PD-Defect: A Field-Captured Smartphone Image Dataset for AI-Based Inspection of Power Distribution Infrastructure

TL;DR

PD-Defect is a field-captured smartphone image dataset for computer vision on power distribution infrastructure, collected with APEPDCL in Eastern Andhra Pradesh, India. The release comprises 20,476 annotated examples across object detection, image classification, keypoint detection, semantic segmentation, and perspective validation. This is the sum of the subset rows below. The same field photograph can appear in more than one row, for example once as a detection image and again as a keypoint image, or once as a pole-top crop and again in perspective validation. Annotations use YOLO boxes, YOLO pose, folder or CSV labels, and COCO polygons. The dataset is released under CC BY 4.0.

---

Key Statistics

• 20,476 examples across all released subsets.
• Geography: Eastern Andhra Pradesh, India.
• Collection: project team and APEPDCL field personnel under operational conditions.

---

Subset Overview

Subset	Task Type	Images	Annotation Type
Pole Detection	Object Detection	2,252	YOLO bounding boxes
Pole Lean Assessment (Classification)	3-Class Image Classification	2,274	Image-level labels
Pole Lean Assessment (Object Detection)	Object Detection	1,801	YOLO bounding boxes
Pole Lean Assessment (Keypoints)	Keypoint / Pose Detection	1,798	YOLO pose (kpt_shape [2, 3])
Crossarm and Top-Cleat Tilt Assessment (Object Detection)	Object Detection	1,686	YOLO bounding boxes
Crossarm and Top-Cleat Tilt Assessment (Keypoints)	Keypoint / Pose Detection	1,686	YOLO pose (kpt_shape [2, 3])
Perspective Validation	Binary Image Classification	4,950 (combined)	Image-level labels
DTR Plinth Clearance	Object Detection (pole–plinth)	476	YOLO bounding boxes
Voltage Level Classification	Object Detection	1,772	YOLO bounding boxes
Vegetation–Conductor	Semantic Segmentation	1,781	COCO polygon annotations

Baseline Benchmarks

Classification baselines use DINOv3 ViT-B/16. Detection baselines use YOLO12m. Keypoint baselines use a YOLO12m pose model evaluated with Object Keypoint Similarity (OKS). The Vegetation–Conductor baseline uses Mask2Former. Full metric tables are reported in each module card.

Module	Model	Split	Primary result
Pole Detection	YOLO12m	valid / test	mAP@0.5 98.81% / 97.08%
Pole Lean Assessment (Object Detection)	YOLO12m	valid / test	mAP@0.5 95.73% / 93.82%
Pole Lean Assessment (Classification)	DINOv3 ViT-B/16	val / test	Acc 83.58% / 84.14%; Macro-F1 80.57% / 81.74%
Pole Lean Assessment (Keypoints)	YOLO12m pose	valid / test	OKS mAP@0.5 98.19% / 97.75%
Crossarm and Top-Cleat Tilt Assessment (Object Detection)	YOLO12m	valid / test	mAP@0.5 92.37% / 92.17%
Crossarm and Top-Cleat Tilt Assessment (Keypoints)	YOLO12m pose	valid / test	OKS mAP@0.5 97.46% / 98.48%
Voltage Level Classification	YOLO12m	valid / test	mAP@0.5 95.73% / 95.72%
DTR Plinth Clearance	YOLO12m	valid / test	mAP@0.5 96.85% / 97.45%
Perspective Validation (Full-image)	DINOv3 ViT-B/16	val / test	Acc 89.36% / 90.43%; Macro-F1 89.33% / 90.41%
Perspective Validation (Pole-top crops)	DINOv3 ViT-B/16	val / test	Acc 95.00% / 96.56%; Macro-F1 94.95% / 96.54%
Perspective Validation (DTR)	DINOv3 ViT-B/16	val / test	Acc 91.11% / 85.19%; Macro-F1 90.89% / 84.89%
Vegetation–Conductor	Mask2Former	valid	mIoU 74.57%; Conductor F1 57.92%

