Road Damage Detection using YOLOv8

Overview

This project implements an object detection system for automated road damage detection using given dataset in competition Crackathon. The model detects and classifies five types of road surface damage using YOLOv8.

Dataset

The competition uses a randomized version of the RDD2022 dataset with the following structure:

Label Format

Labels are provided in YOLO format: <class_id> <x_center> <y_center>

Class Mapping

0 Longitudinal Crack 1 Transverse Crack 2 Alligator Crack 3 Other Corruption 4 Pothole

Model

• YOLOv8 (Ultralytics)
• Transfer learning using pretrained weights

Training

yolo detect train \
  model=yolov8s.pt \
  data=data.yaml \
  epochs=100 \
  imgsz=640 \
  batch=8

Inference
yolo detect predict \
  model=best.pt \
  source=randomized_dataset/test/images \
  save_txt=True \
  save_conf=True \
  conf=0.001

Training and inference can be reproduced using `scripts/train.py` and `scripts/predict.py`.


Submission
Prediction .txt files are generated for each test image and placed in a predictions/ folder. All files are compressed into submission.zip for final submission.
Hardware
Google Colab (Free GPU)
Compatible with 8GB RAM systems

## 🔗 Source Code

This repository contains the full training and inference pipeline used in the competition.

To reproduce results:
1. Install dependencies from `requirements.txt`
2. Open `notebooks/train_yolov8.ipynb`
3. Follow the cells sequentially

The model was trained using the RDD2022 dataset in YOLO format.