YOLO Training Script Documentation

Overview

This document provides an explanation of how to use the train.py script for training object detection models, specifically YOLOv8 and RT-DETR models, using a custom dataset. The script is implemented in Python and uses the ultralytics library to load and train the models.

Prerequisites

• Python 3.x
• ultralytics library (YOLO and RTDETR classes)
• torch library for PyTorch operations

You can install the required packages via pip:

pip install ultralytics torch

1. Model Loading:

   • YOLO Model:

     model = YOLO('yolov8l.pt')

     • This line loads a pre-trained YOLOv8 model using the pre-trained weights file (yolov8l.pt).
     • You can use different versions of YOLO models (e.g., yolov8s.pt, yolov8m.pt, etc.) by specifying the corresponding .pt file.
     • The YOLO model file (.pt) typically contains the complete model architecture and pre-trained weights.
     • If you want to use a non pre-trained model, use a .yaml file

   • RT-DETR Model:

     model = RTDETR('rtdetr-resnet50.yaml')

     • This line loads the RT-DETR model based on a configuration file (rtdetr-resnet50.yaml).
     • The configuration file defines the model architecture (e.g., backbone, number of layers, etc.) and other settings.

---

How to Run eval.py

Verify Python Installation

Make sure Python 3.x is installed. If it is not installed, you can download it from the official Python website.

Install Required Libraries

Install the required libraries using the command below:

pip install numpy opencv-python

The packages numpy and opencv-python are necessary for tasks such as IoU calculation and coordinate transformation in the script.

Run the Script

You can run the eval.py file by typing the following command in the terminal:

python eval.py

Ensure that the paths for gt_folder and pred_folder are correctly set within the script. These paths should point to the folders containing the ground truth data and prediction data, respectively.

Modify Folder Paths and Image Size

Before running the script, you need to modify the following part of the eval.py file:

gt_folder = '/path/to/groundtruths'  # Path to the folder containing ground truth data
pred_folder = '/path/to/predictions' # Path to the folder containing prediction data
image_width = 2048  # Width of the image
image_height = 1080 # Height of the image

• gt_folder and pred_folder should be set to the paths of the folders containing ground truth labels and predicted labels, respectively.
• image_width and image_height should match the actual size of the images in the dataset.

Output Results

After running the script, it will output metrics such as mAP@50 and mAP@50-95. These values are used to evaluate how accurately the model detects objects.

---