:
👉 https://www.kaggle.com/code/themosesab/finetuning-yolov12-comic-panel-detection

This repository contains a complete workflow for fine-tuning a YOLOv12 model to detect panels in comic book pages. The project uses the powerful Ultralytics library and a custom dataset managed on Roboflow.

The primary goal is to demonstrate an end-to-end process: from setting up the environment and loading a pre-trained model to training, validating, and performing inference on new images.

Visit this space to try out the model right now: The_Best_Comic_Panel_Detection.

Key Features

• State-of-the-Art Model: Fine-tunes the yolo12x model, the largest and most performant variant of the YOLOv12 series.
• Custom Dataset: Integrates with Roboflow for easy dataset download, versioning, and management.
• End-to-End Workflow: The entire process is detailed in a single Jupyter Notebook (comic-boundary-detection.ipynb).
• Comprehensive Evaluation: Generates and displays key performance metrics, including loss curves, confusion matrices, and precision-recall curves.
• Reproducibility: Aims for easy setup and execution, allowing others to replicate the results.

File Structure

The repository is organized as follows:

.
├── comic-boundary-detection.ipynb  # The main notebook with all the code
├── my_yolo_train/                  # Directory for all training outputs
│   ├── demonstration/              # Output from the initial model test
│   ├── finetuning/                 # All artifacts from the fine-tuning process
│   │   ├── args.yaml               # Training configuration
│   │   ├── results.csv             # Training metrics per epoch
│   │   ├── results.png             # Plot of training/validation loss & metrics
│   │   ├── confusion_matrix.png    # Performance visualization
│   │   └── weights/
│   │       ├── best.pt             # Best performing model checkpoint (recommended for inference)
│   │       └── last.pt             # Model checkpoint from the final epoch
│   ├── prediction/                 # Inference output on a sample test image
│   └── validation/                 # Validation results on the test set
└── ...                             # Other project files

Training Results

The model was trained for 200 epochs, achieving high performance on the validation set. Below are the key training metrics over time.

Val Pred Visualization

Model Checkpoints

The trained model weights are saved in the my_yolo_train/finetuning/weights/ directory.

• best.pt: The model checkpoint with the best mAP50-95 score on the validation set. This is the recommended model for inference.
• last.pt: The model checkpoint saved at the end of the final training epoch.

Acknowledgements

• Ultralytics for the amazing YOLOv12 model and library.
• Roboflow: for their dataset hosting platform and custom-workflow-3-object-detection-g24r5-fmfkb for compiling and annotating this incredible dataset.

License

This project is licensed under the MIT License. See the LICENSE file for details.