A Semantically Annotated 15-Class Ground Truth Dataset for Substation Equipment

This dataset contains 1660 images of electric substations with 50705 annotated objects. The images were obtained using different cameras, including cameras mounted on Autonomous Guided Vehicles (AGVs), fixed location cameras and those captured by humans using a variety of cameras. A total of 15 classes of objects were identified in this dataset, and the number of instances for each class is provided in the following table:

All images in this dataset were collected from a single electrical distribution substation in Brazil over a period of two years. The images were captured at various times of the day and under different weather and seasonal conditions, ensuring a diverse range of lighting conditions for the depicted objects. A team of experts in Electrical Engineering curated all the images to ensure that the angles and distances depicted in the images are suitable for automating inspections in an electrical substation.

The file structure of this dataset contains the following directories and files:

 images: This directory contains 1660 electrical substation images in JPEG format.

images: This directory contains 1660 electrical substation images in JPEG format.

• labels_json: This directory contains JSON files annotated in the VOC-style polygonal format. Each file shares the same filename as its respective image in the images directory.
• 15_masks: This directory contains PNG segmentation masks for all 15 classes, including the porcelain pin insulator class. Each file shares the same name as its corresponding image in the images directory.
• 14_masks: This directory contains PNG segmentation masks for all classes except the porcelain pin insulator. Each file shares the same name as its corresponding image in the images directory.
• porcelain_masks: This directory contains PNG segmentation masks for the porcelain pin insulator class. Each file shares the same name as its corresponding image in the images directory.
• classes.txt: This text file lists the 15 classes plus the background class used in LabelMe.
• json2png.py: This Python script can be used to generate segmentation masks using the VOC-style polygonal JSON annotations.

The dataset aims to support the development of computer vision techniques and deep learning algorithms for automating the inspection process of electrical substations. The dataset is expected to be useful for researchers, practitioners, and engineers interested in developing and testing object detection and segmentation models for automating inspection and maintenance activities in electrical substations.

The authors would like to thank UTFPR for the support and infrastructure made available for the development of this research and COPEL-DIS for the support through project PD-2866-0528/2020—Development of a Methodology for Automatic Analysis of Thermal Images. We also would like to express our deepest appreciation to the team of annotators who worked diligently to produce the semantic labels for our dataset. Their hard work, dedication and attention to detail were critical to the success of this project.