# Perspective Validation Module - PD-Defect

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

## TL;DR

Binary classification of full images, pole-top crops, and DTR scenes as Valid or Perspective-Distorted. The combined release contains 4,950 samples across three subtasks. The task concerns capture viewpoint only, not lens distortion, image sharpness, object completeness, or defect presence. Borderline cases lie on a continuum and are forced into a binary label.

## Task Definition

Classify each input as geometrically Valid or Perspective-Distorted for the analysis task that depends on that region (lean assessment, component tilt assessment, or distribution-transformer-to-ground clearance).

Images rejected for perspective distortion in the inspection datasets were retained in this release, including the validation and test splits, and were used to train the DINOv3 perspective-validation models.

## Label Definitions

- **Valid** (folder `accepted`): the viewpoint preserves the geometric relationship required by the corresponding analysis task.
- **Perspective-Distorted** (folder `rejected_perspective_issue`): the viewpoint compromises that relationship, so structural interpretation from the image would be unreliable.



## What Perspective Distortion Means in Each Subtask



### Full-image (Pole Lean Assessment)

The pole must be photographed from a viewpoint facing it directly so that its imaged axis reflects true vertical alignment. Under an oblique angle, an upright pole can appear to lean and a leaning pole can appear upright.

### Pole-top crops (Crossarm and Top-Cleat Tilt Assessment)

The pole-top region must be viewed close to head-on so that Crossarm and Top Cleat orientation in the image reflects true structure. Foreshortening under an oblique view can reverse Straight and Tilted appearance. A Valid full-image label does not imply a Valid crop: the pole top occupies only a small fraction of a full frame.

### DTR (plinth–ground clearance)

The pole and plinth must stand at approximately the same distance from the camera so that their relative image heights remain comparable.

## Subtasks



### 1) Full-image

- Input: full-frame smartphone images
- Size: 2,169 images (1,018 Valid, 1,151 Perspective-Distorted)
- Split: train 1,605 (74.00%) / val 282 (13.00%) / test 282 (13.00%)



### 2) Pole-top crops

- Input: pole-top crops from the same crop family as Crossarm and Top-Cleat Tilt Assessment and Voltage Level Classification
- Size: 1,743 images (945 Valid, 798 Perspective-Distorted)
- Split: train 1,221 (70.05%) / val 260 (14.92%) / test 262 (15.03%)



### 3) DTR (`plinth_dtr/`)

- Input: DTR pole–plinth images
- Size: 1,038 images (438 Valid, 600 Perspective-Distorted)
- Split: train 768 (73.99%) / val 135 (13.01%) / test 135 (13.01%)



## Data Acquisition and Labeling Protocol

Perspective labels are assigned during curation from geometric interpretability criteria. The Perspective-Distorted class is supplemented by pathway rejects and by other dedicated or repurposed field captures:

- **Full-image:** frames excluded from Pole Lean Assessment object detection for perspective distortion (also labeled `Rejected` in that module's classification subset), plus other dedicated or repurposed captures
- **Pole-top crops:** crops that fail the Crossarm and Top-Cleat Tilt Assessment pole-top perspective screen
- **DTR:** pole–plinth perspective screening rejects

Visibility-only `Rejected` images from Pole Lean Assessment are not placed in perspective validation.

## Dataset Structure

```text
perspective_validation/
├── full_image/
│   ├── dataset.json
│   ├── labels.csv
│   ├── train/
│   │   ├── accepted/
│   │   ├── rejected_perspective_issue/
│   │   ├── train.csv
│   │   └── train.json
│   ├── val/
│   │   ├── accepted/
│   │   ├── rejected_perspective_issue/
│   │   ├── val.csv
│   │   └── val.json
│   └── test/
│       ├── accepted/
│       ├── rejected_perspective_issue/
│       ├── test.csv
│       └── test.json
├── pole_top_crops/          # same layout as full_image/
└── plinth_dtr/              # same layout as full_image/
```

All three subtasks use `val/` for the validation split.

## Baseline Benchmarks

DINOv3 ViT-B/16 (`facebook/dinov3-vitb16-pretrain-lvd1689m`), fine-tuned independently on each subtask. Split Accuracy is overall. Class-wise Accuracy, Precision, Recall, F1, and support (`n`) are from the same evaluation. Class-wise Accuracy is `correct / n` and equals Recall. Macro-F1 is the unweighted mean of the class-wise F1 scores.

### Full-image

`val` **(282)** — Accuracy 89.36% (252/282); Macro-F1 89.33%


| Class                      | Accuracy | Precision | Recall | F1     | `n` |
| -------------------------- | -------- | --------- | ------ | ------ | --- |
| accepted                   | 89.47%   | 88.15%    | 89.47% | 88.81% | 133 |
| rejected_perspective_issue | 89.26%   | 90.48%    | 89.26% | 89.86% | 149 |


`test` **(282)** — Accuracy 90.43% (255/282); Macro-F1 90.41%


| Class                      | Accuracy | Precision | Recall | F1     | `n` |
| -------------------------- | -------- | --------- | ------ | ------ | --- |
| accepted                   | 91.67%   | 88.32%    | 91.67% | 89.96% | 132 |
| rejected_perspective_issue | 89.33%   | 92.41%    | 89.33% | 90.85% | 150 |




### Pole-top crops

`val` **(260)** — Accuracy 95.00% (247/260); Macro-F1 94.95%


| Class                      | Accuracy | Precision | Recall | F1     | `n` |
| -------------------------- | -------- | --------- | ------ | ------ | --- |
| accepted                   | 97.16%   | 93.84%    | 97.16% | 95.47% | 141 |
| rejected_perspective_issue | 92.44%   | 96.49%    | 92.44% | 94.42% | 119 |


`test` **(262)** — Accuracy 96.56% (253/262); Macro-F1 96.54%


| Class                      | Accuracy | Precision | Recall | F1     | `n` |
| -------------------------- | -------- | --------- | ------ | ------ | --- |
| accepted                   | 97.18%   | 96.50%    | 97.18% | 96.84% | 142 |
| rejected_perspective_issue | 95.83%   | 96.64%    | 95.83% | 96.23% | 120 |




### DTR

`val` **(135)** — Accuracy 91.11% (123/135); Macro-F1 90.89%


| Class                      | Accuracy | Precision | Recall | F1     | `n` |
| -------------------------- | -------- | --------- | ------ | ------ | --- |
| accepted                   | 89.47%   | 89.47%    | 89.47% | 89.47% | 57  |
| rejected_perspective_issue | 92.31%   | 92.31%    | 92.31% | 92.31% | 78  |


`test` **(135)** — Accuracy 85.19% (115/135); Macro-F1 84.89%


| Class                      | Accuracy | Precision | Recall | F1     | `n` |
| -------------------------- | -------- | --------- | ------ | ------ | --- |
| accepted                   | 84.21%   | 81.36%    | 84.21% | 82.76% | 57  |
| rejected_perspective_issue | 85.90%   | 88.16%    | 85.90% | 87.01% | 78  |




## Limitations

Binary labels do not encode fine-grained perspective quality, and borderline cases are rejected conservatively.

## Examples

Random predictions on each subtask's test set. Green **P** / **T** denote prediction and ground truth; red marks an error.

![Full-image perspective validation examples](./examples/full_image.png)

![Pole-top crop perspective validation examples](./examples/pole_top_crops.png)

![DTR plinth perspective validation examples](./examples/plinth_dtr.png)

## Citation

Please cite PD-Defect ([DOI 10.5281/zenodo.18074045](https://doi.org/10.5281/zenodo.18074045)); BibTeX is given on the [main dataset card](../README.md).