A deep learning-powered computer vision system designed to automatically detect potholes in road imagery using the YOLOv9 object detection model. This project aims to streamline road maintenance operations by enabling automated, accurate pothole identification from images, videos, and live feeds.
We used the following dataset for training and validation:
This dataset includes thousands of annotated images for pothole detection and is ideal for training object detection models.
- Model: YOLOv9 (You Only Look Once - v9)
- Framework: PyTorch
- Input Types: Static images, video files, or webcam streams
- Training Platform: Trained on Modal’s cloud infrastructure with NVIDIA A100 GPU
- ✅ High-speed real-time pothole detection
- ✅ Works with multiple input sources (images, videos, webcam)
- ✅ Easy to retrain with your own datasets
- ✅ Lightweight and deployment-ready
This project uses Modal — a serverless cloud platform — to train and deploy YOLOv9 models in the cloud effortlessly.
🚀 Why Modal? ⚡ Fast GPU compute (NVIDIA A100 for training, T4 for inference)
💾 Persistent volumes for dataset and model storage
💸 Generous free tier ($5 in credits — great for experimenting)
✅ No infrastructure hassle (no Docker, no Kubernetes)
Here are some predictions generated by our model:
🧪 The following metrics and charts summarize the model's evaluation on the validation dataset.
git clone https://github.com/swarnranjan/pothole
cd pothole
pip install -r requirements.txt