# Pole Detection Module - PD-Defect

Collection, licensing, and citation are documented on the [main dataset card](../README.md).

## TL;DR

Object detection of distribution poles in full-frame smartphone images. The release contains 2,252 annotated images and 2,425 pole instances.

## Task Definition

Localize poles from top to base in full-frame field images, including scenes that contain more than one pole.

## Label Definitions

- **Pole:** a distribution pole annotated from top to base. The released class name is `Pole` (class 0).

## Annotation Protocol

Only structurally identifiable distribution poles are annotated. Telecommunication poles and other non-distribution vertical structures are not labeled. Bounding boxes run from pole top to base and follow the shaft tightly. Attached components such as crossarms and insulators are included when they overlap the visible pole extent; this differs from [Pole Lean Assessment](../pole_lean_assessment/README.md) object detection, where those fittings are excluded. When the base is only partially visible, annotators estimate the ground-contact location according to protocol. Samples that cannot be interpreted reliably are excluded.

## Annotation Format

YOLO object detection:

```text
<class_id> <center_x> <center_y> <width> <height>
```

Class 0: `Pole`

## Dataset Structure

```text
pole_detection/
├── data.yaml
├── dataset.json
├── train/
│   ├── images/
│   ├── labels/
│   └── train.json
├── valid/
│   ├── images/
│   ├── labels/
│   └── val.json
└── test/
    ├── images/
    ├── labels/
    └── test.json
```

## Train / Validation / Test Split

Train 1,665 (73.93%) / valid 293 (13.01%) / test 294 (13.06%); total 2,252.

## Baseline Benchmarks

YOLO12m (`Pole`). Overall box metrics are given in the split headers; class-wise metrics follow.

**`valid`** — Precision 97.44%; Recall 98.06%; `mAP@0.5` 98.81%; `mAP@0.5:0.95` 88.53%

| Class | Precision | Recall | `mAP@0.5` | `mAP@0.5:0.95` |
| ----- | --------- | ------ | --------- | -------------- |
| Pole  | 97.44%    | 98.06% | 98.81%    | 88.53%         |

**`test`** — Precision 97.33%; Recall 94.36%; `mAP@0.5` 97.08%; `mAP@0.5:0.95` 85.07%

| Class | Precision | Recall | `mAP@0.5` | `mAP@0.5:0.95` |
| ----- | --------- | ------ | --------- | -------------- |
| Pole  | 97.33%    | 94.36% | 97.08%    | 85.07%         |

## Limitations

Base estimation under partial visibility introduces additional uncertainty.

## Examples

Random predictions on the test set. Boxes show predicted class and confidence.

![Pole detection examples](./examples/pole_detection_collage.jpg)

## Citation

Please cite PD-Defect ([DOI 10.6084/m9.figshare.34329402](https://doi.org/10.6084/m9.figshare.34329402)); BibTeX is given on the [main dataset card](../README.md).