Pothole Detection (YOLO11n)
YOLO11n fine-tuned to detect potholes in road images. One class: pothole.
Code, training script and full details: https://github.com/SavanPatel02/pothole-detection-yolo11
Usage
pip install ultralytics huggingface_hub
from huggingface_hub import hf_hub_download
from ultralytics import YOLO
model = YOLO(hf_hub_download("SPatel02/pothole-detection-yolo11", "best.pt"))
results = model.predict("road.jpg", conf=0.25)
results[0].show()
Results
30 epochs at 640 px (auto batch size = 11, seed 42). From results/metrics.json:
| Split | Images | Precision | Recall | mAP50 | mAP50-95 |
|---|---|---|---|---|---|
| Validation | 133 | 0.807 | 0.673 | 0.762 | 0.473 |
| Test | 67 | 0.771 | 0.727 | 0.792 | 0.503 |
Inference: 3.44 ms per image (plus 0.87 ms preprocess and 0.60 ms postprocess), averaged over 10 test images on an NVIDIA RTX 4060 Laptop GPU.
Sample outputs
Held-out test images, confidence 0.25. The last one includes a false positive (traffic cone).
Training data
Pothole dataset v1 on Roboflow Universe: 665 images (465 train / 133 valid / 67 test), created by Atikur Rahman Chitholian. License: ODbL v1.0. The dataset is not included here.
Limitations
- Small training set (465 images).
- Misses many small or distant potholes (validation recall 0.67).
- False positives on objects such as traffic cones.
- Mostly daytime images; night and rain performance not evaluated.
License
Fine-tuned from Ultralytics YOLO11n, which is licensed AGPL-3.0, so these weights are released under AGPL-3.0.
- Downloads last month
- 24