Supported Tasks

1. Pole Detection: localize poles in full-frame imagery, including multi-pole scenes
2. Pole Lean Assessment: leaned versus straight labeling at the image and instance levels, and pole-axis keypoints
3. Crossarm and Top-Cleat Tilt Assessment: six-class Straight/Tilted detection of pole-top components, with structural-axis keypoints
4. DTR Plinth Clearance: pole and plinth detection; an image-plane height ratio may be derived as a clearance-risk proxy (no physical-distance ground truth)
5. Voltage Level Classification: HT or LT from insulator configuration (HT: 11 kV and 33 kV; LT: 220 V and 440 V)
6. Perspective Validation: Valid versus Perspective-Distorted screening at full-image, crop, and DTR views
7. Vegetation–Conductor Segmentation: pixel-level vegetation and conductor masks, with validation-set risk labels

Field Data Collection & Capture Protocols

Field images were collected by utility personnel and the project team across urban and rural locations in Eastern Andhra Pradesh using entry-level Android smartphones under natural lighting conditions:

• Standardized Pole Capture Protocol: Field personnel captured target poles head-on from an approximately centered viewpoint, with both top and base fully visible, at a camera-to-pole distance of approximately 1–1.5 times the pole height.
• Tilt Correction: Initial handheld captures revealed that camera tilt may affect the lean judgment. Early images were manually tilt-corrected. The survey application subsequently integrated automatic tilt correction adapted from the open-source Open Camera project, using smartphone orientation sensors to standardize orientation during acquisition.
• DTR Capture Protocol: Images required the pole top, base, and complete transformer plinth to appear on the same ground plane, visible without mutual overlap.

Annotation Protocol & Quality Assurance

A structured annotation protocol was established to ensure consistency across the different inspection tasks:

• Domain Expert Definition: Domain experts defined the defect categories, visual criteria, and structural boundaries before annotation.
• Annotation Formats: Object-detection and pose datasets were prepared in YOLO format, while vegetation semantic segmentation annotations were prepared in COCO format.
• Quality Control: All annotated datasets were manually reviewed for annotation quality.

Dataset Structure

Data Instances

Object detection (YOLO). Images and labels under train|valid|test/, with data.yaml and optional dataset.json:

<class_id> <center_x> <center_y> <width> <height>

Classification and perspective validation. Folder name equals class label. These subsets use val/ for the validation split. An optional labels.csv is provided.

Keypoint / pose (YOLO pose, kpt_shape: [2, 3]). Released under pole_lean_assessment/keypoint_detection/ and crossarm_top_cleat_tilt_assessment/keypoint_detection/:

<class_id> <cx> <cy> <w> <h> <kx1> <ky1> <v1> <kx2> <ky2> <v2>

Keypoint semantics are defined in the Pole Lean Assessment and Crossarm and Top-Cleat Tilt Assessment module cards (pole / topcleat: bottom to top; crossarm / v-crossarm: left to right).

Segmentation (COCO). _annotations.coco.json with categories 1 = Conductor and 2 = Vegetation; background is implicit. GT_ValidationSet.csv holds Risky/Safe labels for the vegetation validation split.

Images. JPEG, RGB, variable smartphone resolution, typically portrait.

