IDEA (Image Database for Earthquake Damage Annotation) is a dataset developed by Eucentre comprising over 5,400 real images of structural and non-structural damage annotated by structural engineers. These images were captured during post-earthquake inspections in L’Aquila (2009), Emilia (2012) and Central Italy (2016–17), as well as during regular surveys of buildings and infrastructure. Traditional digital cameras and unmanned aerial vehicle (UAV) systems were used to capture the images.
The dataset’s primary objective is to address the shortage of annotated data required for training and validating deep learning algorithms for automatically detecting and classifying structural damage. Unlike many other datasets, IDEA is based on images that are truly annotated (rather than simply classified), collected in real contexts and representative of widespread building types in Southern Europe.
Each image is accompanied by an XML annotation file in Pascal VOC format containing detailed information on:
- presence or absence of damage (damagePresence);
- structural type (e.g. residential buildings, churches or bridges);
- affected element (e.g. reinforced concrete column, beam or masonry wall);
- the type and category of damage (e.g. cracks, spalling, corrosion, collapse)
- type of annotation (rectangular or polygonal);
- The coordinates of the damaged area are also included.
The dataset is divided into two main categories: ‘Damage’, containing images of structural damage, and ‘no_damage’, containing images with no visible damage. This distinction is crucial for reducing false positives in automatic recognition models.
One of IDEA’s strengths is its detailed structural ontology, which was defined by the authors and covers a wide range of damage elements and mechanisms. This system enables annotations to be standardised and allows the dataset to be expanded or new ones to be created according to a shared standard in structural engineering.
IDEA is conceived as a ‘living lab’, designed to be updated with new data over time, and is a reference resource:
- the development and validation of deep convolutional neural networks (DCNNs);
- for structural and infrastructure monitoring;
- post-event inspection, and emergency management;
- It also supports situational awareness for operators and responders.
The open-access dataset is available on Zenodo (DOI: 10.5281/zenodo.15120522).