RoadEye - Unified Pothole Detection Model
Overview
RoadEye is a unified pothole detection system designed for real-world road condition analysis using deep learning. It supports multiple architectures including YOLOv8 and RF-DETR, and integrates road segmentation for improved inference accuracy.
The model is trained on a merged dataset combining multiple sources (Roboflow, Kaggle, Mendeley) and supports end-to-end pipelines for training, evaluation, and deployment.
Model Variants
Medium Model
- Training + Inference: https://drive.google.com/drive/folders/1HoBO5hgAdMVwkSq8ort8UezmWEZDyf5E?usp=sharing
Large Model
- Training + Inference: https://drive.google.com/drive/folders/11Z4Ri7jdXBfsuCX4G6Ahdo7UW7JDeQNj?usp=drive_link
Inference Results
- Images and Videos: https://drive.google.com/drive/folders/1A9bsHYJtYSah6HvPjLE9KpoYdbDtrors?usp=sharing
Final Inference Pipeline
- Colab Notebook (SegFormer + ROI + Detection): https://colab.research.google.com/drive/1K62OVa3uZtw51wlQMypFd_v6qdIyKvYC?usp=sharing
This pipeline includes:
- Road segmentation using SegFormer (Cityscapes)
- Region of Interest (ROI) filtering
- Pothole detection using trained models
Datasets
RF-DETR Dataset: https://huggingface.co/datasets/AkumaDachi/RoadEye_RFDETR_Aug
YOLO Dataset: https://huggingface.co/datasets/AkumaDachi/RoadEye_Yolo_Aug
Repository
Full System (Frontend + Backend + App)
https://github.com/astralranger/road-eye
Training & Augmentation Scripts
This repository contains scripts for:
- Dataset augmentation
- Model training (YOLOv8, RF-DETR)
- Validation and inference pipelines
Intended Use
- Autonomous driving research
- Road condition monitoring
- Smart city infrastructure
- Edge AI deployment for traffic systems
Limitations
- Performance may vary under extreme lighting or weather conditions
- Dataset is aggregated from multiple sources with varying annotation quality
- Not optimized for real-time deployment on low-end hardware without optimization
License
This project uses a combination of datasets from multiple sources (Roboflow, Kaggle, Mendeley, CC-BY-4.0). Hence, the overall license is marked as "other". Please refer to individual dataset licenses for details.
Citation
If you use this work, please consider citing:
@misc{roadeye2026,
title={RoadEye: Unified Pothole Detection System},
author={AkumaDachi},
year={2026}
}