---
license: cc-by-sa-4.0
library_name: ultralytics
pipeline_tag: object-detection
tags:
  - yolo
  - yolov8
  - ultralytics
  - object-detection
  - poles
  - towers
  - streetlevel
  - openstreetmap
  - infrastructure
datasets:
  - hotosm/streetlevel-poles
base_model: Ultralytics/YOLOv8
---

# streetlevel-poles-yolov8s

YOLOv8s object detector for utility **poles** and **towers** in street-level
imagery. Trained on the
[Street-level Poles & Towers](https://huggingface.co/datasets/hotosm/streetlevel-poles)
dataset as part of the HOT-OSM YOLO experiments (curated by UC Berkeley Code For Good cohort).

This is the **50-epoch YOLOv8s** model from the original project deliverable.
For a lightweight 10-epoch sanity-check baseline, see the related
[`kshitijrajsharma/streetlevel-poles-yolo11n`](https://huggingface.co/kshitijrajsharma/streetlevel-poles-yolo11n).

## Classes

`0: pole`, `1: tower`

## Metrics

Best epoch (39 of 49; early-stopped from a 50-epoch schedule, patience 10),
**validation split**, `imgsz=640`:

| metric    | value |
|-----------|------:|
| mAP@50    | 0.685 |
| mAP@50-95 | 0.278 |
| precision | 0.672 |
| recall    | 0.686 |

> **Note on comparability.** These are validation-split metrics logged by
> Ultralytics during training. The related yolo11n baseline reports
> **test-split** metrics (mAP@50 = 0.332), so the two numbers are **not**
> directly comparable. A matched test-split evaluation of this model can be
> added - see `results.csv` for the full per-epoch training history.

Artifacts: `results.png`, `confusion_matrix.png`,
`confusion_matrix_normalized.png`, `pr_curve.png`, `f1_curve.png`,
`val_pred_sample.jpg`, `results.csv`.

## Training

- Base weights: `yolov8s.pt` (Ultralytics)
- Epochs: 50 configured, stopped at 49 (early stopping, patience 10), best epoch 39
- Image size: 640, batch: 16, seed: 42
- Full training code: `training_notebook.ipynb` (original HOT-OSM notebook)

## Files

| file | description |
|------|-------------|
| `best.pt` | best-epoch weights (recommended for inference) |
| `last.pt` | final-epoch weights |
| `metrics.json` | training config + best-epoch validation metrics |
| `data.yaml` | Ultralytics dataset config (class names) |
| `results.csv` | full per-epoch training/validation metrics |
| `training_notebook.ipynb` | end-to-end training notebook |
| `*.png`, `val_pred_sample.jpg` | evaluation plots and a sample prediction |

## Inference

```python
from huggingface_hub import hf_hub_download
from ultralytics import YOLO

weights = hf_hub_download(
    repo_id="hotosm/streetlevel-poles-yolov8s",
    filename="best.pt",
)
model = YOLO(weights)
results = model("street_image.jpg")
results[0].show()
```

## Related repositories

- Dataset: [`hotosm/streetlevel-poles`](https://huggingface.co/datasets/hotosm/streetlevel-poles)
- Lightweight baseline (10-epoch yolo11n): [`kshitijrajsharma/streetlevel-poles-yolo11n`](https://huggingface.co/kshitijrajsharma/streetlevel-poles-yolo11n)
- Demo Space (runs the yolo11n baseline): [`kshitijrajsharma/streetlevel-poles-detector`](https://huggingface.co/spaces/kshitijrajsharma/streetlevel-poles-detector)

## License

Released under **CC-BY-SA-4.0**, matching the dataset. Training imagery derives
from Panoramax / OpenStreetMap-sourced street-level images; share-alike
compatibility is preserved.