README.md

---
license: cc-by-sa-4.0
task_categories:
  - object-detection
language:
  - en
pretty_name: Street-level Poles & Towers
size_categories:
  - 1K<n<10K
tags:
  - object-detection
  - yolo
  - streetlevel
  - panoramax
  - openstreetmap
  - poles
  - towers
  - infrastructure
configs:
  - config_name: default
    data_files:
      - split: train
        path: data/train/*.parquet
      - split: validation
        path: data/val/*.parquet
      - split: test
        path: data/test/*.parquet
---

# Street-level Poles & Towers

Object detection dataset of utility **poles** and **towers** in street-level
imagery, packaged from the *Object Detection (Berkeley)* working set produced
under the HOT-OSM YOLO experiments. The dataset is shipped in two compatible
layouts:

- **HuggingFace parquet** (default loader): images embedded as bytes, with
  COCO-style bounding boxes (`x, y, w, h` in pixels), class labels, and
  per-image source metadata. Loadable with `datasets.load_dataset(...)`.
- **COCO JSON** (`coco/instances_{train,val,test}.json`) alongside the
  original **YOLO** layout description in `data.yaml`, for native use with
  Ultralytics or any COCO-aware trainer.

## Classes

| id | name  |
|----|-------|
| 0  | pole  |
| 1  | tower |

## Splits

| split | images | positive | negative | boxes | pole boxes | tower boxes |
|-------|-------:|---------:|---------:|------:|-----------:|------------:|
| train | 2,001  | 227      | 1,774    | 338   | 199        | 139         |
| val   | 996    | 501      | 495      | 816   | 505        | 311         |
| test  | 503    | 256      | 247      | 409   | 273        | 136         |
| total | 3,500  | 984      | 2,516    | 1,563 | 977        | 586         |

Note: the train split is heavily dominated by **negative** (no-object)
samples and the positive/negative ratios are not balanced across splits.
This reflects the upstream curation rather than a sampling decision; users
training detectors may want to resample for class balance.

## Geographic coverage

Per-image geographic metadata (latitude, longitude, Panoramax id, source
URL) is available on **1,051 negative samples** sourced from
[panoramax.openstreetmap.fr](https://panoramax.openstreetmap.fr). Coverage
spans roughly Western Europe and East Asia (latitude 24.8 to 53.1,
longitude -4.6 to 121.5). Positive samples generally do not carry
per-image coordinates: their geographic origin is encoded only as a
filename prefix (`bryan_*`, `shanghai_*`, `oakland_*`, `berkeley_*`) or
through Roboflow-style identifiers.

Visualizations available under `artifacts/`:

- `geo_map.html` — interactive map of all geotagged samples.
- `geo_map.png` — static scatter plot of latitude vs longitude.
- `split_class_counts.png` — bounding box counts by class and split.
- `image_counts.png` — image counts by split, broken down positive / negative.
- `boxes_per_image.png` — histogram of boxes per positive image.
- `source_distribution.png` — image counts by source prefix.

## Source provenance

| source prefix | images |
|---------------|-------:|
| bryan         | 3,452  |
| roboflow      |    42  |
| shanghai      |     5  |
| oakland       |     1  |

`bryan_*` is the dominant prefix and includes both positive and Panoramax
sourced negative samples. `roboflow` denotes images stored under the
Roboflow-style identifier scheme (`<hash>_jpg.rf.<token>.jpg`).

## Parquet schema

```python
{
  "image":       Image(decode=True),                          # JPEG bytes
  "image_id":    Value("string"),                             # filename stem
  "width":       Value("int32"),
  "height":      Value("int32"),
  "objects": {
    "bbox":      Sequence(Sequence(Value("float32"), length=4)),   # [x, y, w, h] in pixels (COCO)
    "category":  Sequence(ClassLabel(names=["pole", "tower"])),
  },
  "source":         Value("string"),    # bryan | roboflow | shanghai | oakland
  "panoramax_id":   Value("string"),    # populated for geotagged negatives
  "source_url":     Value("string"),    # original Panoramax derivate URL
  "lat":            Value("float64"),
  "lon":            Value("float64"),
}
```

## Loading

```python
from datasets import load_dataset

ds = load_dataset("kshitijrajsharma/streetlevel-poles")
ds["train"][0]["image"]            # PIL.Image
ds["train"][0]["objects"]          # {"bbox": [...], "category": [...]}
```

For the YOLO / Ultralytics workflow, point at the `data.yaml` shipped with
the dataset after downloading the `coco/` and image folders, or regenerate
the YOLO layout from the parquet shards (helper in this repo).

## Baseline

A short YOLOv11n baseline was trained on this dataset for 10 epochs at
`imgsz=640`. Test split results:

| metric          | value |
|-----------------|------:|
| mAP@50          | 0.332 |
| mAP@50-95       | 0.120 |
| precision       | 0.455 |
| recall          | 0.345 |

See `artifacts/baseline_results.png`,
`artifacts/baseline_confusion_matrix.png`, and
`artifacts/baseline_pr_curve.png` for full curves.

This is intended as a sanity-check baseline, not a state-of-the-art result.
The dataset is small, class-imbalanced, and the train split is dominated by
negatives, so longer training and rebalancing are likely to help.

## Reproducing

```bash
just setup
just extract
just convert
just viz
just parquet
just train       # YOLOv11n, 10 epochs
just upload
```

## License

Released under **CC-BY-SA-4.0**. The negative samples derive from
Panoramax / OpenStreetMap-sourced street-level imagery; share-alike
compatibility is preserved.

## Citation

If you use this dataset, please cite the originating work and this
distribution:

```
@misc{streetlevel-poles,
  title  = {Street-level Poles & Towers},
  author = {Sharma, Kshitij Raj and contributors},
  year   = {2026},
  howpublished = {HuggingFace Datasets},
  url    = {https://huggingface.co/datasets/kshitijrajsharma/streetlevel-poles}
}
```