Dataset Card for InsPLAD

This is a FiftyOne dataset with 49706 samples.

Installation

If you haven't already, install FiftyOne:

pip install -U fiftyone

Usage

import fiftyone as fo
from fiftyone.utils.huggingface import load_from_hub

# Load the dataset
# Note: other available arguments include 'max_samples', etc
dataset = load_from_hub("harpreetsahota/InsPLAD")

# Launch the App
session = fo.launch_app(dataset)

Dataset Details

Dataset Description

InsPLAD (Inspection of Power Line Assets Dataset) is a Power Line Asset Inspection Dataset and Benchmark built from 10,561 high-resolution UAV images of real, operating power lines. It covers the three tasks of a typical power line visual inspection pipeline: (1) object detection of 17 unique power line asset categories in full UAV frames (InsPLAD-det), (2) supervised classification of visible defects on cropped asset images for 5 asset categories, and (3) unsupervised anomaly detection (normal vs. anomalous) on cropped asset images for all 17 asset categories. This FiftyOne build merges all three official sub-datasets into a single dataset, distinguished by a task field and per-task saved views.

• Curated by: André Luiz Vieira-e-Silva, Heitor Felix, Francisco Simões, Veronica Teichrieb (Universidade Federal de Pernambuco / Centro de Informática; Universidade Federal Rural de Pernambuco)
• Funded by: Coordenação de Aperfeiçoamento de Pessoal de Nível Superior (CAPES), Brazil; National Council for Scientific and Technological Development (CNPq), Brazil
• Shared by: The original authors, via GitHub and Mendeley Data
• Language(s): en (asset/defect class names; not an NLP dataset)
• License: cc-by-nc-3.0

Dataset Sources

• Repository: https://github.com/andreluizbvs/InsPLAD
• Paper: InsPLAD: A Dataset and Benchmark for Power Line Asset Inspection in UAV Images, International Journal of Remote Sensing (2023), https://arxiv.org/abs/2311.01619. A companion paper using the same data, Attention Modules Improve Image-Level Anomaly Detection for Industrial Inspection: A DifferNet Case Study, WACV 2024, https://arxiv.org/abs/2311.02747
• Original data download: https://data.mendeley.com/datasets/5n3fjgvfyz/1

Uses

Direct Use

Training and benchmarking object detectors on power line hardware categories; training and benchmarking supervised defect classifiers on cropped asset images; training and benchmarking unsupervised anomaly detection methods on cropped asset images; research into rare-class data curation and annotation-efficient workflows (the class distribution is heavily long-tailed, from 3,169 images down to 56 for the rarest raw category).

Out-of-Scope Use

Any commercial use — the source license (CC BY-NC 3.0) is non-commercial only. The dataset is not intended or suitable for identifying individuals; it contains no personal or sensitive data by design (aerial images of power line hardware only). Asset appearance reflects a specific utility company's hardware standards in Brazil and may not generalize to power line hardware from other manufacturers, regions, or voltage classes without further validation.

Dataset Structure

This is a flat image dataset (media_type = "image"), not grouped or video, with 49,706 samples total, combining InsPLAD's three official sub-datasets into one FiftyOne dataset via a task field. Three saved views are provided, one per sub-task: InsPLAD-det (10,561 samples), InsPLAD-fault-supervised (11,525 samples), InsPLAD-anomaly (27,620 samples).

Fields

Field	FiftyOne type	Description
task	StringField	Which sub-dataset this sample belongs to: detection, fault_classification, or anomaly_detection
ground_truth	Detections	Bounding boxes over 18 raw power line asset categories (InsPLAD-det COCO annotations, converted to relative [x, y, w, h]); populated only for task == "detection"
asset	StringField	Power line asset category of the cropped image (5 distinct values for fault_classification, 17 for anomaly_detection); not populated for detection samples
fault	Classification	Supervised fault-classification label: good, rust, missing_cap, or bird_nest; populated only for task == "fault_classification"
anomaly	Classification	Binary label: normal or anomaly; populated only for task == "anomaly_detection"
defect	StringField	Specific defect subtype for anomaly-detection samples: none, rust, missing_cap, bird_nest, torn, or peeling_paint; populated only for task == "anomaly_detection"
tags	list of str	Encodes both the sub-dataset (det, fault, or anomaly) and the original split (train, val, or test)

Label types and why

• ground_truth uses Detections because InsPLAD-det's source annotations are 2D bounding boxes (no segmentation masks, despite the COCO schema having a segmentation field — it is always empty in the source data).
• fault and anomaly use Classification (not Classifications) because each cropped image in these sub-datasets carries exactly one label, encoded by its containing folder in the source data.
• defect is stored as a plain string field, separate from the anomaly Classification, so the specific defect subtype can be filtered/faceted independently of the binary normal/anomaly label.

dataset.info

dataset.info records source (the GitHub repository URL), paper (the arXiv URL), and license (CC BY-NC 3.0).

