YOLOv8 Pothole Detection

Overview

This project utilizes YOLOv8 to detect potholes in images and videos. It involves dataset preparation, model training, evaluation, and inference on test data.

Prerequisites

Ensure you have the following installed:

• Python 3.10+
• PyTorch 2.3.1+cu121
• NVIDIA GPU with CUDA 12.2 (recommended)

Installation

1. Clone the repository (if applicable).
2. Install required dependencies:

   pip install ultralytics==8.0.0 roboflow

3. Check if Ultralytics YOLOv8 is correctly installed:

   import ultralytics
   ultralytics.checks()

Dataset Preparation

1. Create a dataset directory:

   mkdir -p /content/datasets

2. Download the dataset from Roboflow:

   from roboflow import Roboflow
   rf = Roboflow(api_key="YOUR_API_KEY")
   project = rf.workspace("YOUR_NAME").project("yolov8pothole")
   version = project.version(1)
   dataset = version.download("yolov5")

Training the YOLOv8 Model

Train the model using the following command:

cd /content
yolo task=detect mode=train model=yolov8m.pt data=/content/Yolov8Pothole-1/data.yaml epochs=70 imgsz=640

Model Evaluation

1. Confusion Matrix:

   from IPython.display import Image
   Image(filename='/content/runs/detect/train2/confusion_matrix.png', width=900)

2. Training and Validation Loss:

   Image(filename='/content/runs/detect/train2/results.png', width=600)

3. Validation Batch Predictions:

   Image(filename='/content/runs/detect/train2/val_batch0_pred.jpg', width=600)

4. Validate with best weights:

   yolo task=detect mode=val model=/content/runs/detect/train2/weights/best.pt data=/content/datasets/Yolov8Pothole-1/data.yaml

Inference on Test Data

yolo task=detect mode=predict model=/content/runs/detect/train2/weights/best.pt conf=0.25 source=/content/datasets/Yolov8Pothole-1/test/images

Testing on a Demo Video

cp "/content/drive/MyDrive/Pothole Detect/demo2.mp4" .
yolo task=detect mode=predict model=/content/runs/detect/train2/weights/best.pt conf=0.25 source='/content/demo2.mp4'

Display the Demo Video

Watch the demo video

from IPython.display import HTML
from base64 import b64encode
import os

# Input video path
save_path = '/content/runs/detect/predict3/demo2.mp4'
compressed_path = "/content/result_compressed2.mp4"

os.system(f"ffmpeg -i {save_path} -vcodec libx264 {compressed_path}")

# Show video
mp4 = open(compressed_path,'rb').read()
data_url = "data:video/mp4;base64," + b64encode(mp4).decode()
HTML(f"""
<video width=600 controls>
  <source src='{data_url}' type='video/mp4'>
</video>
""")

Troubleshooting

• Ensure all paths exist before executing commands.
• Check that your dataset is downloaded correctly.
• Ensure your GPU is available and properly set up with CUDA.

---

This README provides step-by-step instructions for setting up, training, and evaluating a pothole detection model using YOLOv8. Let me know if you need further refinements!