Data Splits

Dataset	Train	Validation	Test	Total
Pole Detection	1,665 (73.93%)	293 (13.01%)	294 (13.06%)	2,252
Pole Lean Assessment (Object Detection)	1,441 (80.01%)	181 (10.05%)	179 (9.94%)	1,801
Pole Lean Assessment (Classification)	1,706 (75.02%)	341 (15.00%)	227 (9.98%)	2,274
Crossarm and Top-Cleat Tilt Assessment (Object Detection)	1,175 (69.69%)	254 (15.07%)	257 (15.24%)	1,686
Voltage Level Classification	1,337 (75.45%)	218 (12.30%)	217 (12.25%)	1,772
DTR Plinth Clearance	362 (76.05%)	57 (11.97%)	57 (11.97%)	476
Perspective Validation (Full-image)	1,605 (74.00%)	282 (13.00%)	282 (13.00%)	2,169
Perspective Validation (Pole-top crops)	1,221 (70.05%)	260 (14.92%)	262 (15.03%)	1,743
Perspective Validation (DTR)	768 (73.99%)	135 (13.01%)	135 (13.01%)	1,038
Vegetation–Conductor Segmentation	1,324 (74.34%)	457 (25.66%)	—	1,781
Pole Lean Assessment (Keypoints)	1,438 (79.98%)	181 (10.07%)	179 (9.96%)	1,798
Crossarm and Top-Cleat Tilt Assessment (Keypoints)	1,175 (69.69%)	254 (15.07%)	257 (15.24%)	1,686
Total	15,217	2,913	2,346	20,476

Dataset Creation

Source Data

• Captured by the project team and APEPDCL field personnel in Eastern Andhra Pradesh, India
• Various Android smartphones; urban and rural; varying lighting, weather, and infrastructure ages
• No crowdsourcing; no demographic data on producers (APEPDCL–research collaboration)

Annotations

Module-specific annotation rules are given in the linked module cards. Protocols were developed with APEPDCL utility engineers. Annotators were the dataset authors and project personnel at APEPDCL.

Personal and Sensitive Information

Images are public roadside captures and may incidentally include identifiable people or private property. Personal data were not collected intentionally, and no facial recognition use is intended. Users should comply with applicable privacy requirements before redistribution.

Considerations for Using the Data

Intended Use

• Research and benchmarking of detection, classification, keypoint / pose, and segmentation models for power distribution infrastructure
• Not a substitute for field verification or regulatory compliance checks

Out-of-Scope and Misuse

• The dataset does not provide calibrated physical measurements; labels and derived quantities use relative geometric indicators.
• Model outputs must not be the sole basis for safety, maintenance, or compliance decisions without field verification and local validation.
• Perspective-validation labels are specific to distribution-infrastructure inspection and are not a general photographic-quality benchmark.
• Transfer to transmission networks or to utilities outside the collected APEPDCL setting requires adaptation and re-validation.
• Detection and segmentation outputs are not risk classes by themselves; additional modeling is required before they represent clearance or defect-risk decisions.

Other Known Limitations

Limitations specific to each subset are given in that module's card. Across the dataset:

1. Occasional annotation errors and subjective borderline lean or tilt cases remain possible.
2. Image quality varies with phone sensor, lighting, motion blur, and focus; resolution is not standardized.
3. Defect classes are typically less frequent than non-defect classes.
4. Coverage is limited to one utility (APEPDCL), with configurations specific to that network.

Additional Information

Dataset Authors

• Sampath Balaji Masabathula — Eastern Power Distribution Company of Andhra Pradesh Limited (APEPDCL); Andhra University
• Srija Vanapalli — Eastern Power Distribution Company of Andhra Pradesh Limited (APEPDCL); Andhra University
• K. Gangadhar — Eastern Power Distribution Company of Andhra Pradesh Limited (APEPDCL)
• V.A.N. Sreenivasa Rao — Eastern Power Distribution Company of Andhra Pradesh Limited (APEPDCL)
• Sasank Chilamkurthy — Von Neumann AI (JOHNAIC), Bengaluru, India

Annotation and data-collection protocols were developed by the authors with domain expertise from K. Gangadhar and V.A.N. Sreenivasa Rao on distribution-infrastructure defects and their visual identification.

Acknowledgments

We thank Marla Krishna Sumwith, Shreya Vanapalli, Thalada Srinivas, Deepesh Reddy Kalapureddy, and Keerthi Subrahmanyam for assistance with field data collection and annotation, and APEPDCL field personnel for assistance with field data collection.

Licensing Information

Dataset license: CC BY 4.0. Associated code, when released, is under the MIT License. Attribution is required.

Citation Information