Polyp Detection 🩺

This repository contains a complete and reproducible pipeline for benchmarking three object detection architectures on the REAL-Colon dataset:

• YOLOv7
• YOLOv11
• RT-DETR

The goal is to provide a standardized and comparable workflow for evaluating modern detectors in realistic colonoscopy settings.
This repository includes:

1. Dataset Preparation
2. Model Training Configurations
3. Model Evaluation

---

1. Dataset Preparation 📁

This module provides all scripts needed to prepare the REAL-Colon dataset for training YOLOv7, YOLOv11, and RT-DETR. It includes:

• Download and organize dataset files (frames, annotations, CSV metadata)
• Convert annotations to YOLO and COCO formats
• Sample positive and negative frames to create a filtered dataset
• Build final YOLO-ready and RT-DETR-ready datasets

For detailed instructions, scripts, and configuration options, see the Dataset Preparation folder.

---

2. Model Training Configurations 🤖

This module provides the necessary files and instructions to train the three object detection architectures on the REAL-Colon dataset.

Included models:

• YOLOv7: Modified training scripts (train.py) and dataset YAML for colonoscopy-specific training.
• YOLOv11: Dataset configuration and hyperparameters via Ultralytics CLI; no source code modifications required.
• RT-DETR: Configuration files for training on the REAL-Colon dataset.

For detailed instructions, scripts, and configuration options for each model, see the Models folder.

---

3. Model Evaluation 📊

This module includes scripts and notebooks for:

• Evaluating model performance on the test set
• Generating experimental logs and plots
• Comparing YOLOv7, YOLOv11, and RT-DETR in realistic colonoscopy settings

Evaluation results are reproducible following the full pipeline outlined here