Parsing decisions

• Duplicate COCO entries merged. InsPLAD-det's instances_train.json contains 46 file_names that appear under two different image_ids, each holding a disjoint subset of that image's boxes. These were merged into one sample per file with the union of both image_ids' detections. As a result, the true unique image count for InsPLAD-det is 10,561, not the 10,607 implied by the paper's headline image count.
• Bounding box conversion. Source boxes are absolute-pixel COCO [x, y, w, h]; converted to FiftyOne's relative [0, 1] [x, y, w, h] by dividing by each image's width/height.
• sphere category kept as-is. InsPLAD-det's COCO categories list includes an 18th category, sphere (26 boxes, train split only), which is not one of the paper's 17 named asset categories or its per-class benchmark table. It is retained in ground_truth rather than dropped, since the source data includes it.
• Fault-classification label vocabulary normalized. In the source data, supervised_fault_classification's polymer-insulator-upper-shackle asset used Portuguese folder names in its train split (corrosão, normal) but English folder names in val (rust, good) for the same two conditions. Both were mapped to rust/good for consistency with the other four fault-classification asset categories.
• Minor spelling normalized across sub-datasets. The same defect concept was spelled differently between the supervised and unsupervised sub-datasets in the source data (missing-cap vs. missingcap; bird-nest vs. nest). Both were normalized to missing_cap and bird_nest respectively across both fault and defect fields.
• defect uses "none" for normal anomaly-detection samples, rather than leaving the field unset, for consistent filtering/faceting.

Dataset Creation

Curation Rationale

Automated visual inspection of power line assets is an active research area, but most power line inspection datasets are proprietary, undisclosed by the utility companies and agencies that hold them. InsPLAD was created to provide the first large, real-world, publicly available dataset and benchmark covering all three stages of a typical inspection pipeline — asset detection, defect classification, and anomaly detection — on the same underlying asset categories.

Source Data

Data Collection and Processing

Images were captured by UAV (drone) during real-world inspections of operating power lines, at 1920×1080 resolution, under varied environmental conditions, orientations, and distances. InsPLAD-fault's two sub-datasets (supervised_fault_classification and unsupervised_anomaly_detection) were derived from InsPLAD-det: annotated objects for five (supervised) and seventeen (anomaly detection) asset categories were cropped to a near-square proportion and labeled at the image level according to visible condition (normal, or the specific defect present). For supervised fault classification, faulty training samples were augmented roughly 10x using the Albumentations library to balance otherwise rare defect classes; for unsupervised anomaly detection, defective examples were deliberately withheld from training and included only in the test split, consistent with the unsupervised anomaly detection setting.

Who are the source data producers?

UAV inspection imagery of real, operating power lines, collected by the Voxar Labs group at Universidade Federal de Pernambuco.

Annotations

Annotation process

Bounding boxes for InsPLAD-det were drawn using the LabelImg tool by two annotators, instructed by specialists on asset naming conventions and the regions of interest that should be contained in each box, with annotators in constant contact to standardize labeling. Fault and anomaly labels were applied at the image level based on the visible condition of each cropped asset.

Who are the annotators?

Two human annotators, instructed and guided by domain specialists, per the source paper.

Personal and Sensitive Information

Not addressed explicitly in the source paper. Images are aerial captures of power line hardware and surrounding infrastructure; there is no statement in the source material regarding incidental capture of people or other personal data. [More Information Needed]

Citation

BibTeX:

@article{doi:10.1080/01431161.2023.2283900,
   author    = {André Luiz Buarque Vieira e Silva, Heitor de Castro Felix, Franscisco Paulo Magalhães Simões, Veronica Teichrieb, Michel dos Santos, Hemir Santiago, Virginia Sgotti and Henrique Lott Neto},
   title     = {InsPLAD: A Dataset and Benchmark for Power Line Asset Inspection in UAV Images},
   journal   = {International Journal of Remote Sensing},
   volume    = {44},
   number    = {23},
   pages     = {1-27},
   year      = {2023},
   publisher = {Taylor & Francis},
   doi       = {10.1080/01431161.2023.2283900},
   URL       = {https://doi.org/10.1080/01431161.2023.2283900},
   eprint    = {https://doi.org/10.1080/01431161.2023.2283900},
}

@InProceedings{Vieira_2024_WACV,
    author    = {e Silva, Andr\'e Luiz Vieira and Sim\~oes, Francisco and Kowerko, Danny and Schlosser, Tobias and Battisti, Felipe and Teichrieb, Veronica},
    title     = {Attention Modules Improve Image-Level Anomaly Detection for Industrial Inspection: A DifferNet Case Study},
    booktitle = {Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV)},
    month     = {January},
    year      = {2024},
    pages     = {8246-8255}
}

APA:

Vieira-e-Silva, A. L. B., de Castro Felix, H., Simões, F. P. M., Teichrieb, V., dos Santos, M., Santiago, H., Sgotti, V., & Lott Neto, H. (2023). InsPLAD: A Dataset and Benchmark for Power Line Asset Inspection in UAV Images. International Journal of Remote Sensing, 44(23), 1-27.

More Information

This FiftyOne build merges the three official InsPLAD sub-datasets (InsPLAD-det.zip, supervised_fault_classification.zip, unsupervised_anomaly_detection.zip) into one dataset. See the source repository (https://github.com/andreluizbvs/InsPLAD) for the original, separately-distributed sub-datasets and additional documentation.

Dataset Card Authors

Harpreet Sahota (FiftyOne / Voxel51 conversion and card)

Dataset Card Contact

harpreetsahota